Skip to content

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

docs: breaking changes page lists each change under the release that shipped it (#15592)

Bartosz Majewski70e0505fd257
diff --git a/docs/install/architecture/benchmark.mdx b/docs/install/architecture/benchmark.mdx
index 2e12a577966..c4f55b6af8d 100644
--- a/docs/install/architecture/benchmark.mdx
+++ b/docs/install/architecture/benchmark.mdx
@@ -118,7 +118,7 @@ This is the time the **worker** is occupied per job — and at concurrency 1 it
- **Worker**: 0.5 vCPU / 1 GB, concurrency 1, `SANDBOX_CODE_ONLY` (Node fork + `isolated-vm`). Idle RSS ~145 Mi
- **App**: 1 vCPU / 1 GB
- **Object store**: same-region GCS bucket (`europe-west1`) over the S3-interop endpoint, path-style SigV4 presigned URLs (`AP_S3_USE_SIGNED_URLS=true`)
-- **Piece bundles**: official tarballs served from the Activepieces CDN (`AP_USE_CDN_FOR_BUNDLES=true`)
+- **Piece bundles**: official tarballs served from the Activepieces CDN (`AP_USE_CDN_FOR_BUNDLES=true`; the default is `false`, which serves from npm)
- **Postgres + Redis**: in-cluster singletons, deliberately generous — Postgres requesting 3 vCPU / 3 GB with `max_connections=2000` (the default 100 would starve the app pools past ~10 apps), durability off, and its data dir on tmpfs; Redis requesting 2 vCPU / 2 GB with `io-threads`. Neither has a CPU *limit*, so both may burst above their request; Postgres consumed 2738m at the top tier
- **Load**: `hey` **run inside the cluster**, against the app Service, concurrency matched to worker count (40/80/120/160) so requests don't queue behind the concurrency-1 workers — latency reflects real service time, not backlog. 400 requests per worker per tier (16k/32k/48k/64k), preceded by an unmeasured warmup pass
diff --git a/docs/install/reference/breaking-changes.mdx b/docs/install/reference/breaking-changes.mdx
index 57db79da402..d230c1c3e43 100644
--- a/docs/install/reference/breaking-changes.mdx
+++ b/docs/install/reference/breaking-changes.mdx
@@ -8,46 +8,82 @@ icon: "hammer"
### What has changed?
-#### Worker group runs no longer fall back to shared workers
+#### The AI provider config endpoint is removed
-Runs for a project assigned to a worker group used to fall back to the shared queue whenever the group had no online workers, so shared workers would execute them. Runs now always stay on the group's dedicated queue (`project-<group-name>-jobs`) and wait there until a worker carrying the matching `AP_WORKER_GROUP_ID` connects — shared workers never pick them up.
+`GET /v1/ai-providers/:provider/config` returned a provider's decrypted credentials so that an AI step running inside the engine could build the model itself and call the provider directly. AI steps no longer work that way: the worker resolves the credentials and makes the call, which is what lets Activepieces charge managed AI on what the call actually cost. Nothing in the product calls the endpoint any more, and it now returns `404`.
+
+The route only ever accepted an engine token, so it was never reachable with a user token, an API key or from the browser. `GET /v1/ai-providers` and `GET /v1/ai-providers/:provider/models` are unchanged.
#### What you need to do
-If you assign projects to worker groups, keep at least one worker online per assigned group. If a group is down, reassign the project to another group or unassign it (set its `workerGroupId` to `null`) — any assignment change moves the project's waiting runs to the new queue.
-#### The legacy product-embed connection-key and app-credential endpoints are removed
+Nothing on a normal upgrade. Every flow is moved to the AI piece version that runs on the worker the first time its version is read, before it can run.
-The original product-embed flow is gone. It let a host application store per-project OAuth client credentials and a public key, then mint an app connection by presenting a token signed with the matching private key. These endpoints are removed:
+If you serve pieces from your own registry, make sure it carries the AI piece at `0.11.0` or later. A flow left on an earlier version would still try to fetch its provider credentials over this endpoint and fail.
-- `GET`, `POST` and `DELETE` `/v1/connection-keys/app-connections`
-- `GET`, `POST` and `DELETE` `/v1/connection-keys`
-- `GET`, `POST` and `DELETE` `/v1/app-credentials`
+#### AI steps are billed on what each call costs, not on a fixed price per model
-They were registered on the Cloud edition only, so a self-hosted installation never served them, on either Community or Enterprise. The user interface that drove them was removed several releases ago and no SDK or documented flow referenced them. They now return `404`.
+The managed Activepieces AI provider used to charge a fixed number of credits per model per step, taken from a table calibrated at roughly a thousand tokens per call. A short call was charged about right; a long one was charged the same as a short one while costing us many times more.
-The current embedding stack is unaffected: signing keys, managed authentication, JWT user provisioning and the embed SDK are a separate feature and continue to work unchanged.
+Every managed AI call is now billed on the dollar cost the provider reports for that call, converted at `AP_AI_CREDIT_USD_VALUE` (default `0.0005`, so one credit is a twentieth of a cent of model spend). A call on your own API key is billed a flat one credit per model call. Tool calls stay at one credit each.
+
+This changes what the same flow costs in both directions. A short prompt on an expensive model gets much cheaper; a long document through Extract Structured Data gets dearer. The credit cost shown in the model picker and in the credits breakdown is gone, because there is no longer a fixed number to show.
+
+A step or a chat turn on your own API key costs the same as it did: one credit, plus one per tool call. An agent still counts as one credit however many times it goes back to the model.
#### What you need to do
-Nothing for a self-hosted installation, and no environment variable or migration. The `app_credential` and `connection_key` tables are left in place by this release and are simply no longer read; a later release drops them.
+Nothing to configure. Self-hosters who want a different conversion rate can set `AP_AI_CREDIT_USD_VALUE` to the dollar value of one credit.
-If you call one of the endpoints above against Activepieces Cloud, move to the current embedding flow, which provisions users and connections through a platform signing key. See [Embedding](/embedding/overview).
+#### AI steps now consume credits on your own API key too
-#### Brevo's Create or Update Contact stops clearing blacklist flags and attributes it was not given
+`POST /v1/ai/execute` refuses to start an AI step when the platform is out of credits, whatever provider the step uses. Previously only the managed Activepieces provider was gated, so a platform that had exhausted its credits kept running AI steps as long as they used its own key.
-Four changes to the Brevo (formerly Sendinblue) piece's Create or Update Contact action alter what an existing, unchanged step does.
+Credits are only enforced where a billing provider is configured, so a Community self-host is unaffected.
-The action used to strip every falsy value from the request before sending it, so `Email Blacklisted` and `SMS Blacklisted` could be switched on but never switched back off — the `false` was dropped and the contact stayed blacklisted with no error. Those flags are now sent when you set them, and neither checkbox defaults to off any more, so a checkbox you never touched is omitted from the request instead of quietly un-blacklisting the contact on every run.
+#### What you need to do
-The `Attributes` field used to come pre-filled with nine empty values (`FIRST_NAME`, `LAST_NAME`, `SMS`, `CIV`, `DOB`, `ADDRESS`, `ZIP_CODE`, `CITY`, `AREA`). Because they were sent on every call, running the step overwrote whatever those attributes held in Brevo with empty strings. The default is removed, and only the attributes you list are written.
+Nothing, unless you run AI steps on your own provider key on a platform that sits at zero credits. Those steps now fail with a credits error instead of running. Top up, or raise the plan's credit allowance.
-`List IDs` was a free-text list of numbers and is now a dropdown of the lists in your account. A stored numeric id is still sent exactly as before, so existing steps keep working unchanged.
+#### Files passed to an AI step are capped at `AP_MAX_FILE_SIZE_MB`
-The `SMTP Blacklist Sender` checkbox is replaced by a `Blocked Sender Addresses` list. Brevo expects a list of sender email addresses here, not a true or false, and rejects a boolean outright with *"Invalid smtpBlacklistSender format"* — so ticking that checkbox always failed the step, and leaving it clear did nothing at all. The field never worked. It is now a list of addresses, which Brevo accepts, and any value your step already had is ignored rather than sent.
+Extract Structured Data and Generate Image used to read file inputs straight from the flow's memory, so no Activepieces-side size limit applied. Those files are now stored and fetched by the worker that runs the model, which puts them under the same per-file limit as every other file the engine writes: `AP_MAX_FILE_SIZE_MB`, 25 MB by default.
+
+The limit is per file, not per step, so several large files in one step are fine as long as each is under it. On Activepieces Cloud the limit is 10 MB and is not configurable.
#### What you need to do
-Nothing to configure or migrate, and no existing step stops working. Review any flow whose Create or Update Contact step relied on the old behaviour: a step that was silently wiping attributes will now leave them intact, and a step you were relying on to blacklist contacts must have the checkbox explicitly ticked. To block a sender for a contact, list that sender's address in `Blocked Sender Addresses` — it must be an active sender in your Brevo account, or Brevo answers *"One of the sender is invalid or inactive"*.
+Nothing unless an AI step handles files larger than 25 MB. Those steps now fail with a size error naming the limit. Raise `AP_MAX_FILE_SIZE_MB` on a self-hosted installation if you need more.
+
+## 0.91.0
+
+### What has changed?
+
+#### WhatsScale media actions take a file input, capping media at the instance file-size limit
+
+The 17 WhatsScale actions that send media — image, video, document and audio sends across the contact, group, channel and CRM-contact variants, plus the image and video Status posts — used to take a plain text URL and hand it straight to WhatsScale, which fetched it itself. Those fields are now file inputs: the engine resolves whatever you pass (a URL, or a file from a previous step) into a file, stores it on your instance, and gives WhatsScale that copy to fetch.
+
+The field keys are unchanged (`imageUrl`, `videoUrl`, `documentUrl`, `audioUrl`), so existing flows keep their configured value and a stored URL still works. Two behaviour changes come with the new path:
+
+- **Media is now capped by `AP_MAX_FILE_SIZE_MB` — 10 MB on Cloud, 25 MB by default self-hosted.** The file travels through your instance instead of going vendor-to-vendor, so a URL pointing at something larger than that limit now fails where it previously sent. This mostly affects video and document sends. Streaming does not lift the ceiling: writing into storage is counted against the cap either way, and producing a URL for WhatsScale to fetch requires writing into storage. See [Limits](/install/reference/limits).
+- **The URL WhatsScale fetches is now one your instance serves.** It is a signed, expiring link on your public API URL. An instance that is not reachable from the internet cannot hand WhatsScale a URL it can fetch, so media sends need a publicly resolvable `AP_FRONTEND_URL` on self-hosted installs.
+
+#### What you need to do
+
+Nothing if your media is comfortably under the cap and your instance is reachable from the internet. On Cloud that cap is 10 MB, so check any flow that forwards video or large documents.
+
+If you send files larger than the cap on a self-hosted install, raise `AP_MAX_FILE_SIZE_MB` to cover the largest media you send and redeploy. On Cloud the 10 MB cap is not configurable. If your instance is not publicly reachable, keep media sends on a host that is, or leave these actions unused; there is no way for WhatsScale to fetch a file from a private address.
+
+#### WhatsScale "Add a Tag to a CRM Contact" now takes a list of tags
+
+The two add-tag actions ("Add Tags to a CRM Contact" and its By-ID twin) accepted a single tag in a text field named `tag`. They now accept a list, in a field named `tags`, and send one request per tag so existing tags are kept.
+
+Because the field was renamed, **a step configured before this release loses its tag value** and shows "Tags" as an empty required field. The field could not keep its old name: a stored single string does not validate as a list.
+
+Adding several tags is **not atomic** — each tag is its own request, so a failure part-way through leaves the earlier tags applied. Re-running is safe: adding a tag a contact already has converges on the same set.
+
+#### What you need to do
+
+Re-enter the tag on any flow that uses either add-tag action, as a list entry rather than free text. The removal actions are unchanged and still take one tag at a time.
#### A custom AI provider is now checked when you save it
@@ -69,124 +105,139 @@ An agent step, saved agent, or agent run that names a model outside the offered
Nothing, if your agent steps use the model picker. If a flow, a saved agent, or an API call names a model id directly while running on Activepieces AI credits, switch it to one of the offered models or to your own provider key. The error message lists the available ids.
-#### `AP_FLOW_TIMEOUT_SECONDS` now also bounds a single action run
+#### The legacy product-embed connection-key and app-credential endpoints are removed
-An action run — an agent's configured piece tool, a chat action, or MCP `ap_run_action` — used to be capped at a hardcoded 120 seconds, and a flow run started as an MCP tool at a hardcoded 300 seconds. Neither could be changed. Any agent tool whose action took longer than a minute failed outright with `RPC [executePieceTool] failed (timeout: 60000ms)`, because a second hardcoded limit in the worker cut it off first.
+The original product-embed flow is gone. It let a host application store per-project OAuth client credentials and a public key, then mint an app connection by presenting a token signed with the matching private key. These endpoints are removed:
-Both hardcoded limits are gone. All of them now follow `AP_FLOW_TIMEOUT_SECONDS`, the same setting that bounds a flow run, which defaults to 600 seconds. An action run therefore gets the same budget as a flow step.
+- `GET`, `POST` and `DELETE` `/v1/connection-keys/app-connections`
+- `GET`, `POST` and `DELETE` `/v1/connection-keys`
+- `GET`, `POST` and `DELETE` `/v1/app-credentials`
-The limit going up is the change to be aware of: a slow or hung action now occupies a worker slot for up to `AP_FLOW_TIMEOUT_SECONDS` instead of 120 seconds, and action runs are not covered by the per-project concurrency limiter. On a small worker pool, several slow actions can hold execution capacity for longer than they used to.
+They were registered on the Cloud edition only, so a self-hosted installation never served them, on either Community or Enterprise. The user interface that drove them was removed several releases ago and no SDK or documented flow referenced them. They now return `404`.
+
+The current embedding stack is unaffected: signing keys, managed authentication, JWT user provisioning and the embed SDK are a separate feature and continue to work unchanged.
#### What you need to do
-Nothing to configure or migrate, and no new environment variable — a default installation gets the fix with no action. Two things to consider if they apply to you. If you raised `AP_FLOW_TIMEOUT_SECONDS` for long flows, note that agent tools, chat actions and MCP action runs now inherit that same ceiling; lower it if you do not want them held that long. If you put Activepieces behind a reverse proxy and use MCP `ap_run_action`, an action that runs past the proxy's read timeout (nginx defaults to 60 seconds, Cloudflare to 100) returns a gateway error to the MCP client while the action keeps running — raise `proxy_read_timeout` if you rely on long action runs over MCP.
+Nothing for a self-hosted installation, and no environment variable or migration. The `app_credential` and `connection_key` tables are left in place by this release and are simply no longer read; a later release drops them.
-#### A Run Agent step whose turn dies now fails its flow run
+If you call one of the endpoints above against Activepieces Cloud, move to the current embedding flow, which provisions users and connections through a platform signing key. See [Embedding](/embedding/overview).
-A Run Agent step whose turn died — no AI provider configured for the chosen vendor, a revoked API key, a model the provider no longer serves — used to report success. The step returned the failed agent result rather than raising it, so the step read SUCCEEDED, the run completed SUCCEEDED, and later steps acted on a payload that carried the failure inside it. Such a run is now FAILED, and the agent's error message is the step's error.
+#### Kimai connections now authenticate with an API token instead of a username and API password
-Two cases are deliberately unaffected, because the agent did produce usable output:
+Kimai is deprecating its legacy `X-AUTH-USER` / `X-AUTH-TOKEN` authentication (API passwords), with removal planned no later than July 2026 — see [Kimai's announcement](https://www.kimai.org/en/blog/2026/removing-api-passwords). The Kimai piece's connection now sends `Authorization: Bearer <token>` instead, matching Kimai's current API. This also fixes the Activity dropdown on the Create Timesheet action, which ignored the selected project because it tried to filter on a GET request body that the piece's HTTP client cannot send.
-- A turn that finished but had a single tool call error. The agent may have recovered from it and reported on it, so those runs keep succeeding exactly as before.
-- A turn that was cut short by the output limit or the run's usage budget. It still returns the work it completed, and the flow still continues.
+The connection's `Username` and `API Password` fields are removed and replaced with a single `API Token` field.
#### What you need to do
-Nothing to configure or migrate, and no new environment variable. Two things to check if either applies to you. If a flow relied on continuing past a dead agent step, tick **Continue on failure** on that step — it now takes effect, where previously the step never failed for it to apply to. If a flow branches on the agent result (a Router reading the step's `status`), that branch no longer runs when the turn dies, because the step raises instead of returning; move that handling onto the step's failure path.
-
-If you alert on run status, expect agent flows that were silently completing on a misconfigured provider to start showing as failed. That is the misconfiguration surfacing, not new breakage.
+Only if you have an existing Kimai connection. Reconnect it: in your Kimai instance, generate an API token from your user profile's "API Access" page, then re-enter your Kimai connection in Activepieces with the Server URL and that token. Existing connections using the old username/API-password fields will fail to authenticate until reconnected.
-#### Workers no longer pre-warm the flow cache on startup by default
+#### Worker group runs no longer fall back to shared workers
-Since v0.86.1 every worker pre-filled its local piece and code cache on startup by resolving and compiling every enabled flow on the platform. That warm-up costs memory and CPU proportional to the number of enabled flows: on instances with many flows it pinned each worker at its CPU limit for the duration and spiked memory enough to OOM-kill small workers, especially during upgrades when all workers restart at once. The warm-up is now opt-in behind the new `AP_PREWARM_CACHE_ON_STARTUP` worker environment variable, which defaults to `false`. When disabled, caches fill lazily on each flow's first run after a worker starts, exactly as they did before v0.86.1.
+Runs for a project assigned to a worker group used to fall back to the shared queue whenever the group had no online workers, so shared workers would execute them. Runs now always stay on the group's dedicated queue (`project-<group-name>-jobs`) and wait there until a worker carrying the matching `AP_WORKER_GROUP_ID` connects — shared workers never pick them up.
#### What you need to do
-Nothing, unless you want to keep the pre-v0.88 warm-up behaviour. Set `AP_PREWARM_CACHE_ON_STARTUP=true` on your worker containers to restore it — recommended only if your instance has a modest number of enabled flows and your workers have memory headroom. See [Environment Variables](/install/reference/environment-variables) for details.
-
-#### A new workspace is named after the company in the sign-up email
+If you assign projects to worker groups, keep at least one worker online per assigned group. If a group is down, reassign the project to another group or unassign it (set its `workerGroupId` to `null`) — any assignment change moves the project's waiting runs to the new queue.
-Completing sign-up used to always name the new platform after the person, as `"Ahmad's Platform"`. It now reads the email domain first: `[email protected]` creates a platform called `Activepieces`, and the project alongside it becomes `Activepieces's Project`.
+## 0.90.4
-The person-based name is still what you get from a consumer address. `[email protected]` continues to produce `"Ahmad's Platform"`, as do the other common providers (Outlook, Yahoo, iCloud, Proton, GMX, QQ and similar). Whether an address counts as a work address is decided by a denylist of consumer providers, so anything not on that list is treated as a company.
+### What has changed?
-This only affects platforms created from here on. Existing platforms keep their names, and renaming stays available in settings.
+#### Telemetry moves from `AP_TELEMETRY_ENABLED` to the platform admin UI
-#### What you need to do
+Product analytics used to be governed only by the `AP_TELEMETRY_ENABLED` environment variable, read once when the app started. It is now a per-platform setting stored in the database, editable at **Platform → Infrastructure → Configurations**.
-Nothing. No configuration, no new environment variable, and nothing to migrate. If you script or test first-run sign-up and assert on the generated platform or project name, update that assertion: a work-domain address no longer yields `"<FirstName>'s Platform"`.
+Upgrading does not change what you collect. Each platform's setting is created the first time it is read, taking `AP_TELEMETRY_ENABLED` as its value, so an instance running with `AP_TELEMETRY_ENABLED=false` stays opted out — on an existing install and on a brand-new one alike.
-#### Piece builds fail when piece code uses `__dirname` without declaring `bundleForkedEntries`
+After a platform's setting exists, the environment variable no longer governs it. Editing the variable has no effect on that platform, and switching the toggle on in the UI turns product analytics on even if the variable still reads `false`.
-The piece bundler now emits files a piece loads by path at runtime (for example a `child_process.fork` target) beside the main bundle, when they are declared in a `bundleForkedEntries` array in the piece's `package.json`, and it keeps dependencies that only those files import in the published manifest. Because an undeclared `__dirname`-relative file access always breaks after publishing — this is exactly how `@activepieces/piece-oracle-database` 0.1.11/0.1.12 shipped with every new connection failing — the build now fails loudly when a piece's source references `__dirname` and declares no forked entries. Previously such a piece built successfully and shipped broken.
+The setting is per platform, so on an instance hosting several platforms each one is controlled separately. The page is hidden on Activepieces Cloud.
#### What you need to do
-Nothing for catalog pieces — oracle-database is the only piece that forks a sibling file, and this change fixes it (0.1.13). If you build custom pieces and one references `__dirname`, either declare the runtime-loaded file in `bundleForkedEntries` (see [Bundling Pieces](/build-pieces/misc/bundling-pieces)) or remove the `__dirname` usage. Already-published pieces keep working; the check runs at build time only.
+Nothing, if you want to keep collecting what you collect today, and nothing if you rely on `AP_TELEMETRY_ENABLED=false` — that value is carried into each platform's setting as it is created. What changes is that the variable becomes a starting value rather than a permanent lock: once a platform has a setting, an admin can switch product analytics on from the UI regardless of what the variable says. If that matters to you, restrict who holds platform-admin rights.
-#### MCP OAuth: revoking a token requires a client identity, and registration issues a usable secret
+## 0.90.1
-Two changes to the MCP OAuth endpoints:
+### What has changed?
-- `POST /revoke` used to revoke a refresh token for a caller that presented no client identity at all, so anyone who learned a refresh token could disconnect its owner. It now requires the request to identify the client, and a confidential client to authenticate, per RFC 7009. A revocation that sends only `token` now returns `400 invalid_client`.
-- Dynamic client registration (`POST /register`) that omits `token_endpoint_auth_method` now records the client as `client_secret_basic` and issues a client secret, per RFC 7591. It previously issued a secret but recorded the client as public, so that secret was never checked. A client registered from now on must present it when calling `/token` and `/revoke`. Concretely: a client that omits `token_endpoint_auth_method` **and** discards the `client_secret` it is handed now gets `400 invalid_client` at `/token`, where it previously succeeded. Such a client must either store and send that secret, or register explicitly with `token_endpoint_auth_method: "none"` to stay public.
+#### AWS connections using the IAM Role / OIDC method are now verified against AWS before they save
-#### What you need to do
+Saving an AWS connection (Bedrock, S3, Secrets Manager) with the **IAM Role / OIDC** method previously only checked that the Role ARN was well formed, so any well-formed ARN saved with status "Connected" even when the role did not exist or could not be assumed. The connection now performs the same `sts:AssumeRoleWithWebIdentity` call the flow makes at runtime, and the save fails with the AWS error when the role cannot be assumed.
-Nothing for already-connected clients. A client's authentication method is read from the row it registered with, so existing registrations are untouched, including the ones created before this fix that hold a secret while recorded as public. Those keep working whether or not they send it.
+Two consequences. A connection that was accepted before is now rejected if its IAM OIDC provider, trust policy, or `AP_FRONTEND_URL` issuer is not set up correctly. And because the save now calls AWS, an instance with no outbound access to `sts.<region>.amazonaws.com` cannot create these connections at all.
-Claude, Cursor and anything else built on the MCP SDK are unaffected: the SDK stores and sends the `client_secret` it is given, and it never calls the revocation endpoint. A confidential client's secret is accepted from either the request body or the `Authorization: Basic` header, so a client is never rejected for choosing one over the other.
+The region on these connections is also validated against the AWS region format at both save time and every flow step (the region flows into the STS endpoint hostname, so the check is enforced everywhere it is used). The STS call now times out after 10 seconds instead of waiting for the job timeout.
-If you maintain your own MCP OAuth client, check two things before upgrading: that it sends `client_id` when it revokes a token, and that it stores and presents the `client_secret` that `/register` returns. Note that disconnecting and re-adding an MCP server creates a fresh registration, so an existing user meets the new behaviour when they reconnect rather than at upgrade.
+#### What you need to do
-#### Agent steps move from the flow engine to the server
+Only if you use the IAM Role / OIDC method on an AWS piece. Allow outbound HTTPS to `sts.<region>.amazonaws.com` from the machine running the worker, and make sure `AP_FRONTEND_URL` matches the Provider URL of your IAM OIDC identity provider, since the issuer in the signed token is derived from it. Existing saved connections keep working and are not revalidated automatically — pressing **Recheck** on a connection whose trust policy is wrong will now move it to an error state, which reflects what already happens when a flow runs.
-An agent step used to run its loop inside the engine, holding a worker for as long as the agent took. It now pauses the flow, runs on the server, and resumes when the agent finishes. This upgrade rewrites every existing agent step to the new version.
+## 0.90.0
-The step's inputs and its output shape are unchanged, so anything reading the step's result or its structured output keeps working. Sub-flow, MCP server, knowledge base and piece action tools all run on the new path.
+### What has changed?
-Two differences are worth knowing before you upgrade:
+#### Front's Attachments field takes files instead of URLs
-- **Web search is no longer a step setting.** It follows what the platform has configured rather than a per-step toggle. A step that had it switched on keeps working; the stored value is ignored.
-- **Max steps is honoured up to a ceiling.** The value you set still applies, but the server caps a single run at 50 turns. A step configured above that runs 50.
+The Front piece's Send Message, Send Reply, Create Draft and Create Draft Reply actions described `Attachments` as a list of attachment URLs and sent those strings inside a JSON body. Front only accepts attachments on a `multipart/form-data` request, as the bytes of the file, so it took the message and dropped the attachments — HTTP 202, no error, no attachment. The field is now a list of files, the same shape the Gmail and Discord pieces use, and the request is sent as multipart when a message carries attachments.
+
+A URL is still all you need to supply: the engine downloads it and hands the piece the resolved file. A step that already holds a file — an earlier download, a trigger's attachment — can be wired straight in. An entry whose file cannot be resolved is skipped rather than failing the send.
#### What you need to do
-Nothing on upgrade. Check any agent step that relied on the per-step web search toggle, or that set max steps above 50, since those now follow the platform and the server ceiling.
+Re-pick the attachment in any Front step that used the field. The stored value is a plain string and the field now holds a file, so it does not carry across. A flow that never set Attachments is unaffected and keeps its JSON request path. Nothing to configure on the server, and no new environment variable.
-#### The container no longer bundles PM2 — crashes exit the container
+#### A file URL that fails to download is no longer treated as a file
-The Docker image used to run the app and worker under PM2, which restarted a crashed process *inside* the container. PM2 is removed; the entrypoint now launches the processes with plain `node`. When a process crashes (or is OOM-killed), the container exits instead of silently recycling the process in place, and your orchestrator restarts it.
+Any piece property that takes a file also accepts a URL, and the engine buffered whatever came back from that URL without checking the response status. A 4xx or 5xx produced a file named after the URL whose contents were the server's error page, and a piece could not tell that apart from the real thing — an expired signed link became an email with the storage provider's XML error attached. A failed download now fails the step, which is what the streaming path already did.
#### What you need to do
-Nothing if you deploy with Docker Compose, Kubernetes/Helm, or any orchestrator that already restarts failed containers (all official deployments do). If you run the image with a bare `docker run` and relied on PM2 to keep the container alive across process crashes, add a restart policy: `docker run --restart unless-stopped ...`.
+Nothing to configure. A flow that was silently passing on error pages will start failing at that step instead, and the fix is the URL itself — usually a link that has expired or that the instance is not authorized to read.
-#### The Microsoft Teams Bot piece no longer needs a server-side installation record
+#### Signing in with an emailed code is now offered on Activepieces Cloud only
-`POST /v1/teams-bot/webhook` and `POST /v1/teams-bot/send` are removed, and the `teams_bot_installation` table is dropped. The webhook existed only to capture Microsoft's per-tenant `serviceUrl` at install time so the send endpoint could look it up.
+`POST /v1/authentication/otp/request` and `POST /v1/authentication/otp/verify` are no longer served on a self-hosted
+instance. They return `404`, and the sign-in card no longer offers the "email me a code" step — it opens on the
+password form instead. Versions 0.88.2 through 0.89.0 did serve them on any edition with SMTP configured.
+`POST /v1/authentication/complete-sign-up` is unaffected and still served everywhere.
-The piece now posts to Microsoft's documented global Teams service endpoint directly, so it holds no server-side state and its messaging endpoint moves to the shared app-webhooks route, `/api/v1/app-events/microsoft-teams-bot`. Because the installed-or-not check is no longer a cached row, it is answered by Microsoft at send time: posting to a team the bot is not a member of now fails with `403 The bot is not part of the conversation roster` instead of a stale local lookup.
+A six-digit code is a million possibilities, so what keeps guessing it expensive is the captcha in front of the
+request endpoint. Cloudflare Turnstile is opt-in and unset out of the box, so a self-hosted instance served both
+endpoints with nothing in front of them. Rather than require a Cloudflare account to make a sign-in method safe,
+that method is no longer served where the captcha is not there.
-The piece is published as `0.1.0` with a minimum supported release of `0.88.4`, so releases below `0.88.4` keep being offered `0.0.2`, which still has the server route it depends on.
+On Cloud the same rule now applies: the routes exist only when both `AP_TURNSTILE_SITE_KEY` and
+`AP_TURNSTILE_SECRET_KEY` are set. Without them the instance starts normally and simply does not offer emailed
+codes. Activepieces Cloud has both configured, so nothing changes for it — the code step stays on the sign-in card
+and accounts that use it keep signing in exactly as before.
#### What you need to do
-Nothing for existing connections; they keep working untouched, and sending no longer reads the dropped table.
-
-If you already have a flow using **Microsoft Teams Bot** piece version `0.0.2`, open the step and upgrade the piece to `0.1.0`. Flow steps pin an exact piece version, so an existing step stays on `0.0.2` after you upgrade, and `0.0.2` calls the `/v1/teams-bot/send` route this release removes. That step will fail with a 404 until it is upgraded.
-
-Optionally repoint the **Messaging endpoint** on your Azure Bot resource from `/api/v1/teams-bot/webhook` to `/api/v1/app-events/microsoft-teams-bot`; an endpoint left on the old path returns 404 and is otherwise harmless. If you call `/v1/teams-bot/send` directly from your own code, switch to the piece's **Send Channel Message as Bot** action.
-
-#### Table record filters compare date columns chronologically
+Nothing, if your users sign in with a password or with Google — those paths are unchanged, and so are `/sign-up`,
+`/sign-in` and `/switch-platform`.
-The `gt`, `gte`, `lt` and `lte` operators on `GET /v1/records` used to parse every cell value as a number. On a Date column that meant `2026-08-12T14:30:00Z` was read as `2026`, so two dates in the same year always compared equal and a range filter matched nothing. Those four operators now compare Date and Date & Time columns as instants.
+**If you ran 0.88.2–0.89.0 with SMTP configured, assume some accounts were created through the code flow.** Those
+accounts hold a random password nobody was ever shown, so removing the code flow removes the only way they could
+sign in. There is no reliable marker for them in the database — a verified code deletes its `otp` row — so treat any
+`EMAIL`-provider account whose owner cannot recall setting a password as one of them.
-This affects the Tables piece's Find Records action and any direct API call that filters a Date column with a range operator. `eq`, `neq` and `co` are unchanged and still compare the stored text exactly, so two spellings of the same instant (`...T14:30:00Z` and `...T14:30:00.000Z`) do not match each other.
+On **Enterprise**, send those users through **Forgot password** once to set a password; that flow needs the same SMTP
+you already have configured.
-#### What you need to do
+On **Community** there is no way back in the product. Password reset lives in the enterprise module, which Community
+does not register, so `POST /v1/authn/local/reset-password` is not served — and nothing else writes a password on any
+edition, as there is no change-password endpoint. Do not be misled by the first half of that flow appearing to work:
+`POST /v1/otp` **is** served on Community and answers `204`, but Community sends no OTP email of any type except a
+login code, so no reset mail ever arrives. Set a password for each affected account directly —
+`UPDATE user_identity SET password = '<bcrypt hash>' WHERE email = '…'` — and hand it to its owner over a channel you
+trust, since they cannot change it themselves afterwards. Do this **before** you upgrade if you can, so nobody is
+locked out in between.
-Nothing on upgrade. Re-check any flow or API integration that filters a Date column with `gt`, `gte`, `lt` or `lte` — it now returns the rows the filter actually describes, which may be more or fewer than before.
+If you script or test the sign-in page, note that the emailed-code step is gone and the two endpoints above
+answer `404`.
#### A failed run on a synchronous webhook answers with 500 instead of 408 after the full timeout
@@ -202,234 +253,229 @@ Two cases are deliberately unchanged. A run that succeeds without reaching a Ret
Nothing to configure or migrate. If you have a caller or reverse proxy that treats `408` as the signal for a failed synchronous flow, or retry logic keyed on `408`, switch it to `5xx` — a failed run no longer produces a `408`, and the reply now arrives in seconds rather than after the timeout.
-#### AWS connections using the IAM Role / OIDC method are now verified against AWS before they save
+## 0.89.0
-Saving an AWS connection (Bedrock, S3, Secrets Manager) with the **IAM Role / OIDC** method previously only checked that the Role ARN was well formed, so any well-formed ARN saved with status "Connected" even when the role did not exist or could not be assumed. The connection now performs the same `sts:AssumeRoleWithWebIdentity` call the flow makes at runtime, and the save fails with the AWS error when the role cannot be assumed.
+### What has changed?
-Two consequences. A connection that was accepted before is now rejected if its IAM OIDC provider, trust policy, or `AP_FRONTEND_URL` issuer is not set up correctly. And because the save now calls AWS, an instance with no outbound access to `sts.<region>.amazonaws.com` cannot create these connections at all.
+#### Piece builds fail when piece code uses `__dirname` without declaring `bundleForkedEntries`
-The region on these connections is also validated against the AWS region format at both save time and every flow step (the region flows into the STS endpoint hostname, so the check is enforced everywhere it is used). The STS call now times out after 10 seconds instead of waiting for the job timeout.
+The piece bundler now emits files a piece loads by path at runtime (for example a `child_process.fork` target) beside the main bundle, when they are declared in a `bundleForkedEntries` array in the piece's `package.json`, and it keeps dependencies that only those files import in the published manifest. Because an undeclared `__dirname`-relative file access always breaks after publishing — this is exactly how `@activepieces/piece-oracle-database` 0.1.11/0.1.12 shipped with every new connection failing — the build now fails loudly when a piece's source references `__dirname` and declares no forked entries. Previously such a piece built successfully and shipped broken.
#### What you need to do
-Only if you use the IAM Role / OIDC method on an AWS piece. Allow outbound HTTPS to `sts.<region>.amazonaws.com` from the machine running the worker, and make sure `AP_FRONTEND_URL` matches the Provider URL of your IAM OIDC identity provider, since the issuer in the signed token is derived from it. Existing saved connections keep working and are not revalidated automatically — pressing **Recheck** on a connection whose trust policy is wrong will now move it to an error state, which reflects what already happens when a flow runs.
+Nothing for catalog pieces — oracle-database is the only piece that forks a sibling file, and this change fixes it (0.1.13). If you build custom pieces and one references `__dirname`, either declare the runtime-loaded file in `bundleForkedEntries` (see [Bundling Pieces](/build-pieces/misc/bundling-pieces)) or remove the `__dirname` usage. Already-published pieces keep working; the check runs at build time only.
-#### Kimai connections now authenticate with an API token instead of a username and API password
+#### A new workspace is named after the company in the sign-up email
-Kimai is deprecating its legacy `X-AUTH-USER` / `X-AUTH-TOKEN` authentication (API passwords), with removal planned no later than July 2026 — see [Kimai's announcement](https://www.kimai.org/en/blog/2026/removing-api-passwords). The Kimai piece's connection now sends `Authorization: Bearer <token>` instead, matching Kimai's current API. This also fixes the Activity dropdown on the Create Timesheet action, which ignored the selected project because it tried to filter on a GET request body that the piece's HTTP client cannot send.
+Completing sign-up used to always name the new platform after the person, as `"Ahmad's Platform"`. It now reads the email domain first: `[email protected]` creates a platform called `Activepieces`, and the project alongside it becomes `Activepieces's Project`.
-The connection's `Username` and `API Password` fields are removed and replaced with a single `API Token` field.
+The person-based name is still what you get from a consumer address. `[email protected]` continues to produce `"Ahmad's Platform"`, as do the other common providers (Outlook, Yahoo, iCloud, Proton, GMX, QQ and similar). Whether an address counts as a work address is decided by a denylist of consumer providers, so anything not on that list is treated as a company.
-#### What you need to do
+This only affects platforms created from here on. Existing platforms keep their names, and renaming stays available in settings.
-Only if you have an existing Kimai connection. Reconnect it: in your Kimai instance, generate an API token from your user profile's "API Access" page, then re-enter your Kimai connection in Activepieces with the Server URL and that token. Existing connections using the old username/API-password fields will fail to authenticate until reconnected.
+#### What you need to do
-#### The AI provider config endpoint is removed
+Nothing. No configuration, no new environment variable, and nothing to migrate. If you script or test first-run sign-up and assert on the generated platform or project name, update that assertion: a work-domain address no longer yields `"<FirstName>'s Platform"`.
-`GET /v1/ai-providers/:provider/config` returned a provider's decrypted credentials so that an AI step running inside the engine could build the model itself and call the provider directly. AI steps no longer work that way: the worker resolves the credentials and makes the call, which is what lets Activepieces charge managed AI on what the call actually cost. Nothing in the product calls the endpoint any more, and it now returns `404`.
+#### Workers no longer pre-warm the flow cache on startup by default
-The route only ever accepted an engine token, so it was never reachable with a user token, an API key or from the browser. `GET /v1/ai-providers` and `GET /v1/ai-providers/:provider/models` are unchanged.
+Since 0.86.2 every worker pre-filled its local piece and code cache on startup by resolving and compiling every enabled flow on the platform. That warm-up costs memory and CPU proportional to the number of enabled flows: on instances with many flows it pinned each worker at its CPU limit for the duration and spiked memory enough to OOM-kill small workers, especially during upgrades when all workers restart at once. The warm-up is now opt-in behind the new `AP_PREWARM_CACHE_ON_STARTUP` worker environment variable, which defaults to `false`. When disabled, caches fill lazily on each flow's first run after a worker starts, exactly as they did before 0.86.2.
#### What you need to do
-Nothing on a normal upgrade. Every flow is moved to the AI piece version that runs on the worker the first time its version is read, before it can run.
+Nothing, unless you want to keep the warm-up behaviour of 0.86.2 through 0.88.4. Set `AP_PREWARM_CACHE_ON_STARTUP=true` on your worker containers to restore it — recommended only if your instance has a modest number of enabled flows and your workers have memory headroom. See [Environment Variables](/install/reference/environment-variables) for details.
-If you serve pieces from your own registry, make sure it carries the AI piece at `0.11.0` or later. A flow left on an earlier version would still try to fetch its provider credentials over this endpoint and fail.
+#### The Microsoft Teams Bot piece no longer needs a server-side installation record
-#### AI steps are billed on what each call costs, not on a fixed price per model
+`POST /v1/teams-bot/webhook` and `POST /v1/teams-bot/send` are removed, and the `teams_bot_installation` table is dropped. The webhook existed only to capture Microsoft's per-tenant `serviceUrl` at install time so the send endpoint could look it up.
-The managed Activepieces AI provider used to charge a fixed number of credits per model per step, taken from a table calibrated at roughly a thousand tokens per call. A short call was charged about right; a long one was charged the same as a short one while costing us many times more.
+The piece now posts to Microsoft's documented global Teams service endpoint directly, so it holds no server-side state and its messaging endpoint moves to the shared app-webhooks route, `/api/v1/app-events/microsoft-teams-bot`. Because the installed-or-not check is no longer a cached row, it is answered by Microsoft at send time: posting to a team the bot is not a member of now fails with `403 The bot is not part of the conversation roster` instead of a stale local lookup.
-Every managed AI call is now billed on the dollar cost the provider reports for that call, converted at `AP_AI_CREDIT_USD_VALUE` (default `0.0005`, so one credit is a twentieth of a cent of model spend). A call on your own API key is billed a flat one credit per model call. Tool calls stay at one credit each.
+The piece is published as `0.1.0` with a minimum supported release of `0.88.4`, so releases below `0.88.4` keep being offered `0.0.2`, which still has the server route it depends on.
-This changes what the same flow costs in both directions. A short prompt on an expensive model gets much cheaper; a long document through Extract Structured Data gets dearer. The credit cost shown in the model picker and in the credits breakdown is gone, because there is no longer a fixed number to show.
+#### What you need to do
-A step or a chat turn on your own API key costs the same as it did: one credit, plus one per tool call. An agent still counts as one credit however many times it goes back to the model.
+Nothing for existing connections; they keep working untouched, and sending no longer reads the dropped table.
-#### What you need to do
+If you already have a flow using **Microsoft Teams Bot** piece version `0.0.2`, open the step and upgrade the piece to `0.1.0`. Flow steps pin an exact piece version, so an existing step stays on `0.0.2` after you upgrade, and `0.0.2` calls the `/v1/teams-bot/send` route this release removes. That step will fail with a 404 until it is upgraded.
-Nothing to configure. Self-hosters who want a different conversion rate can set `AP_AI_CREDIT_USD_VALUE` to the dollar value of one credit.
+Optionally repoint the **Messaging endpoint** on your Azure Bot resource from `/api/v1/teams-bot/webhook` to `/api/v1/app-events/microsoft-teams-bot`; an endpoint left on the old path returns 404 and is otherwise harmless. If you call `/v1/teams-bot/send` directly from your own code, switch to the piece's **Send Channel Message as Bot** action.
-#### AI steps now consume credits on your own API key too
+#### Brevo's Create or Update Contact stops clearing blacklist flags and attributes it was not given
-`POST /v1/ai/execute` refuses to start an AI step when the platform is out of credits, whatever provider the step uses. Previously only the managed Activepieces provider was gated, so a platform that had exhausted its credits kept running AI steps as long as they used its own key.
+Four changes to the Brevo (formerly Sendinblue) piece's Create or Update Contact action alter what an existing, unchanged step does.
-Credits are only enforced where a billing provider is configured, so a Community self-host is unaffected.
+The action used to strip every falsy value from the request before sending it, so `Email Blacklisted` and `SMS Blacklisted` could be switched on but never switched back off — the `false` was dropped and the contact stayed blacklisted with no error. Those flags are now sent when you set them, and neither checkbox defaults to off any more, so a checkbox you never touched is omitted from the request instead of quietly un-blacklisting the contact on every run.
+
+The `Attributes` field used to come pre-filled with nine empty values (`FIRST_NAME`, `LAST_NAME`, `SMS`, `CIV`, `DOB`, `ADDRESS`, `ZIP_CODE`, `CITY`, `AREA`). Because they were sent on every call, running the step overwrote whatever those attributes held in Brevo with empty strings. The default is removed, and only the attributes you list are written.
+
+`List IDs` was a free-text list of numbers and is now a dropdown of the lists in your account. A stored numeric id is still sent exactly as before, so existing steps keep working unchanged.
+
+The `SMTP Blacklist Sender` checkbox is replaced by a `Blocked Sender Addresses` list. Brevo expects a list of sender email addresses here, not a true or false, and rejects a boolean outright with *"Invalid smtpBlacklistSender format"* — so ticking that checkbox always failed the step, and leaving it clear did nothing at all. The field never worked. It is now a list of addresses, which Brevo accepts, and any value your step already had is ignored rather than sent.
#### What you need to do
-Nothing, unless you run AI steps on your own provider key on a platform that sits at zero credits. Those steps now fail with a credits error instead of running. Top up, or raise the plan's credit allowance.
+Nothing to configure or migrate, and no existing step stops working. Review any flow whose Create or Update Contact step relied on the old behaviour: a step that was silently wiping attributes will now leave them intact, and a step you were relying on to blacklist contacts must have the checkbox explicitly ticked. To block a sender for a contact, list that sender's address in `Blocked Sender Addresses` — it must be an active sender in your Brevo account, or Brevo answers *"One of the sender is invalid or inactive"*.
-#### Files passed to an AI step are capped at `AP_MAX_FILE_SIZE_MB`
+#### A Run Agent step whose turn dies now fails its flow run
-Extract Structured Data and Generate Image used to read file inputs straight from the flow's memory, so no Activepieces-side size limit applied. Those files are now stored and fetched by the worker that runs the model, which puts them under the same per-file limit as every other file the engine writes: `AP_MAX_FILE_SIZE_MB`, 25 MB by default.
+A Run Agent step whose turn died — no AI provider configured for the chosen vendor, a revoked API key, a model the provider no longer serves — used to report success. The step returned the failed agent result rather than raising it, so the step read SUCCEEDED, the run completed SUCCEEDED, and later steps acted on a payload that carried the failure inside it. Such a run is now FAILED, and the agent's error message is the step's error.
-The limit is per file, not per step, so several large files in one step are fine as long as each is under it. On Activepieces Cloud the limit is 10 MB and is not configurable.
+Two cases are deliberately unaffected, because the agent did produce usable output:
+
+- A turn that finished but had a single tool call error. The agent may have recovered from it and reported on it, so those runs keep succeeding exactly as before.
+- A turn that was cut short by the output limit or the run's usage budget. It still returns the work it completed, and the flow still continues.
#### What you need to do
-Nothing unless an AI step handles files larger than 25 MB. Those steps now fail with a size error naming the limit. Raise `AP_MAX_FILE_SIZE_MB` on a self-hosted installation if you need more.
+Nothing to configure or migrate, and no new environment variable. Two things to check if either applies to you. If a flow relied on continuing past a dead agent step, tick **Continue on failure** on that step — it now takes effect, where previously the step never failed for it to apply to. If a flow branches on the agent result (a Router reading the step's `status`), that branch no longer runs when the turn dies, because the step raises instead of returning; move that handling onto the step's failure path.
-## 0.91.0
+If you alert on run status, expect agent flows that were silently completing on a misconfigured provider to start showing as failed. That is the misconfiguration surfacing, not new breakage.
-### What has changed?
+#### `AP_FLOW_TIMEOUT_SECONDS` now also bounds a single action run
-#### Telemetry moves from `AP_TELEMETRY_ENABLED` to the platform admin UI
+An action run — an agent's configured piece tool, a chat action, or MCP `ap_run_action` — used to be capped at a hardcoded 120 seconds, and a flow run started as an MCP tool at a hardcoded 300 seconds. Neither could be changed. Any agent tool whose action took longer than a minute failed outright with `RPC [executePieceTool] failed (timeout: 60000ms)`, because a second hardcoded limit in the worker cut it off first.
-Product analytics used to be governed only by the `AP_TELEMETRY_ENABLED` environment variable, read once when the app started. It is now a per-platform setting stored in the database, editable at **Platform → Infrastructure → Configurations**.
+Both hardcoded limits are gone. All of them now follow `AP_FLOW_TIMEOUT_SECONDS`, the same setting that bounds a flow run, which defaults to 600 seconds. An action run therefore gets the same budget as a flow step.
-Upgrading does not change what you collect. Each platform's setting is created the first time it is read, taking `AP_TELEMETRY_ENABLED` as its value, so an instance running with `AP_TELEMETRY_ENABLED=false` stays opted out — on an existing install and on a brand-new one alike.
+The limit going up is the change to be aware of: a slow or hung action now occupies a worker slot for up to `AP_FLOW_TIMEOUT_SECONDS` instead of 120 seconds, and action runs are not covered by the per-project concurrency limiter. On a small worker pool, several slow actions can hold execution capacity for longer than they used to.
-After a platform's setting exists, the environment variable no longer governs it. Editing the variable has no effect on that platform, and switching the toggle on in the UI turns product analytics on even if the variable still reads `false`.
+#### What you need to do
-The setting is per platform, so on an instance hosting several platforms each one is controlled separately. The page is hidden on Activepieces Cloud.
+Nothing to configure or migrate, and no new environment variable — a default installation gets the fix with no action. Two things to consider if they apply to you. If you raised `AP_FLOW_TIMEOUT_SECONDS` for long flows, note that agent tools, chat actions and MCP action runs now inherit that same ceiling; lower it if you do not want them held that long. If you put Activepieces behind a reverse proxy and use MCP `ap_run_action`, an action that runs past the proxy's read timeout (nginx defaults to 60 seconds, Cloudflare to 100) returns a gateway error to the MCP client while the action keeps running — raise `proxy_read_timeout` if you rely on long action runs over MCP.
-#### What you need to do
+#### Sign-up addresses are checked with ZeroBounce instead of a bundled blocklist
-Nothing, if you want to keep collecting what you collect today, and nothing if you rely on `AP_TELEMETRY_ENABLED=false` — that value is carried into each platform's setting as it is created. What changes is that the variable becomes a starting value rather than a permanent lock: once a platform has a setting, an admin can switch product analytics on from the UI regardless of what the variable says. If that matters to you, restrict who holds platform-admin rights.
+The bundled list of disposable-email domains introduced in 0.88.2 is gone. Sign-up now asks ZeroBounce about the address, on every password sign-up and on the emailed-code path only when the address has no account yet, and refuses it when ZeroBounce reports it as disposable, a spam trap, abusive, or on a global suppression list. The refusal is silent, so the response looks the same as a successful one. When ZeroBounce cannot be reached, or the key is invalid or out of credits, the address is let through and the reason is logged.
-## 0.88.2
+The check runs only when `AP_ZEROBOUNCE_API_KEY` is set. `AP_ALLOW_DISPOSABLE_EMAILS` is removed; the key is the switch. An instance without a key accepts throwaway addresses again, as it did before 0.88.2.
-### What has changed?
+#### What you need to do
-#### Sign-in and sign-up are one screen, and an emailed code signs people in wherever SMTP is configured
+Nothing, unless you relied on the 0.88.2 blocklist to keep throwaway addresses out. To keep refusing them, create a ZeroBounce account and set `AP_ZEROBOUNCE_API_KEY` on the app container. Remove `AP_ALLOW_DISPOSABLE_EMAILS` from your configuration if you set it; it is ignored. See [Environment Variables](/install/reference/environment-variables).
-`/sign-in` and `/sign-up` now render the same card, and `/sign-up` redirects to `/sign-in` carrying its query string, so existing links — invitation URLs included — keep working.
+## 0.88.2
-The card's primary path is a six-digit code emailed to the address typed into it. It is offered whenever the instance has SMTP configured and email auth is enabled, on every edition including Community, so a self-hosted instance that already sends transactional email gains a way to sign in that needs no password. An instance with no SMTP is unaffected: the card opens directly on the classic password form, with Google and SAML above it exactly as before.
+### What has changed?
-Two things change on an instance that does have SMTP. Password sign-in moves behind a **Use password** link on the card, and password sign-*up* is offered only where the card is in sign-up mode — the first account on a fresh instance, or an invitation link (`/sign-up?email=…`) — so someone who simply opens `/sign-in` on an instance that already has accounts is given the emailed code and no way to switch to a password sign-up form. Existing passwords are untouched: a verified identity keeps its password, and signing in with a code does not discard it.
+#### The container no longer bundles PM2 — crashes exit the container
-Who may get in does not change, and the existing guards all run before a code is sent. Joining still requires `AP_ALLOW_OPEN_SIGN_UP=true`, an accepted invitation, or existing membership, and that refusal is silent — the response is identical to an accepted one and no code is sent, so the endpoint reveals nothing about who has an account. A platform's allowed-domains list still applies too, but it answers with an explicit domain-not-allowed error rather than silently, and like the platform's Email-auth setting it is inert on Community and on any platform whose plan does not include SSO. On a fresh install, before the first platform exists, there is no platform for either check to run against and both are skipped.
+The Docker image used to run the app and worker under PM2, which restarted a crashed process *inside* the container. PM2 is removed; the entrypoint now launches the processes with plain `node`. When a process crashes (or is OOM-killed), the container exits instead of silently recycling the process in place, and your orchestrator restarts it.
#### What you need to do
-Nothing to configure or migrate, and no new environment variable: emailed-code sign-in turns itself on once SMTP is configured and Email is enabled as an authentication method.
+Nothing if you deploy with Docker Compose, Kubernetes/Helm, or any orchestrator that already restarts failed containers (all official deployments do). If you run the image with a bare `docker run` and relied on PM2 to keep the container alive across process crashes, add a restart policy: `docker run --restart unless-stopped ...`.
-If you do not want emailed-code sign-in, turn off **Email** as an authentication method for the platform (Enterprise and Cloud, under SSO settings, and only where your plan includes SSO); that disables password sign-in along with it, since both are the same method. On Community that toggle is inert, so the only lever there is leaving SMTP unconfigured.
+#### Table record filters compare date columns chronologically
-If you script or end-to-end test the sign-in screen, re-check it: the same URL now renders a different form depending on whether SMTP is configured and whether any account exists yet. Password reset moved inside the card wherever the emailed-code path is active; an instance without SMTP still sends people to `/forget-password`, and Community shows no reset control at all.
+The `gt`, `gte`, `lt` and `lte` operators on `GET /v1/records` used to parse every cell value as a number. On a Date column that meant `2026-08-12T14:30:00Z` was read as `2026`, so two dates in the same year always compared equal and a range filter matched nothing. Those four operators now compare Date and Date & Time columns as instants.
-#### Emailed sign-in codes allow ten wrong guesses per account per hour
+This affects the Tables piece's Find Records action and any direct API call that filters a Date column with a range operator. `eq`, `neq` and `co` are unchanged and still compare the stored text exactly, so two spellings of the same instant (`...T14:30:00Z` and `...T14:30:00.000Z`) do not match each other.
-The six-digit login code already allowed five wrong guesses, but that budget lived on the code itself, and the fifth wrong guess threw the code away — so asking for a new code handed out five fresh guesses immediately, with no ceiling on how often that could repeat. A six-digit code is only a million possibilities, so unlimited retries reduce it to a matter of hours.
+#### What you need to do
-Wrong guesses are now counted per account over a rolling hour, independently of how many codes get sent. Entering the right code clears the counter, so someone who fumbles a few digits and then succeeds is unaffected.
+Nothing on upgrade. Re-check any flow or API integration that filters a Date column with `gt`, `gte`, `lt` or `lte` — it now returns the rows the filter actually describes, which may be more or fewer than before.
-Two consequences worth knowing. Someone who spends ten wrong guesses on an account within an hour cannot sign in **with an emailed code** until the hour is up; password and Google sign-in are unaffected. And because the counter is keyed on the account rather than the caller, anyone who knows an address can spend that budget on the owner's behalf — a temporary nuisance for the owner, and the trade the cap is worth making.
+#### Official piece bundles are served from npm again by default
+
+`AP_USE_CDN_FOR_BUNDLES` goes back to defaulting to `false`, reversing the change in 0.87.0. Official piece bundles are downloaded from `registry.npmjs.org` instead of `cdn.activepieces.com`. When the flag is switched on, the CDN serves the repackaged, self-contained tarballs from `https://cdn.activepieces.com/pieces/bundled/`. The same release stops mirroring bundles into your own S3 bucket: a bundle that was copied there is no longer served from it, and every download goes to npm or the CDN.
#### What you need to do
-Nothing. The limit applies out of the box and needs no configuration. If your support team sees a report of "the code keeps saying it's wrong", have them check whether the account has burned its hourly budget, and point the user at password or Google sign-in in the meantime.
+If your network policy restricts outbound traffic, allow `registry.npmjs.org`: it is the default source, and it stays the fallback for any bundle the CDN does not have even when `AP_USE_CDN_FOR_BUNDLES=true`. Set the flag to `true` to prefer the CDN, and allow `cdn.activepieces.com` as well in that case. Otherwise nothing.
-#### One-time codes are no longer stored in a readable form
+#### MCP OAuth: revoking a token requires a client identity, and registration issues a usable secret
-The `otp` table used to hold the six-digit sign-in code as plain text, so anyone who could read the database — a replica, a backup, a support query — could sign in as any account for the ten minutes that code was alive, without a password. Codes are now stored as a digest keyed with a server-held secret, so reading the table no longer yields anything you can sign in with. The code itself is held only long enough to re-send it if the person asks for it again.
+Two changes to the MCP OAuth endpoints:
-A code found to be expired is thrown away when it is next presented, rather than sitting in the table until the same person happens to request another one.
+- `POST /revoke` used to revoke a refresh token for a caller that presented no client identity at all, so anyone who learned a refresh token could disconnect its owner. It now requires the request to identify the client, and a confidential client to authenticate, per RFC 7009. A revocation that sends only `token` now returns `400 invalid_client`.
+- Dynamic client registration (`POST /register`) that omits `token_endpoint_auth_method` now records the client as `client_secret_basic` and issues a client secret, per RFC 7591. It previously issued a secret but recorded the client as public, so that secret was never checked. A client registered from now on must present it when calling `/token` and `/revoke`. Concretely: a client that omits `token_endpoint_auth_method` **and** discards the `client_secret` it is handed now gets `400 invalid_client` at `/token`, where it previously succeeded. Such a client must either store and send that secret, or register explicitly with `token_endpoint_auth_method: "none"` to stay public.
#### What you need to do
-Nothing, and no new configuration: the key is derived from a secret your instance already has. Codes written by an older build are recorded as such and keep working until they expire, so a sign-in already underway when you deploy still completes, an email verification link still opens, and a rolling deploy where both builds are serving at once behaves the same.
-
-Rolling back costs at most the codes issued after the deploy: the older build cannot read those, so whoever holds one asks for a fresh code. Nothing is rewritten and nothing is deleted, so no cleanup is needed either way.
+Nothing for already-connected clients. A client's authentication method is read from the row it registered with, so existing registrations are untouched, including the ones created before this fix that hold a secret while recorded as public. Those keep working whether or not they send it.
-#### WhatsScale media actions take a file input, capping media at the instance file-size limit
+Claude, Cursor and anything else built on the MCP SDK are unaffected: the SDK stores and sends the `client_secret` it is given, and it never calls the revocation endpoint. A confidential client's secret is accepted from either the request body or the `Authorization: Basic` header, so a client is never rejected for choosing one over the other.
-The 17 WhatsScale actions that send media — image, video, document and audio sends across the contact, group, channel and CRM-contact variants, plus the image and video Status posts — used to take a plain text URL and hand it straight to WhatsScale, which fetched it itself. Those fields are now file inputs: the engine resolves whatever you pass (a URL, or a file from a previous step) into a file, stores it on your instance, and gives WhatsScale that copy to fetch.
+If you maintain your own MCP OAuth client, check two things before upgrading: that it sends `client_id` when it revokes a token, and that it stores and presents the `client_secret` that `/register` returns. Note that disconnecting and re-adding an MCP server creates a fresh registration, so an existing user meets the new behaviour when they reconnect rather than at upgrade.
-The field keys are unchanged (`imageUrl`, `videoUrl`, `documentUrl`, `audioUrl`), so existing flows keep their configured value and a stored URL still works. Two behaviour changes come with the new path:
+#### Disposable email addresses can no longer sign up
-- **Media is now capped by `AP_MAX_FILE_SIZE_MB` — 10 MB on Cloud, 25 MB by default self-hosted.** The file travels through your instance instead of going vendor-to-vendor, so a URL pointing at something larger than that limit now fails where it previously sent. This mostly affects video and document sends. Streaming does not lift the ceiling: writing into storage is counted against the cap either way, and producing a URL for WhatsScale to fetch requires writing into storage. See [Limits](/install/reference/limits).
-- **The URL WhatsScale fetches is now one your instance serves.** It is a signed, expiring link on your public API URL. An instance that is not reachable from the internet cannot hand WhatsScale a URL it can fetch, so media sends need a publicly resolvable `AP_FRONTEND_URL` on self-hosted installs.
+Sign-up now refuses addresses from throwaway providers such as `mailinator.com` and `guerrillamail.com`, on both the emailed-code flow and the password form. Federated sign-in (Google, SAML, JWT), managed authentication and SCIM provisioning are unaffected, since those addresses come from an identity provider you already trust.
#### What you need to do
-Nothing if your media is comfortably under the cap and your instance is reachable from the internet. On Cloud that cap is 10 MB, so check any flow that forwards video or large documents.
+Only if your members sign up with addresses from a disposable email provider. On 0.88.2 through 0.88.4, set `AP_ALLOW_DISPOSABLE_EMAILS=true` to keep accepting them.
-If you send files larger than the cap on a self-hosted install, raise `AP_MAX_FILE_SIZE_MB` to cover the largest media you send and redeploy. On Cloud the 10 MB cap is not configurable. If your instance is not publicly reachable, keep media sends on a host that is, or leave these actions unused; there is no way for WhatsScale to fetch a file from a private address.
+<Note>
+Superseded in 0.89.0: the bundled blocklist was replaced by a ZeroBounce check that runs only when `AP_ZEROBOUNCE_API_KEY` is set, and `AP_ALLOW_DISPOSABLE_EMAILS` is no longer read. See the 0.89.0 entry.
+</Note>
-#### WhatsScale "Add a Tag to a CRM Contact" now takes a list of tags
+#### Sign-in and sign-up are one screen, and an emailed code signs people in wherever SMTP is configured
-The two add-tag actions ("Add Tags to a CRM Contact" and its By-ID twin) accepted a single tag in a text field named `tag`. They now accept a list, in a field named `tags`, and send one request per tag so existing tags are kept.
+`/sign-in` and `/sign-up` now render the same card, and `/sign-up` redirects to `/sign-in` carrying its query string, so existing links — invitation URLs included — keep working.
-Because the field was renamed, **a step configured before this release loses its tag value** and shows "Tags" as an empty required field. The field could not keep its old name: a stored single string does not validate as a list.
+The card's primary path is a six-digit code emailed to the address typed into it. It is offered whenever the instance has SMTP configured and email auth is enabled, on every edition including Community, so a self-hosted instance that already sends transactional email gains a way to sign in that needs no password. An instance with no SMTP is unaffected: the card opens directly on the classic password form, with Google and SAML above it exactly as before.
-Adding several tags is **not atomic** — each tag is its own request, so a failure part-way through leaves the earlier tags applied. Re-running is safe: adding a tag a contact already has converges on the same set.
+Two things change on an instance that does have SMTP. Password sign-in moves behind a **Use password** link on the card, and password sign-*up* is offered only where the card is in sign-up mode — the first account on a fresh instance, or an invitation link (`/sign-up?email=…`) — so someone who simply opens `/sign-in` on an instance that already has accounts is given the emailed code and no way to switch to a password sign-up form. Existing passwords are untouched: a verified identity keeps its password, and signing in with a code does not discard it.
+
+Who may get in does not change, and the existing guards all run before a code is sent. Joining still requires `AP_ALLOW_OPEN_SIGN_UP=true`, an accepted invitation, or existing membership, and that refusal is silent — the response is identical to an accepted one and no code is sent, so the endpoint reveals nothing about who has an account. A platform's allowed-domains list still applies too, but it answers with an explicit domain-not-allowed error rather than silently, and like the platform's Email-auth setting it is inert on Community and on any platform whose plan does not include SSO. On a fresh install, before the first platform exists, there is no platform for either check to run against and both are skipped.
#### What you need to do
-Re-enter the tag on any flow that uses either add-tag action, as a list entry rather than free text. The removal actions are unchanged and still take one tag at a time.
+Nothing to configure or migrate, and no new environment variable: emailed-code sign-in turns itself on once SMTP is configured and Email is enabled as an authentication method.
-#### Front's Attachments field takes files instead of URLs
+If you do not want emailed-code sign-in, turn off **Email** as an authentication method for the platform (Enterprise and Cloud, under SSO settings, and only where your plan includes SSO); that disables password sign-in along with it, since both are the same method. On Community that toggle is inert, so the only lever there is leaving SMTP unconfigured.
-The Front piece's Send Message, Send Reply, Create Draft and Create Draft Reply actions described `Attachments` as a list of attachment URLs and sent those strings inside a JSON body. Front only accepts attachments on a `multipart/form-data` request, as the bytes of the file, so it took the message and dropped the attachments — HTTP 202, no error, no attachment. The field is now a list of files, the same shape the Gmail and Discord pieces use, and the request is sent as multipart when a message carries attachments.
+If you script or end-to-end test the sign-in screen, re-check it: the same URL now renders a different form depending on whether SMTP is configured and whether any account exists yet. Password reset moved inside the card wherever the emailed-code path is active; an instance without SMTP still sends people to `/forget-password`, and Community shows no reset control at all.
-A URL is still all you need to supply: the engine downloads it and hands the piece the resolved file. A step that already holds a file — an earlier download, a trigger's attachment — can be wired straight in. An entry whose file cannot be resolved is skipped rather than failing the send.
+#### Emailed sign-in codes allow ten wrong guesses per account per hour
+
+The six-digit login code already allowed five wrong guesses, but that budget lived on the code itself, and the fifth wrong guess threw the code away — so asking for a new code handed out five fresh guesses immediately, with no ceiling on how often that could repeat. A six-digit code is only a million possibilities, so unlimited retries reduce it to a matter of hours.
+
+Wrong guesses are now counted per account over a rolling hour, independently of how many codes get sent. Entering the right code clears the counter, so someone who fumbles a few digits and then succeeds is unaffected.
+
+Two consequences worth knowing. Someone who spends ten wrong guesses on an account within an hour cannot sign in **with an emailed code** until the hour is up; password and Google sign-in are unaffected. And because the counter is keyed on the account rather than the caller, anyone who knows an address can spend that budget on the owner's behalf — a temporary nuisance for the owner, and the trade the cap is worth making.
#### What you need to do
-Re-pick the attachment in any Front step that used the field. The stored value is a plain string and the field now holds a file, so it does not carry across. A flow that never set Attachments is unaffected and keeps its JSON request path. Nothing to configure on the server, and no new environment variable.
+Nothing. The limit applies out of the box and needs no configuration. If your support team sees a report of "the code keeps saying it's wrong", have them check whether the account has burned its hourly budget, and point the user at password or Google sign-in in the meantime.
-#### A file URL that fails to download is no longer treated as a file
+#### One-time codes are no longer stored in a readable form
-Any piece property that takes a file also accepts a URL, and the engine buffered whatever came back from that URL without checking the response status. A 4xx or 5xx produced a file named after the URL whose contents were the server's error page, and a piece could not tell that apart from the real thing — an expired signed link became an email with the storage provider's XML error attached. A failed download now fails the step, which is what the streaming path already did.
+The `otp` table used to hold the six-digit sign-in code as plain text, so anyone who could read the database — a replica, a backup, a support query — could sign in as any account for the ten minutes that code was alive, without a password. Codes are now stored as a digest keyed with a server-held secret, so reading the table no longer yields anything you can sign in with. The code itself is held only long enough to re-send it if the person asks for it again.
-#### What you need to do
+A code found to be expired is thrown away when it is next presented, rather than sitting in the table until the same person happens to request another one.
-Nothing to configure. A flow that was silently passing on error pages will start failing at that step instead, and the fix is the URL itself — usually a link that has expired or that the instance is not authorized to read.
+#### What you need to do
-#### Signing in with an emailed code is now offered on Activepieces Cloud only
+Nothing, and no new configuration: the key is derived from a secret your instance already has. Codes written by an older build are recorded as such and keep working until they expire, so a sign-in already underway when you deploy still completes, an email verification link still opens, and a rolling deploy where both builds are serving at once behaves the same.
-`POST /v1/authentication/otp/request` and `POST /v1/authentication/otp/verify` are no longer served on a self-hosted
-instance. They return `404`, and the sign-in card no longer offers the "email me a code" step — it opens on the
-password form instead. Versions 0.88.2 through 0.88.4 did serve them on any edition with SMTP configured.
-`POST /v1/authentication/complete-sign-up` is unaffected and still served everywhere.
+Rolling back costs at most the codes issued after the deploy: the older build cannot read those, so whoever holds one asks for a fresh code. Nothing is rewritten and nothing is deleted, so no cleanup is needed either way.
-A six-digit code is a million possibilities, so what keeps guessing it expensive is the captcha in front of the
-request endpoint. Cloudflare Turnstile is opt-in and unset out of the box, so a self-hosted instance served both
-endpoints with nothing in front of them. Rather than require a Cloudflare account to make a sign-in method safe,
-that method is no longer served where the captcha is not there.
+## 0.88.0
-On Cloud the same rule now applies: the routes exist only when both `AP_TURNSTILE_SITE_KEY` and
-`AP_TURNSTILE_SECRET_KEY` are set. Without them the instance starts normally and simply does not offer emailed
-codes. Activepieces Cloud has both configured, so nothing changes for it — the code step stays on the sign-in card
-and accounts that use it keep signing in exactly as before.
+### What has changed?
-#### What you need to do
+#### Agent steps move from the flow engine to the server
-Nothing, if your users sign in with a password or with Google — those paths are unchanged, and so are `/sign-up`,
-`/sign-in` and `/switch-platform`.
+An agent step used to run its loop inside the engine, holding a worker for as long as the agent took. It now pauses the flow, runs on the server, and resumes when the agent finishes. This upgrade rewrites every existing agent step to the new version.
-**If you ran 0.88.2–0.88.4 with SMTP configured, assume some accounts were created through the code flow.** Those
-accounts hold a random password nobody was ever shown, so removing the code flow removes the only way they could
-sign in. There is no reliable marker for them in the database — a verified code deletes its `otp` row — so treat any
-`EMAIL`-provider account whose owner cannot recall setting a password as one of them.
+The step's inputs and its output shape are unchanged, so anything reading the step's result or its structured output keeps working. Sub-flow, MCP server, knowledge base and piece action tools all run on the new path.
-On **Enterprise**, send those users through **Forgot password** once to set a password; that flow needs the same SMTP
-you already have configured.
+Two differences are worth knowing before you upgrade:
-On **Community** there is no way back in the product. Password reset lives in the enterprise module, which Community
-does not register, so `POST /v1/authn/local/reset-password` is not served — and nothing else writes a password on any
-edition, as there is no change-password endpoint. Do not be misled by the first half of that flow appearing to work:
-`POST /v1/otp` **is** served on Community and answers `204`, but Community sends no OTP email of any type except a
-login code, so no reset mail ever arrives. Set a password for each affected account directly —
-`UPDATE user_identity SET password = '<bcrypt hash>' WHERE email = '…'` — and hand it to its owner over a channel you
-trust, since they cannot change it themselves afterwards. Do this **before** you upgrade if you can, so nobody is
-locked out in between.
+- **Web search is no longer a step setting.** It follows what the platform has configured rather than a per-step toggle. A step that had it switched on keeps working; the stored value is ignored.
+- **Max steps is honoured up to a ceiling.** The value you set still applies, but the server caps a single run at 50 turns. A step configured above that runs 50.
-If you script or test the sign-in page, note that the emailed-code step is gone and the two endpoints above
-answer `404`.
+#### What you need to do
+Nothing on upgrade. Check any agent step that relied on the per-step web search toggle, or that set max steps above 50, since those now follow the platform and the server ceiling.
## 0.87.0
### What has changed?
-#### Disposable email addresses can no longer sign up
-
-Sign-up now refuses addresses from throwaway providers such as `mailinator.com` and `guerrillamail.com`, on both the emailed-code flow and the password form. Federated sign-in (Google, SAML, JWT), managed authentication and SCIM provisioning are unaffected, since those addresses come from an identity provider you already trust.
-
#### Plans and credits are now managed by our billing service
Two license-key endpoints are removed: `GET /v1/license-keys/:licenseKey` and `POST /v1/license-keys/verify`. Applying a key is now `POST /v1/platform-billing/activate`.
@@ -446,18 +492,22 @@ When a platform has no credits left, new production runs are recorded with the `
In the AI piece, picking the Activepieces-provided AI now lists only the named tiers (Fast, Expert, Heavy) rather than the full upstream model catalogue. Existing steps keep running on the model they already have, but that model no longer appears in the dropdown, so re-saving the step moves it onto one of the tiers.
### Do you need to take action?
-- Only if your members sign up with addresses from a disposable email provider. Set `AP_ALLOW_DISPOSABLE_EMAILS=true` to keep accepting them.
- Only if you run the Enterprise edition without a license key. Enter your key so your contracted limits apply instead of the free plan's.
- Only if you call `GET /v1/license-keys/:licenseKey` or `POST /v1/license-keys/verify`. Both are removed — use `POST /v1/platform-billing/activate` instead.
- Only if you read `SHOW_BILLING_PAGE`, `CAN_BUY_ACTIVE_FLOWS`, `CAN_BUY_AI_CREDITS` or `SHOW_BILLING_LIMITS_ON_SIDEBAR` from `GET /v1/flags`. They no longer exist.
- Only if a flow uses Activepieces-provided AI on a model outside the Fast/Expert/Heavy tiers. Choose a tier the next time you edit that step.
- No database action is required. The migration only adds columns and leaves the existing ones in place.
+
#### Official piece bundles are served from the Activepieces CDN by default
`AP_USE_CDN_FOR_BUNDLES` now defaults to `true`. When an official piece bundle is not already cached in your own S3 bucket, the engine is redirected to `cdn.activepieces.com` for the tarball instead of `registry.npmjs.org`. If the CDN does not have that piece version, the download falls back to npm automatically, so installs keep working either way. Custom pieces and pieces from a private registry never use the CDN and are unaffected.
### Do you need to take action?
- Only if your network policy restricts outbound traffic. Allow `cdn.activepieces.com`, or set `AP_USE_CDN_FOR_BUNDLES=false` to keep serving official bundles from npm.
+<Note>
+Reverted in 0.88.2: `AP_USE_CDN_FOR_BUNDLES` defaults to `false` again, so official bundles are served from npm by default. Set it to `true` to opt back into the CDN. The S3 cache mentioned above was removed in the same release.
+</Note>
+
## 0.86.0
### What has changed?
@@ -476,6 +526,10 @@ The `AP_PRE_WARM_CACHE` environment variable no longer exists. It used to make a
The up-front install is no longer needed because piece installs are now fast: each piece is a self-contained bundle fetched as a single file. If you have S3 enabled, official piece bundles are cached to your S3 bucket on first use and served from there via signed links — so when your bucket is co-located in the same region as your workers, subsequent installs are a low-latency download from a nearby object store instead of a registry round-trip.
+<Note>
+Removed in 0.88.2: bundles are no longer mirrored into your S3 bucket. Every download goes to npm or, with `AP_USE_CDN_FOR_BUNDLES=true`, to the CDN.
+</Note>
+
### Do you need to take action?
- Only if you set up Google SSO from the admin panel. Move your Google OAuth2 credentials into the `AP_GOOGLE_CLIENT_ID` and `AP_GOOGLE_CLIENT_SECRET` environment variables.
- Only if you build dashboards or alerts on Activepieces logs. Update them to the new grouped names, e.g. `runId` → `flowRun.id` and `err` → `error`.