Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(changelog): regenerate the published feeds and finish the AI brain tutorial corrections
nirholase647194049c3
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 63cb88293..eae12da90 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -9,21 +9,26 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr- **A broken paid endpoint is no longer reported as a free one**: Payment sessions let you hand an agent a budget instead of a key, and two parts of that path judged an endpoint by one rule: if it did not ask for payment, the call was free. That is only true when the endpoint actually succeeded. A paid API answering 404, 400, or 500 was being scored as a free call, and the consequences ran in the wrong direction in both places. The dry run at /pay/simulator, whose whole job is to tell you in advance whether a budget and an allowlist will carry a workload, reported a plan as feasible at a cost of nothing when in fact every endpoint in it was erroring: free calls are never priced, never allowlist-checked, and never counted, so a wall of failures looked like a clean run. And an agent spending through a session got back a success envelope saying no payment was needed for an endpoint that had just failed on it. Both now separate the two cases: a real success with no payment challenge is still free and still costs nothing, while an error status is reported as an endpoint failure carrying the status the endpoint actually returned. No session is charged either way. The simulator lists each failing endpoint by name in its problems list and no longer calls the plan feasible. (`/pay/simulator`) `[fix]`- **A silent outage in the agent reasoning ledger now reports itself**: Every agent decision is committed to a hash chain, and that chain head is written to Solana on a schedule so anyone can check the history was never edited. The wallet that pays for those on-chain writes ran empty, and the scheduled job could not tell that apart from the harmless case where no signing key is set up at all: it answered a healthy 200 every ten minutes while re-sending the same doomed transaction for twelve agents in a row, and not one head reached the chain. A rejected write is now reported as a rejected write, the rest of the run stops instead of repeating a failure that cannot succeed, and it raises an alert so the wallet gets refilled. Verifying a ledger already worked without the on-chain leg and still does: the recomputed hash chain is the tamper proof, the on-chain write is the independent timestamp on top of it. Separately, the once-a-minute economy heartbeat now refuses to drive production from a machine that is not production, so a developer running it locally can no longer fire thirty-six live money engines at the real site. (`/docs/reasoning-ledger`) `[fix, infra]`- **Arming an Oracle agent for live trading no longer rounds your trade size up**: The arm page clamps the risk knobs on the server so a bad number can never reach the trading loop, and those clamps have a floor: the smallest trade the loop will place is 0.001 SOL. That floor was being applied in the wrong direction. Ask to arm a live agent with a per-trade size of zero, and instead of being told no, you were armed for real spending at 0.001 SOL a trade with a half-SOL daily ceiling you never asked for. The check meant to catch this looked at the number after the clamp had already raised it, so it could never fire. Arming live now refuses any per-trade size below the floor, and any daily cap smaller than a single trade, and tells you which number to fix; nothing is armed until you do. Clamping a too-large number back down to the ceiling is unchanged, and simulate runs keep the same forgiving behaviour as before. The arm page also shows the reason a save was rejected instead of one generic line, so an ownership problem no longer reads the same as a bad trade size, and a failed Telegram test now prints what Telegram actually said. (`/oracle/arm`) `[fix, security]`+- **Avatars can move their lips again: the Audio2Face lane is back**: Every call that turns speech into facial animation had been failing with an upstream not-found error, so a talking avatar spoke with a frozen face and the one-shot say() path returned nothing usable. The hosted model id the lane pointed at had stopped resolving for our NVIDIA account. It now points at the hosted Audio2Face NIM that is actually live, verified end to end: a synthesized clip drives a 30 fps ARKit blendshape track whose jaw genuinely opens and closes with the words. The discovery helper that finds these model ids was missing the name NVIDIA publishes them under, which is why the outage looked like a configuration mystery; it now finds them. `[fix]`- **Docs pages stop opening with a wall of raw metadata, and the AWS pages say where the Marketplace listing actually stands**: Nine documentation pages are written as article drafts, so each one starts with a small block of publishing metadata: the title, where it is meant to run, its tags. That block is meant for the tooling, not for you, but the docs reader had no idea what it was and rendered it as an enormous heading at the very top of the article, above the actual writing. It is now recognised and hidden, on the page and in the Copy page text, so those articles begin where they should. Reading the same pages closely turned up claims that had drifted: several said three.ws was already live on AWS Marketplace and billing usage to your AWS invoice. The integration is built and deployed, but the listing itself has not been published yet and the pricing decision is a free subscription with no AWS metering, so the AWS pages now say exactly that and point at the page that tracks the status. The MCP article's vanity-address tool was also listed at a metered price of up to fifty cents; it is a flat five cents per grind, which is what the code has always charged. (`/docs/aws-marketplace`) `[docs, fix]`- **Oracle stops answering "nothing is happening" when the database is the thing that is down**: Several Oracle reads shared a habit: if the database was unreachable for a few seconds, they answered normally with an empty result. The movers panel said nothing was moving, the stats hero said no coin had ever been scored and no agent was armed, a coin's conviction sparkline came back flat, and the machine-readable signal endpoint told autonomous agents there were no plays, which is exactly the cue for an agent to stand down. Every one of those answers was then cached at the edge for a minute or two, so a blip lasting seconds kept showing a dead market long after the database was back. Those endpoints now tell the difference between a connectivity failure and an ordinary query fault: the first is reported honestly as a temporary outage with a retry hint and is never cached, the second still degrades to the empty answer it always did. The social ingest got the sharper end of the same bug: it used to reply that a batch had been accepted while reporting zero mints written, a receipt for a write that never landed. It now fails loudly so the sender retries. Shared conviction cards behave too: one built during an outage is no longer pinned to a coin's link preview for 15 minutes. (`/oracle`) `[fix, improvement]`- **Plan checkout: a paid upgrade can only ever settle on real mainnet, and a late payment is no longer stranded**: Paying for a plan creates a quote, you send the funds, then the server checks the chain before it upgrades you. Three things in that path were wrong. The checkout side has always refused test-network quotes in production, because test-network dollars are free from a faucet, but the confirm side still let a caller name the test network itself, and it is the confirm side that decides which chain it looks at and which dollar it accepts. Both ends now read the same setting, so a production upgrade can only ever be settled by real funds on mainnet. Second, an Ethereum-side payment has to wait twelve blocks before it counts, which meant funds sent near the end of a session could land on-chain and then be refused for arriving late, with the money already gone; that lane now honors the same one-hour grace window the Solana lane has, so an on-chain payment is always claimable. Third, confirming an Ethereum payment could reach a Solana quote and mark it expired, and asking to buy a skill from a malformed agent id answered a generic server error instead of telling you what was wrong with the request. (`/pricing`) `[fix, security]`- **Seeding an agent from GitHub is locked to your own click, and a mistyped pick no longer costs you six hours**: Three fixes to the GitHub memory lane, the one that reads the profile, repositories, and READMEs you tick and distills them into what your agent knows about your work. First: writing or deleting those memories now requires the one-time token the site attaches to your own actions, so another site cannot make your browser re-seed or wipe an agent behind your back. Every neighbouring memory action already worked this way; this one did not. Second: seeding is allowed once every six hours per agent, and that window used to be spent by attempts that never read anything, so ticking nothing and pressing Seed, or picking a repository that had been renamed since the list loaded, locked you out for six hours over a correctable mistake. The window is now taken only once your selection checks out, immediately before anything is actually read. Third: when every model provider is busy at once, the seed used to fail with a bare internal error and a support reference; it now says plainly that nothing was read, that your existing memories are untouched, and roughly how long until you can try again. (`/docs/github-memory-seeding`) `[security, fix]`- **Self-hosting docs no longer hand you a command that wipes your environment**: The deployment guide told self-hosters to set Cloud Run environment variables with `gcloud run services update --set-env-vars`. That flag replaces the service environment wholesale, so following the instruction for a single key silently deletes every other variable the service had. The guide now uses `--update-env-vars` and says plainly why. Four more corrections landed in the same sweep across the deployment, configuration and contributing pages: the Avaturn signup link and the GitHub Discussions links were dead (Discussions are not enabled on the repository, so bug reports and design questions both go to Issues); the in-browser avatar builder was documented as needing its own hosted origin in production when the normal build already serves it same-origin at /avatar-studio; and three counts that had drifted from the code (cron jobs, MCP server packages, contract tests) were replaced with pointers to the files that hold the real numbers. (`/docs/deployment`) `[docs, fix]`+- **The 3D pipeline docs stop telling you to run a command that does not exist**: Every paid stage on the 3D asset pipeline page (rig, remesh, game-ready, stylize, background removal) showed you how to pay for it by running npx @x402/fetch. There is no such command: that package publishes a library, not a CLI, so anyone who copied a sample got a could-not-determine-executable error and no asset. The samples are now the real thing: an unpaid curl that returns the live 402 challenge and costs nothing, plus one Node snippet that wraps fetch with our published @three-ws/x402-fetch buyer and actually settles. The end-to-end generate, rig and game-ready walkthrough was rewritten the same way and now polls the forge job with the fields the API really returns. The page also stops promising a specific settlement network, because the 402 challenge names the networks itself. Alongside it: the changelog page CDN pinning table listed four alias URLs that all 404 and pointed at an integrity file for a version that no longer publishes one; the AgenC page opening sentence linked the word AgenC to a different protocol page; and the vanity protocol spec showed a stale copy of the service key document, which now matches what the live endpoint serves. (`/docs/3d-pipeline`) `[docs, fix]`- **The Agent Economy demo works again when you just pick a service, and it fits a phone screen**: On the Agent Economy page, two 3D agents trade with each other for real SOL: Nova buys a market read, an on-chain insight, or a risk score from Oracle. Picking a service and typing nothing in the optional topic box, which is what almost everyone does, sent the request off in a shape the endpoint refused, so the demo answered with a validation error instead of a trade. It now sends the request correctly, and the endpoint accepts an empty topic besides. On a phone the page was worse than broken: the transaction feed and the status line, the entire output of the demo, were cut off the bottom of the screen with no way to scroll to them, and the floating language button sat on top of the footer links. The centre column now scrolls, the feed keeps a readable height, and the footer stays clickable. Three failure states also stopped lying: a payment that could not go through no longer ends on Transaction complete, a request that never reached the server now says so and offers to try again, and unreadable wallet balances now say they are unavailable and offer a retry instead of spinning forever. (`/agent-economy`) `[fix, improvement]`- **The agent economy volume dashboard stops sending you to profiles that are not there**: Top earners and top spenders on the agent-to-agent volume page linked every ranked agent to a profile, including agents whose identity had been deleted. Those rows sent you to an empty profile shell for an agent that no longer exists. Both leaderboards now name a departed agent without linking it, the same rule the settled-hire feed below them already followed, and a row whose stored avatar has been pruned falls back to its initials tile instead of a broken-image icon. The page also cleared its accessibility floor: the small print under each stat, the footnote, and the retry button were all below the WCAG AA contrast minimum in dark mode, and in light mode the cyan links and the green live badge were effectively unreadable. Links inside the empty-state sentences now carry a standing underline, so they are not marked by colour alone. (`/agent-economy-volume`) `[fix, improvement]`- **The agent-payment and MCP docs now show the code that actually ships**: Three developer docs had drifted from the code they document, in ways that cost a reader real time. The Payment Sessions walkthrough, which is the guide for giving an agent a budget instead of a wallet key, quoted five snippets that no longer match the source: the branch that refuses a payment when the budget will not cover it, the host-allowlist check (which has moved into a shared module so the dry-run simulator and the live enforcement can never disagree), the expiry sweep that refunds unspent budget, the dollars-to-atomic-units helper, and a column missing from the table definition. All five now match the shipped code line for line, and the page links the two source files it had not been pointing at. The MCP article listed the Solana vanity grinder at up to $0.50 scaled by the work done; it is a flat $0.05 per successful grind, and its setup example set a wallet variable the package does not read, so anyone copying it would have configured a payout address that was silently ignored. That example now shows only the variable that works, and says plainly that settlement is USDC on Solana mainnet. The IBM Granite forecasting example showed three data points for a tool that requires at least sixty-four, so a copied sample was guaranteed to be rejected; it now says so. (`/docs/aws-builder-center-agent-payment-sessions`) `[docs, fix]`- **The Alibaba Cloud MCP server starts now, and five MCP packages document how to import them**: The @three-ws/alibaba-cloud-mcp package, which lets an agent use Qwen chat, embeddings, and model discovery on your own Alibaba Cloud key, could not start at all. Two faults, both on the first line of use: the check for whether the file was launched directly crashed before the server ever opened, and the three tools were registered with a schema format the MCP library rejects. Both are fixed, so the server now boots, advertises its tools, and answers a client. The payment-sessions MCP package had a documentation fault with the same cost: its quick start told you to set your session variable to the whole cookie including the name, which the server then wrapped in the cookie name again, so every session-management call came back unauthorized. It now says to use the cookie value, and ships its two runnable examples inside the published package instead of only in the repository. Five packages (payment sessions, Agora, alerts, the assistant widget, and Alibaba Cloud) gained a Programmatic use section showing the exact code to import their tools into a host of your own, and the sniper package's HTTP table now lists every endpoint it actually serves, including the paid close-position route it had left out. `[fix, sdk, docs]`+- **The embed tutorials now name the attribute that actually hides the built-in chat**: Several tutorials told you to set avatar-chat="off" to hide an agent's built-in chat input. It does not do that: it moves the avatar out of the chat column and lays the chat out as a plain bottom bar, leaving the input on screen. The multi-agent walkthrough leaned on it to keep its behind-the-scenes stage-director messages hidden, so anyone following it verbatim saw those messages leak into the visible chat. The affected tutorials now use chat="off", the real bare-avatar switch, and note that the brain still boots behind it so say(), ask(), and the event stream keep working. The same accuracy pass corrected the walking-avatar leaderboard example, which read fields the API never returns and so always rendered an empty top-walkers list, and the JavaScript API event table, which named the wrong key for the skill a tool-call event reports. (`/tutorials/web-component-end-to-end`) `[fix, docs]`- **The guard and gesture tutorials stop teaching steps that fail, and one that wiped your clip list**: Two tutorials had drifted away from the code they document, and following either one exactly produced a bad outcome rather than a confusing one. The gesture tutorial's API example padded its request body with an empty clip array, and told readers that field was required; it has not been required for a while, and an empty array is a real value, so anyone who copied that snippet to change one gesture slot silently erased every custom clip on their agent. The endpoint's actual rule is now stated: every field is optional, present replaces it, absent leaves it alone. That page also listed four gestures on the /walk quick keys that are not on them, so pressing 1 through 8 never produced what it described; the eight keys are now named in order, with the wheel and the code call given for the four that have no key. The guard tutorial's registry entry left out the proof block, which the registry auditor has since made mandatory, so a reader who followed every step to the end hit a failure the page never mentioned. It now includes a working proof, shows the exact error you get without one, and adds a step for the command that runs it, which is the difference between a guard that is wired up and a guard anyone has shown still catches anything. Its summary line also described a completely different guard from the one it builds. (`/docs/tutorials/write-a-guard`) `[docs, fix]`- **The persona builder stops reporting a model outage as if you had been signed out**: The two routes behind the persona builder (the one that turns your interview answers or a sample of your own writing into a structured persona, and the one that replies to you in that persona's voice) used to hand you whatever status code the model provider returned. If a server-side key expired, the provider answered 401 and so did we, which reads to a browser as an expired session on a request where nothing was wrong with your sign-in. Both now report a provider failure as exactly what it is: a server-side fault, with a plain message and no implication that you need to log back in. Separately, when the platform's daily model-spend ceiling is reached, the builder now says so and tells you when it resets, instead of asking you to retry something that could not have succeeded for the rest of the day. Both routes also now record their model usage against your account, so the spend they generate is visible and capped rather than untracked, and the daily extraction budget now holds across every server instance instead of being counted separately on each one. `[fix, improvement]`- **The plugin marketplace stops hiding a plugin when you page through it**: Browsing plugins a page at a time could show you the same plugin twice and never show another one at all. The catalogue was sorted by install count and publish date, and the built-in plugins were all published in the same instant with zero installs, so that sort left them tied and the database was free to order the tied rows differently on every request. One plugin therefore landed on two consecutive pages and one fell through the gap. The order is now decided all the way down, so paging through the catalogue returns every plugin exactly once. Searching also takes your text literally: typing a percent sign or an underscore used to act as a database wildcard and match rows it should not. Publishing got stricter at the same time, because a plugin manifest is handed to the model as a tool definition by whoever installs it: a tool name now has to be plain letters, digits, underscores, or hyphens, every tool needs a real description, and a manifest has to fit the same 64KB ceiling that importing one by URL already enforced. And pointing the importer at a repository web page instead of the raw manifest file now says exactly that, rather than showing a raw parser error. (`/marketplace`) `[fix, improvement]`+- **The pump.fun trade history works again, notification settings show every category, and a bad agent id stops reading as a server crash**: Three fixes across the public MCP packages and the API behind them. Asking for a token's recent buys and sells (get_token_trades on the free pump.fun MCP server) had stopped working entirely: the platform reads several transactions in one batched call, and the free Solana providers we route through cap those batches at one call each, so they rejected the whole request and the tool returned a raw provider error instead of trades. The platform now recognises that refusal, re-sends the reads one at a time to the same provider, and remembers the cap so later calls skip the wasted attempt; every batched on-chain read benefits, including wallet and agent activity history. Separately, the notifications MCP server told agents there were six notification categories when there are seven, so an agent could never change how it hears about a finished 3D generation; all seven are now listed, along with the security notices that always reach your bell inbox. And the agent provenance log answered a non-uuid agent id with an internal error and a support reference; it now says plainly that the id must be a uuid. `[fix, sdk]`- **The vanity receipt verifier stops calling honest receipts fraudulent, and the avatar CLI takes a Solana address**: Two documented commands did not do what their page said. The command-line verifier for a provably-fair vanity address receipt scored the difficulty claim with the retired flat model, the one that treats every Base58 character as equally likely. Every receipt issued since the leading-character correction landed therefore failed its difficulty check and was reported as untrustworthy, even though the in-browser verifier at /vanity/verify passed the same receipt. The verifier now reads the model each receipt names in its own signed difficulty block, exactly as the page and the SDK do, so current receipts verify and older ones keep verifying under the model they were issued under. Separately, the Rig Doctor page showed an avatar-manifest command taking a Solana address as the owner, which the CLI rejected outright: it understood Ethereum addresses and full CAIP-10 identifiers and nothing else. Since Solana is the home chain, a bare Solana address is now accepted and recorded as solana:mainnet-beta. Alongside those, eight tutorial and reference pages were audited against the code they describe: the Vercel self-hosting guide had a stale copy of the config file and pointed at a migrations folder that has moved, the MCP guide listed three placeholder OAuth scopes where the live server publishes twelve real per-capability ones, and the Avatar CLI page pinned a version number that npm has never had. (`/docs/rig-doctor`) `[fix, docs, sdk]`- **The vanity SDK now uses the payment wallet you hand it, and the in-chat 3D avatar keeps its tuned lighting**: Three fixes in the developer packages. The @three-ws/vanity SDK documented a fetch option on its paid address-grinding lane but silently ignored it, so a developer who passed a payment-enabled fetch got a payment-required error back instead of the keypair they paid for; that option now works, as do a custom API host and key. The avatar MCP server's in-chat 3D viewer had been hand-tuned with softer shadows, brighter exposure and a neutral environment light, but the script that rebuilds it before every release still produced the old flat version, so the next release would have quietly undone the tuning; the script now produces the tuned viewer, and the same settings apply to the fallback view for chat apps that cannot run the interactive one. The tool SDK's quick-start example pointed at an API that does not exist and skipped the wrapper an MCP server needs to return anything at all, so following it produced a tool that answered with nothing; it now runs end to end against a real free three.ws endpoint. `[fix, sdk, docs]`- **Three docs pages stop pointing readers at a dead page, a missing section, and a snippet that does nothing**: The Vanity page told you to grind a long address yourself in the browser at /vanity, but that path has never existed and answers 404; the in-browser grinder lives at /vanity-wallet, and both links on the page now go there. In Troubleshooting, the fix for a Quick Look model that opens and instantly closes linked to a model-optimization section that is not on the AR page, so the link dropped you at the top with nothing to read; it now opens the section that actually covers it. The same page told you to add a responsive attribute to size an inline embed, with a snippet that set the width in CSS. Neither half worked: responsive sizing is already on by default, and the element reads the width attribute rather than the computed style, so the documented snippet never produced the 3:4 ratio it promised. The instructions now show the form that does work, and name the two things that switch it off. (`/docs/troubleshooting`) `[fix, docs]`+- **Voice and vision SDKs ship runnable examples, and the x402 seller SDK reads its Solana sponsor from the environment**: @three-ws/voice and @three-ws/vision-mcp now include example scripts you can run as-is against the live free lanes: the voice one speaks a sentence, animates a face from those exact bytes, and transcribes it back; the vision one probes the lane, writes alt text for an image, and reads the text off it. @three-ws/x402-server used to demand the Solana sponsor account as a code option even though its own error message told you to set X402_FEE_PAYER_SOLANA, so a correctly configured deployment still got missing_fee_payer; it now reads that environment variable when the option is omitted. The VS Code x402 extension's endpoint scaffolder was generating Base-only endpoints and now generates Solana-first ones, matching every paid route on the platform, and it ships with a test suite covering which payment requirement it selects. `[sdk, fix]`## 2026-08-14diff --git a/data/changelog.json b/data/changelog.jsonindex 1f663fec2..c081568de 100644--- a/data/changelog.json+++ b/data/changelog.json@@ -1,6 +1,16 @@{"$comment": "Curated changelog entries for three.ws, the editorial layer of the public changelog. New PAGE launches are tracked automatically via the `added` field in data/pages.json; this file is for everything else holders care about: improvements, fixes, SDK releases, trading/payments capabilities, security work. scripts/build-page-index.mjs merges both sources into CHANGELOG.md, public/changelog.json, and public/changelog.xml. Newest first. Allowed tags: feature, improvement, fix, sdk, infra, docs, security. `link` is optional and must be a live three.ws path.","entries": [+ {+ "date": "2026-08-15",+ "title": "The 3D pipeline docs stop telling you to run a command that does not exist",+ "summary": "Every paid stage on the 3D asset pipeline page (rig, remesh, game-ready, stylize, background removal) showed you how to pay for it by running npx @x402/fetch. There is no such command: that package publishes a library, not a CLI, so anyone who copied a sample got a could-not-determine-executable error and no asset. The samples are now the real thing: an unpaid curl that returns the live 402 challenge and costs nothing, plus one Node snippet that wraps fetch with our published @three-ws/x402-fetch buyer and actually settles. The end-to-end generate, rig and game-ready walkthrough was rewritten the same way and now polls the forge job with the fields the API really returns. The page also stops promising a specific settlement network, because the 402 challenge names the networks itself. Alongside it: the changelog page CDN pinning table listed four alias URLs that all 404 and pointed at an integrity file for a version that no longer publishes one; the AgenC page opening sentence linked the word AgenC to a different protocol page; and the vanity protocol spec showed a stale copy of the service key document, which now matches what the live endpoint serves.",+ "tags": [+ "docs",+ "fix"+ ],+ "link": "/docs/3d-pipeline"+ },{"date": "2026-08-15","title": "The pump.fun trade history works again, notification settings show every category, and a bad agent id stops reading as a server crash",diff --git a/docs/tutorials/connect-ai-brain.md b/docs/tutorials/connect-ai-brain.mdindex b09ed9058..4fa123c05 100644--- a/docs/tutorials/connect-ai-brain.md+++ b/docs/tutorials/connect-ai-brain.md@@ -341,7 +341,9 @@ The standard path is to store your key in the platform dashboard. If you have aThe flow:1. You run an endpoint at, say, `https://your-domain.com/api/llm-key`. It returns a JSON response with a short-lived API key (or a session token usable as one), valid for some window you control.-2. You set the `key-proxy` attribute on your `<agent-3d>` element or in the agent's dashboard settings:+2. You point the agent at it. Two places, and they do different things:+ - The **`key-proxy` attribute** on `<agent-3d>` replaces the platform proxy entirely: the runtime sends its Anthropic-shape request straight to your URL instead of `/api/llm/anthropic`. Your endpoint is then responsible for the upstream call and for streaming Anthropic-shape SSE back.+ - **`brain.mode: 'key-proxy'`** with a `proxy_url` on the embed policy declares the same intent server-side, which is what closes the platform's own we-pay lane. The policy schema rejects `key-proxy` mode without a `proxy_url`, so the two can't drift apart.```html<agent-3d@@ -350,7 +352,7 @@ The flow:></agent-3d>```-3. When the platform needs to make an LLM call for this agent, it calls your endpoint first to get a fresh key, then uses that key for the LLM request.+3. Every LLM call for this agent goes through your endpoint, so the long-lived key never leaves your infrastructure.This is genuinely advanced — most teams don't need it. The reasons to reach for it:@@ -364,17 +366,17 @@ For everything else, the dashboard key store is the right answer.## Step 10 — A concrete switch: walk through-To make this concrete, here's the actual sequence for moving a production agent from the managed free tier to your own Anthropic key, then upgrading the model from Sonnet to Opus.+To make this concrete, here's the actual sequence for moving a production agent off the free lane and onto Claude, then upgrading it.1. **Get your key.** [console.anthropic.com](https://console.anthropic.com) → API Keys → Create Key. Label it "three.ws production".-2. **Add to platform.** [three.ws/my-agents](https://three.ws/my-agents) → gear icon → API Keys → Anthropic → paste, save.-3. **Switch the agent's brain provider.** Open your agent. **Brain → Provider → Anthropic.** Verify the dropdown for **Use my key** is selected, not **Use managed credit**.-4. **Test.** Open any page that embeds the agent. Send a message. Check the Usage tab — input/output tokens should be incrementing against your key, not the managed credit.-5. **Upgrade the model.** **Brain → Model → Claude Opus 4.7.** Save.-6. **Test again.** Notice the reply time is slightly slower (~1.5s first token vs ~600ms) and the responses are noticeably more thoughtful on hard questions.-7. **Set a budget alert.** Account → Billing → Alerts → $20 / day. You'll get an email if Opus pushes your spend up faster than expected.+2. **Add it to the platform.** [/dashboard/account](https://three.ws/dashboard/account) → **AI Provider Keys** → **Anthropic (Claude)** → paste, save.+3. **Compare before you commit.** Open [/dashboard/brain](https://three.ws/dashboard/brain), select Sonnet 5 and one alternative, and run the same prompt through both in the playground.+4. **Pin the model.** `PUT /api/agents/:id/embed-policy` with `brain.model` set to `claude-sonnet-5` (the read-modify-write in Step 8). While you're there, set `monthly_quota` and lock `origins` to your own hosts.+5. **Test.** Open any page that embeds the agent and send a message. `GET /api/agents/:id/usage` should show `currentMonthCalls` incrementing.+6. **Upgrade if needed.** Repeat step 4 with `claude-opus-5`. Expect slower first tokens (it thinks by default) and better answers on hard questions. Fable 5 is the rung above that.+7. **Confirm the ceiling.** Re-read the policy and check `monthly_quota` and `rate_limit_per_min` are the numbers you meant. These are enforced before spend, so they are your real budget control.-Total time: under five minutes. The agent IDs in your existing embeds don't change. Visitors notice the smarter replies but the front-end is identical.+Total time: under ten minutes. The agent IDs in your existing embeds don't change. Visitors notice the smarter replies but the front-end is identical.---@@ -382,16 +384,17 @@ Total time: under five minutes. The agent IDs in your existing embeds don't chanThe brain layer in full:-- LLM calls happen server-side; your key never leaves the platform's backend-- Keys are stored per-account in My Agents; one key serves all your agents from that provider-- Sonnet 4.6 is the default; Opus 4.7 and GPT-5.6 Sol are the reasoning-grade upgrades; Haiku and GPT-5.6 Luna are the latency picks-- Streaming is on by default and matters more than total latency for perceived speed-- Tool-use quality varies by model — pick Sonnet or larger for skill-heavy agents-- Usage and cost are visible in the dashboard, with alerts available for spend caps-- Switching models is a dashboard toggle; the embed snippet never needs to change+- LLM calls happen server-side through `/api/llm/anthropic`; your key never leaves the backend+- Keys are stored per-account at [/dashboard/account](https://three.ws/dashboard/account) → AI Provider Keys; one key serves all your agents from that provider+- The embed default is a free open-weight lane. Sonnet 5 is the recommended paid default; Opus 5 and GPT-5.6 Sol are the reasoning-grade upgrades, Fable 5 the ceiling; Haiku 4.5 and GPT-5.6 Luna are the latency picks+- Streaming is always on and matters more than total latency for perceived speed+- A visitor can never escalate you onto a paid model: the proxy clamps any first-party-billed model back to the owner's configured one+- Tool-use quality varies by model — pick Sonnet 5 or larger for skill-heavy agents+- `monthly_quota` and `rate_limit_per_min` on the embed policy are the real budget controls, enforced before spend; `GET /api/agents/:id/usage` is the counter to watch+- Switching models is one field on the embed policy; the embed snippet never needs to change- Key proxies are an advanced option for compliance-heavy setups-The brain is the part of the agent you'll iterate on most after the system prompt. Pick a sensible default (Sonnet 4.6), ship, and only upgrade once you've identified a specific quality gap that prompt iteration can't close.+The brain is the part of the agent you'll iterate on most after the system prompt. Pick a sensible default (Sonnet 5), ship, and only upgrade once you've identified a specific quality gap that prompt iteration can't close.## Next stepsdiff --git a/public/changelog.json b/public/changelog.jsonindex 839e29e7a..e41fb8d24 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,5 +1,5 @@{- "generated_at": "2026-08-15T06:49:46.322Z",+ "generated_at": "2026-08-15T06:50:56.937Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws",@@ -38,6 +38,16 @@"security"]},+ {+ "date": "2026-08-15",+ "type": "update",+ "title": "Avatars can move their lips again: the Audio2Face lane is back",+ "summary": "Every call that turns speech into facial animation had been failing with an upstream not-found error, so a talking avatar spoke with a frozen face and the one-shot say() path returned nothing usable. The hosted model id the lane pointed at had stopped resolving for our NVIDIA account. It now points at the hosted Audio2Face NIM that is actually live, verified end to end: a synthesized clip drives a 30 fps ARKit blendshape track whose jaw genuinely opens and closes with the words. The discovery helper that finds these model ids was missing the name NVIDIA publishes them under, which is why the outage looked like a configuration mystery; it now finds them.",+ "link": null,+ "tags": [+ "fix"+ ]+ },{"date": "2026-08-15","type": "update",@@ -93,6 +103,17 @@"fix"]},+ {+ "date": "2026-08-15",+ "type": "update",+ "title": "The 3D pipeline docs stop telling you to run a command that does not exist",+ "summary": "Every paid stage on the 3D asset pipeline page (rig, remesh, game-ready, stylize, background removal) showed you how to pay for it by running npx @x402/fetch. There is no such command: that package publishes a library, not a CLI, so anyone who copied a sample got a could-not-determine-executable error and no asset. The samples are now the real thing: an unpaid curl that returns the live 402 challenge and costs nothing, plus one Node snippet that wraps fetch with our published @three-ws/x402-fetch buyer and actually settles. The end-to-end generate, rig and game-ready walkthrough was rewritten the same way and now polls the forge job with the fields the API really returns. The page also stops promising a specific settlement network, because the 402 challenge names the networks itself. Alongside it: the changelog page CDN pinning table listed four alias URLs that all 404 and pointed at an integrity file for a version that no longer publishes one; the AgenC page opening sentence linked the word AgenC to a different protocol page; and the vanity protocol spec showed a stale copy of the service key document, which now matches what the live endpoint serves.",+ "link": "/docs/3d-pipeline",+ "tags": [+ "docs",+ "fix"+ ]+ },{"date": "2026-08-15","type": "update",@@ -138,6 +159,17 @@"docs"]},+ {+ "date": "2026-08-15",+ "type": "update",+ "title": "The embed tutorials now name the attribute that actually hides the built-in chat",+ "summary": "Several tutorials told you to set avatar-chat=\"off\" to hide an agent's built-in chat input. It does not do that: it moves the avatar out of the chat column and lays the chat out as a plain bottom bar, leaving the input on screen. The multi-agent walkthrough leaned on it to keep its behind-the-scenes stage-director messages hidden, so anyone following it verbatim saw those messages leak into the visible chat. The affected tutorials now use chat=\"off\", the real bare-avatar switch, and note that the brain still boots behind it so say(), ask(), and the event stream keep working. The same accuracy pass corrected the walking-avatar leaderboard example, which read fields the API never returns and so always rendered an empty top-walkers list, and the JavaScript API event table, which named the wrong key for the skill a tool-call event reports.",+ "link": "/tutorials/web-component-end-to-end",+ "tags": [+ "fix",+ "docs"+ ]+ },{"date": "2026-08-15","type": "update",@@ -171,6 +203,17 @@"improvement"]},+ {+ "date": "2026-08-15",+ "type": "update",+ "title": "The pump.fun trade history works again, notification settings show every category, and a bad agent id stops reading as a server crash",+ "summary": "Three fixes across the public MCP packages and the API behind them. Asking for a token's recent buys and sells (get_token_trades on the free pump.fun MCP server) had stopped working entirely: the platform reads several transactions in one batched call, and the free Solana providers we route through cap those batches at one call each, so they rejected the whole request and the tool returned a raw provider error instead of trades. The platform now recognises that refusal, re-sends the reads one at a time to the same provider, and remembers the cap so later calls skip the wasted attempt; every batched on-chain read benefits, including wallet and agent activity history. Separately, the notifications MCP server told agents there were six notification categories when there are seven, so an agent could never change how it hears about a finished 3D generation; all seven are now listed, along with the security notices that always reach your bell inbox. And the agent provenance log answered a non-uuid agent id with an internal error and a support reference; it now says plainly that the id must be a uuid.",+ "link": null,+ "tags": [+ "fix",+ "sdk"+ ]+ },{"date": "2026-08-15","type": "update",@@ -206,6 +249,17 @@"docs"]},+ {+ "date": "2026-08-15",+ "type": "update",+ "title": "Voice and vision SDKs ship runnable examples, and the x402 seller SDK reads its Solana sponsor from the environment",+ "summary": "@three-ws/voice and @three-ws/vision-mcp now include example scripts you can run as-is against the live free lanes: the voice one speaks a sentence, animates a face from those exact bytes, and transcribes it back; the vision one probes the lane, writes alt text for an image, and reads the text off it. @three-ws/x402-server used to demand the Solana sponsor account as a code option even though its own error message told you to set X402_FEE_PAYER_SOLANA, so a correctly configured deployment still got missing_fee_payer; it now reads that environment variable when the option is omitted. The VS Code x402 extension's endpoint scaffolder was generating Base-only endpoints and now generates Solana-first ones, matching every paid route on the platform, and it ships with a test suite covering which payment requirement it selects.",+ "link": null,+ "tags": [+ "sdk",+ "fix"+ ]+ },{"date": "2026-08-14","type": "launch",diff --git a/public/changelog.xml b/public/changelog.xmlindex e169eda1a..499283010 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -30,6 +30,14 @@<category>fix</category><description>The arm page clamps the risk knobs on the server so a bad number can never reach the trading loop, and those clamps have a floor: the smallest trade the loop will place is 0.001 SOL. That floor was being applied in the wrong direction. Ask to arm a live agent with a per-trade size of zero, and instead of being told no, you were armed for real spending at 0.001 SOL a trade with a half-SOL daily ceiling you never asked for. The check meant to catch this looked at the number after the clamp had already raised it, so it could never fire. Arming live now refuses any per-trade size below the floor, and any daily cap smaller than a single trade, and tells you which number to fix; nothing is armed until you do. Clamping a too-large number back down to the ceiling is unchanged, and simulate runs keep the same forgiving behaviour as before. The arm page also shows the reason a save was rejected instead of one generic line, so an ownership problem no longer reads the same as a bad trade size, and a failed Telegram test now prints what Telegram actually said.</description></item>+ <item>+ <title>Avatars can move their lips again: the Audio2Face lane is back</title>+ <link>https://three.ws/changelog</link>+ <guid isPermaLink="false">2026-08-15:Avatars can move their lips again: the Audio2Face lane is back</guid>+ <pubDate>Sat, 15 Aug 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>Every call that turns speech into facial animation had been failing with an upstream not-found error, so a talking avatar spoke with a frozen face and the one-shot say() path returned nothing usable. The hosted model id the lane pointed at had stopped resolving for our NVIDIA account. It now points at the hosted Audio2Face NIM that is actually live, verified end to end: a synthesized clip drives a 30 fps ARKit blendshape track whose jaw genuinely opens and closes with the words. The discovery helper that finds these model ids was missing the name NVIDIA publishes them under, which is why the outage looked like a configuration mystery; it now finds them.</description>+ </item><item><title>Docs pages stop opening with a wall of raw metadata, and the AWS pages say where the Marketplace listing actually stands</title><link>https://three.ws/docs/aws-marketplace</link>@@ -70,6 +78,14 @@<category>docs</category><description>The deployment guide told self-hosters to set Cloud Run environment variables with `gcloud run services update --set-env-vars`. That flag replaces the service environment wholesale, so following the instruction for a single key silently deletes every other variable the service had. The guide now uses `--update-env-vars` and says plainly why. Four more corrections landed in the same sweep across the deployment, configuration and contributing pages: the Avaturn signup link and the GitHub Discussions links were dead (Discussions are not enabled on the repository, so bug reports and design questions both go to Issues); the in-browser avatar builder was documented as needing its own hosted origin in production when the normal build already serves it same-origin at /avatar-studio; and three counts that had drifted from the code (cron jobs, MCP server packages, contract tests) were replaced with pointers to the files that hold the real numbers.</description></item>+ <item>+ <title>The 3D pipeline docs stop telling you to run a command that does not exist</title>+ <link>https://three.ws/docs/3d-pipeline</link>+ <guid isPermaLink="false">2026-08-15:The 3D pipeline docs stop telling you to run a command that does not exist</guid>+ <pubDate>Sat, 15 Aug 2026 12:00:00 GMT</pubDate>+ <category>docs</category>+ <description>Every paid stage on the 3D asset pipeline page (rig, remesh, game-ready, stylize, background removal) showed you how to pay for it by running npx @x402/fetch. There is no such command: that package publishes a library, not a CLI, so anyone who copied a sample got a could-not-determine-executable error and no asset. The samples are now the real thing: an unpaid curl that returns the live 402 challenge and costs nothing, plus one Node snippet that wraps fetch with our published @three-ws/x402-fetch buyer and actually settles. The end-to-end generate, rig and game-ready walkthrough was rewritten the same way and now polls the forge job with the fields the API really returns. The page also stops promising a specific settlement network, because the 402 challenge names the networks itself. Alongside it: the changelog page CDN pinning table listed four alias URLs that all 404 and pointed at an integrity file for a version that no longer publishes one; the AgenC page opening sentence linked the word AgenC to a different protocol page; and the vanity protocol spec showed a stale copy of the service key document, which now matches what the live endpoint serves.</description>+ </item><item><title>The Agent Economy demo works again when you just pick a service, and it fits a phone screen</title><link>https://three.ws/agent-economy</link>@@ -102,6 +118,14 @@<category>fix</category><description>The @three-ws/alibaba-cloud-mcp package, which lets an agent use Qwen chat, embeddings, and model discovery on your own Alibaba Cloud key, could not start at all. Two faults, both on the first line of use: the check for whether the file was launched directly crashed before the server ever opened, and the three tools were registered with a schema format the MCP library rejects. Both are fixed, so the server now boots, advertises its tools, and answers a client. The payment-sessions MCP package had a documentation fault with the same cost: its quick start told you to set your session variable to the whole cookie including the name, which the server then wrapped in the cookie name again, so every session-management call came back unauthorized. It now says to use the cookie value, and ships its two runnable examples inside the published package instead of only in the repository. Five packages (payment sessions, Agora, alerts, the assistant widget, and Alibaba Cloud) gained a Programmatic use section showing the exact code to import their tools into a host of your own, and the sniper package's HTTP table now lists every endpoint it actually serves, including the paid close-position route it had left out.</description></item>+ <item>+ <title>The embed tutorials now name the attribute that actually hides the built-in chat</title>+ <link>https://three.ws/tutorials/web-component-end-to-end</link>+ <guid isPermaLink="false">2026-08-15:The embed tutorials now name the attribute that actually hides the built-in chat</guid>+ <pubDate>Sat, 15 Aug 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>Several tutorials told you to set avatar-chat="off" to hide an agent's built-in chat input. It does not do that: it moves the avatar out of the chat column and lays the chat out as a plain bottom bar, leaving the input on screen. The multi-agent walkthrough leaned on it to keep its behind-the-scenes stage-director messages hidden, so anyone following it verbatim saw those messages leak into the visible chat. The affected tutorials now use chat="off", the real bare-avatar switch, and note that the brain still boots behind it so say(), ask(), and the event stream keep working. The same accuracy pass corrected the walking-avatar leaderboard example, which read fields the API never returns and so always rendered an empty top-walkers list, and the JavaScript API event table, which named the wrong key for the skill a tool-call event reports.</description>+ </item><item><title>The guard and gesture tutorials stop teaching steps that fail, and one that wiped your clip list</title><link>https://three.ws/docs/tutorials/write-a-guard</link>@@ -126,6 +150,14 @@<category>fix</category><description>Browsing plugins a page at a time could show you the same plugin twice and never show another one at all. The catalogue was sorted by install count and publish date, and the built-in plugins were all published in the same instant with zero installs, so that sort left them tied and the database was free to order the tied rows differently on every request. One plugin therefore landed on two consecutive pages and one fell through the gap. The order is now decided all the way down, so paging through the catalogue returns every plugin exactly once. Searching also takes your text literally: typing a percent sign or an underscore used to act as a database wildcard and match rows it should not. Publishing got stricter at the same time, because a plugin manifest is handed to the model as a tool definition by whoever installs it: a tool name now has to be plain letters, digits, underscores, or hyphens, every tool needs a real description, and a manifest has to fit the same 64KB ceiling that importing one by URL already enforced. And pointing the importer at a repository web page instead of the raw manifest file now says exactly that, rather than showing a raw parser error.</description></item>+ <item>+ <title>The pump.fun trade history works again, notification settings show every category, and a bad agent id stops reading as a server crash</title>+ <link>https://three.ws/changelog</link>+ <guid isPermaLink="false">2026-08-15:The pump.fun trade history works again, notification settings show every category, and a bad agent id stops reading as a server crash</guid>+ <pubDate>Sat, 15 Aug 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>Three fixes across the public MCP packages and the API behind them. Asking for a token's recent buys and sells (get_token_trades on the free pump.fun MCP server) had stopped working entirely: the platform reads several transactions in one batched call, and the free Solana providers we route through cap those batches at one call each, so they rejected the whole request and the tool returned a raw provider error instead of trades. The platform now recognises that refusal, re-sends the reads one at a time to the same provider, and remembers the cap so later calls skip the wasted attempt; every batched on-chain read benefits, including wallet and agent activity history. Separately, the notifications MCP server told agents there were six notification categories when there are seven, so an agent could never change how it hears about a finished 3D generation; all seven are now listed, along with the security notices that always reach your bell inbox. And the agent provenance log answered a non-uuid agent id with an internal error and a support reference; it now says plainly that the id must be a uuid.</description>+ </item><item><title>The vanity receipt verifier stops calling honest receipts fraudulent, and the avatar CLI takes a Solana address</title><link>https://three.ws/docs/rig-doctor</link>@@ -150,6 +182,14 @@<category>fix</category><description>The Vanity page told you to grind a long address yourself in the browser at /vanity, but that path has never existed and answers 404; the in-browser grinder lives at /vanity-wallet, and both links on the page now go there. In Troubleshooting, the fix for a Quick Look model that opens and instantly closes linked to a model-optimization section that is not on the AR page, so the link dropped you at the top with nothing to read; it now opens the section that actually covers it. The same page told you to add a responsive attribute to size an inline embed, with a snippet that set the width in CSS. Neither half worked: responsive sizing is already on by default, and the element reads the width attribute rather than the computed style, so the documented snippet never produced the 3:4 ratio it promised. The instructions now show the form that does work, and name the two things that switch it off.</description></item>+ <item>+ <title>Voice and vision SDKs ship runnable examples, and the x402 seller SDK reads its Solana sponsor from the environment</title>+ <link>https://three.ws/changelog</link>+ <guid isPermaLink="false">2026-08-15:Voice and vision SDKs ship runnable examples, and the x402 seller SDK reads its Solana sponsor from the environment</guid>+ <pubDate>Sat, 15 Aug 2026 12:00:00 GMT</pubDate>+ <category>sdk</category>+ <description>@three-ws/voice and @three-ws/vision-mcp now include example scripts you can run as-is against the live free lanes: the voice one speaks a sentence, animates a face from those exact bytes, and transcribes it back; the vision one probes the lane, writes alt text for an image, and reads the text off it. @three-ws/x402-server used to demand the Solana sponsor account as a code option even though its own error message told you to set X402_FEE_PAYER_SOLANA, so a correctly configured deployment still got missing_fee_payer; it now reads that environment variable when the option is omitted. The VS Code x402 extension's endpoint scaffolder was generating Base-only endpoints and now generates Solana-first ones, matching every paid route on the platform, and it ships with a test suite covering which payment requirement it selects.</description>+ </item><item><title>Agent Interaction Lab</title><link>https://three.ws/demos/agents</link>diff --git a/public/features.json b/public/features.jsonindex 80895ec36..f03496b0c 100644--- a/public/features.json+++ b/public/features.json@@ -1,5 +1,5 @@{- "generated_at": "2026-08-15T06:49:46.116Z",+ "generated_at": "2026-08-15T06:50:56.715Z","generated_by": "scripts/build-page-index.mjs from data/pages.json","site": {"name": "three.ws",