Skip to content

armgit / E2B

Public Imported from GitHub
0

Open-source, secure environment with real-world tools for enterprise-grade agents.

50.2 KB
Raw Blame

@e2b/python-sdk

2.59.0

Minor Changes

  • e5a9fa6: Add dnsServers to the network options (dns_servers in Python): the DNS servers a sandbox can reach when allowOut contains a domain. Leave it out to allow the default, 8.8.8.8, pass a list of IPs to allow those instead, or pass [] to allow none. It only opens the firewall and doesn't change the sandbox's /etc/resolv.conf. updateNetwork/update_network reset it to the default when it is left out, like the other fields, and getInfo/get_info report it when set.

Patch Changes

  • 5a714f0: Deprecate the apiUrl / api_url and sandboxUrl / sandbox_url connection options and their E2B_API_URL and E2B_SANDBOX_URL environment variables. They keep working for now and will be removed in the next major version.
  • 61b6b55: Use the unified https://sandbox.<domain> endpoint for sandbox traffic on e2b-tango.dev, e2b-juliett.dev and foxtrot-2.e2b.dev.
  • 2cd2305: fix(python-sdk): don't crash the build error path on a non-numeric reason.step

2.58.0

Minor Changes

  • 3c4c7cf: - JS: sandbox.commands.streamInput(pid) / sandbox.pty.streamInput(pid) return a CommandInputHandle that sends ordered stdin/PTY input over one StreamInput RPC.
    • Python async: stream_input(pid) returns an AsyncCommandInputHandle; Python sync: stream_input(pid, data) blocks while it sends an iterable of input.

Patch Changes

  • d7ed073: Remove Python 3.10 compatibility shims (Python >=3.11 is already required).

2.57.0

Minor Changes

  • a1ad326: Add tls to the egress proxy options as true or { serverName?, caCert? } (server_name/ca_cert in Python); getInfo/get_info reports it without the CA.

Patch Changes

  • f572bc7: Bump protobuf-py (including native protobuf-py-ext) to 0.6.0 and connectrpc to 0.12.1
  • 999782e: Share the process kill implementation between commands and PTY, share the CLI --language option parser between template init and template migrate, and drop unused internal exports

2.56.0

Minor Changes

  • 547decc: Require Python 3.11 or newer (Python 3.10 is no longer supported).
  • 775f017: Balance HTTP/2 streams with pyqwest's native connection pool instead of a Python-side balancer over multiple transports. The pool dials a new connection once every existing one is at the server-advertised concurrent-stream limit, capped at E2B_MAX_CONNECTIONS (default 200) connections per resolved address. E2B_STREAMS_PER_CONNECTION is removed — the limit now comes from the server's SETTINGS_MAX_CONCURRENT_STREAMS.

Patch Changes

2.55.1

Patch Changes

  • 29064bc: ConnectionConfig.get_sandbox_url / get_sandbox_direct_url take an optional port (defaults to the envd port).

2.55.0

Minor Changes

  • b52cd30: Add mode: 'full' | 'filesystem' (exported as the SnapshotMode type) to createSnapshot / create_snapshot. With 'filesystem', only the filesystem is persisted: the snapshot is smaller and faster to take, and sandboxes created from it cold-boot (start fresh from disk) instead of restoring memory. The source sandbox keeps running either way. Omitted, the API default (a full memory snapshot) applies.

    pause() and the lifecycle.onTimeout / lifecycle["on_timeout"] pause object form accept the same mode. Their keepMemory / keep_memory option is deprecated in favor of it (keepMemory: false is mode: 'filesystem') and keeps working; passing both is an InvalidArgumentError / InvalidArgumentException.

Patch Changes

  • 0894595: Remove dead empty-chunk guards from command output handling; envd never sends empty output chunks, so behavior is unchanged.

2.54.0

Minor Changes

  • cc85f69: Replace the third-party Dockerfile parsers (dockerfile-ast, dockerfile-parse) used by Template.fromDockerfile / Template.from_dockerfile with a built-in parser ported from BuildKit. Quoted and escaped values in ENV/ARG/COPY, whitespace inside RUN/CMD arguments, mixed-case AS aliases, comments in line continuations, the # escape= directive, COPY --chmod, heredocs and ENTRYPOINT + CMD combination are now handled consistently in both SDKs. A .dockerignore line consisting of a lone ! is now rejected, as Docker does, instead of being skipped.

Patch Changes

2.53.1

Patch Changes

  • cd4c62b: Publish the changes from the previous release, which was versioned but failed to publish to npm and PyPI.

2.53.0

Minor Changes

  • 1de92fc: Add SandboxNotRunningError / SandboxNotRunningException and SandboxUnreachableError / SandboxUnreachableException for envd requests that fail because the sandbox cannot be reached. Both extend TimeoutError / TimeoutException, which these cases raised before, so existing instanceof TimeoutError / except TimeoutException handling keeps working.

    • SandboxNotRunning* — the proxy answered 502 "The sandbox was not found" (killed, paused or timed out), either directly or from the health probe the SDK runs after a dropped connection or an unexplained Unavailable (e.g. the sandbox was killed mid-command). Retrying will not help.
    • SandboxUnreachable* — the sandbox is not confirmed stopped but cannot be reached: 502 "port is not open", other 502s, or a connection failure whose health probe got no answer. The original error is the cause / __cause__.
    • A connection failure whose health probe finds the sandbox running still raises the original transport error. The probe itself runs without connection retries.

    Breaking (JS): commands.run / connect, pty.create / connect and files.watchDir on a gone sandbox now throw SandboxNotRunningError instead of SandboxNotFoundError, like every other call. SandboxNotFoundError is only raised by the control plane API.

Patch Changes

  • 8f51c9f: Share the sandbox create/connect/fork response mapping (JS and Python) and the Python Commands.run / Commands.connect handle setup

  • b3bb653: Remove the E2B_USER_AGENT_SOURCE environment variable. The SDKs no longer add a source/<value> User-Agent token, append ?source= to Code Interpreter requests, or log trace IDs and add (trace_id=…) to error messages when it is set to ci, so SDK behavior no longer depends on that variable.

    Code Interpreter errors for statuses other than 404 and 502 now use the same <status> <reason>[: <body>] message in JS and Python: JS includes the response body, and Python falls back to the reason phrase when the body is empty.

2.52.1

Patch Changes

  • cc29617: Match BuildKit when filtering the template build context with .dockerignore: a leading ** followed by literal characters (e.g. **.txt) now matches at any depth, a negated pattern matching only a parent directory (e.g. src/app.ts then !src) no longer re-includes the path, and in the JS SDK ? and bracket expressions match characters outside the BMP (e.g. emoji) as a single character.
  • 3ba1d05: Retry a 502 response only for operations that are safe to replay: a gateway may answer 502 after the API already processed the request, so a 502 from sandbox creation, fork, snapshot or another resource-creating POST is returned instead of being retried (which could create a duplicate). 503 is still retried for every operation. Secret updates (POST /secrets/{id}) are no longer replayed after a 502 or a dropped connection, as each update appends a new version.

2.52.0

Minor Changes

  • 2de0cc3: Cap sandbox fork count at 20. e2b sandbox fork --count, JavaScript Sandbox.fork({ count }), and Python Sandbox.fork(count=...) reject a count outside 1–20 before the API call. Counts from 21 through 100 used to reach the API. Omitting count still leaves the field off the request so the API default applies.

  • 03887ad: Apply .dockerignore and fileIgnorePatterns / file_ignore_patterns the way Docker does when copying files into a template, so ignored files are no longer uploaded or included in the files hash:

    • A pattern that matches a directory (node_modules, .git, dist/) now excludes everything under it, and a leading / is ignored.
    • ! patterns re-include paths; the last matching pattern wins.
    • In Python, patterns now also apply when the copied path contains . or .. segments (for example copy(".")).
    • fileIgnorePatterns / file_ignore_patterns are applied after the .dockerignore lines, so they take precedence.
    • Absolute patterns pointing into the context directory are treated as relative to it.
    • Brace expansion ({a,b}) is not supported, as in Docker. In JS, fileIgnorePatterns such as **/*.{env,pem} previously expanded and now match {env,pem} literally; list each pattern separately instead (**/*.env, **/*.pem).
    • In Python, copying a symlink to a directory copies the link itself instead of the directory's contents, as in JS.
    • Copying a path inside an ignored directory now fails with "No files found", as in Docker. An invalid pattern (such as an unterminated [) now raises an error.

    Directory sizes are no longer part of the files hash, since they depend on the filesystem rather than on the copied files. The files hash of copy() steps that copy directories or are affected by ignore patterns changes once, so those steps are rebuilt on the next build.

  • 2f92cc3: Add an http_version (Python) / httpVersion (JS) connection option, "1.1" or "2" (default), also settable with the E2B_HTTP_VERSION environment variable, to pin requests to the E2B API, to sandboxes (commands, filesystem, PTY) and to volume content to HTTP/1.1. It can also be bound once on a client: E2B(http_version="1.1") / new E2B({ httpVersion: '1.1' }).

Patch Changes

  • 97522f8: Retry control-plane HTTP requests after 502 and 503 responses, in addition to 429, and after a network error once the request was written (dropped connection) — except for POST operations not known to be safe to replay (sandbox creation, fork, snapshot, volume/secret/API-key/webhook creation), which the server may already have processed. Failures to establish the connection keep being retried for every operation (E2B_CONNECTION_RETRIES). Retry-After is honored when the server sends one; otherwise retries back off exponentially with jitter, starting at 0.1 seconds and capped at 10 seconds. The existing retries option (default 3, 0 to disable) and request-timeout budget apply to all of them.
  • f713914: Sandbox commands/files traffic is now balanced across HTTP/2 connections the way the JS SDK's undici agent does it, instead of being sharded into a fixed number of pools by sandbox ID. Every request goes to the connection with the fewest streams in flight to its host; a new connection is dialed only when all existing ones already carry E2B_STREAMS_PER_CONNECTION streams to that host (default 100), up to E2B_MAX_CONNECTIONS (default 200). Envd RPC and HTTP calls share the same connections. E2B_ENVD_POOL_SHARDS and the pool_shard argument of the internal transport factories (get_transport, get_httpx_transport, get_pyqwest_transport, get_envd_transport) are removed.
  • 4975149: Remove unused internal constants and attributes, and share the command handle setup between Commands.run and Commands.connect

2.51.0

Minor Changes

  • 5c0f6eb: Remove SDK-side defaults from API request payloads so the API defaults apply when options are omitted. Sandbox create/fork/connect no longer preset a 5-minute timeout, fork no longer presets count: 1, create no longer presets allow_internet_access, pause no longer presets keeping memory, and template builds no longer preset CPU/memory. Explicitly provided values are still sent unchanged.

    Sandbox create and connect now use the v2 API endpoints (POST /v2/sandboxes, POST /v2/sandboxes/{id}/connect), which default timeout to 5 minutes and always secure envd access. The secure option on Sandbox.create is deprecated: every sandbox is secured, so the option is still accepted but ignored.

Patch Changes

  • 5c0f6eb: Remove client-side validation of the fork count argument. The API validates the requested fork count and rejects invalid values.

2.50.0

Minor Changes

  • 5b015ad: Removed the V1 template build operations and schemas from the generated API clients. The API no longer serves them (runtime 87968fc1e1fa); control planes carrying that change answer 410 Gone. Template builds go through the Template SDK.

    Visible removal, classified minor: the JS paths namespace loses POST /templates, POST /templates/{templateID}, POST /templates/{templateID}/builds/{buildID} and POST /v2/templates, and components['schemas'] loses TemplateLegacy, TemplateBuildRequest and TemplateBuildRequestV2; the Python e2b.api.client.models package loses TemplateLegacy, TemplateBuildRequest and TemplateBuildRequestV2, and e2b.api.client.api.templates loses the post_templates, post_templates_template_id, post_templates_template_id_builds_build_id and post_v2_templates modules. No SDK method accepted or returned them; code that imported these names directly must drop the import.

    The regenerated clients also pick up a documented 429 on most operations, a 409 on template create (v3), the upload-request headers on the build file-upload link, minLength: 1 on the v2 build source fields, and a deprecation marker on the always-empty logs field of the build status.

Patch Changes

  • 956e3ab: Apply the request headers the API returns with a template layer-file upload link. Azure Blob Storage requires x-ms-blob-type on the upload request, which its signed URL cannot carry, so COPY instructions failed on Azure-backed clusters. GCS- and S3-backed clusters return no headers and are unaffected.

2.49.1

Patch Changes

  • 5e418dc: Document that onResume / on_resume needs a control plane that knows the option: an older self-hosted or BYOC control plane drops the memory field and restores memory while reporting success, instead of rejecting the request.
  • 9136603: Retry control-plane HTTP requests up to three times after 429 responses using the server's delta-seconds Retry-After delay. Retries can be configured or disabled with retries, and stop when waiting would exhaust the request timeout. Envd requests, including filesystem operations, and volume-content requests are not retried.

2.49.0

Minor Changes

  • f842aa8: Expose a configurable minimum free-disk target with minFreeDiskMb in JavaScript, min_free_disk_mb in Python, and --min-free-disk-mb in template create. Omission uses the team default, while explicit zero requests no minimum growth. Growth is best effort and never shrinks an existing filesystem.

2.48.0

Minor Changes

  • 08efa36: Add httpsPorts (JS) / https_ports (Python) to the sandbox network config. Ports listed there have their public URLs proxied to the sandbox over HTTPS — use it when the service listening on the port serves TLS itself. This is not TLS passthrough: traffic is still terminated at the E2B proxy and re-encrypted on the hop to the sandbox, and the backend certificate is not verified, so self-signed certificates work. The configured ports are also returned in the sandbox info network config.
  • 6b759bf: Add ServiceBusyError (JavaScript) / ServiceBusyException (Python) for HTTP 503 responses: the API refused the operation because the service or the node running the sandbox is temporarily busy, the sandbox is unchanged, and the call can be retried. A refused pause() is the first case. The base SandboxError / SandboxException also carries the HTTP status as statusCode / status_code when the error came from an API response, so callers can branch on the status without parsing the message. Like AuthenticationError / AuthenticationException, the new class does not subclass the sandbox base error: it is raised for every 503 whatever the operation, so catch it explicitly.

Patch Changes

  • 58c81f1: onTimeout / on_timeout and onResume / on_resume now raise InvalidArgumentError / InvalidArgumentException for a value outside their two literals, instead of silently resolving it to the other one. Both are resolved into a boolean before the request is built, so the value never reaches the API and a typo cannot be rejected server-side: on_timeout="Pause" previously resolved to kill and deleted the sandbox and its snapshot at timeout, and on_resume="Reboot" previously restored the memory the caller asked to skip. A nullish value still means "not configured" and leaves the choice to the API.

2.47.0

Minor Changes

  • 1980d6b: Add onResume / on_resume to Sandbox.connect(): 'reboot' resumes a paused sandbox from its disk state alone, leaving the memory snapshot untouched, for the case where restoring that memory wedges the guest. 'restore' stays the default. Where filesystem-only resume is not enabled, a 'reboot' that would actually drop memory is rejected with an error rather than silently restoring it.
  • 043d050: Removed generated types for endpoints the SDKs never exposed. The JS components['schemas'] namespace and the Python e2b.api.client.models package no longer include the admin, cluster/rig, node, team-API-key and access-token schemas (Node, NodeDetail, NodeMetrics, NodeStatus, NodeStatusChange, MachineInfo, DiskMetrics, Rig*, Admin*, TeamAPIKey, NewTeamAPIKey, CreatedTeamAPIKey, UpdateTeamAPIKey, IdentifierMaskingDetails, VolumeToken; Python additionally Team, TeamMetric, MaxTeamMetric, CreatedAccessToken, NewAccessToken). No SDK method ever accepted or returned them, so code that uses the SDK through its methods is unaffected; code that imported these type names directly must drop the import.

2.46.4

Patch Changes

  • 6146cd0: Improve reliability for high-concurrency sandbox workloads by spreading envd traffic across four HTTP/2 connection pools. Set E2B_ENVD_POOL_SHARDS before importing the SDK to adjust the pool count.

2.46.3

Patch Changes

  • c275393: Improve reliability for high-concurrency sandbox workloads by spreading envd traffic across four HTTP/2 connection pools. Set E2B_ENVD_POOL_SHARDS before importing the SDK to adjust the pool count.

2.46.2

Patch Changes

  • 7dd5321: Spread envd traffic across four HTTP/2 connection pools by default so high-concurrency, long-running streams do not all contend for one connection's stream limit. Set E2B_ENVD_POOL_SHARDS before importing the SDK to tune the pool count.

2.46.1

Patch Changes

  • d4a7f41: Deprecate the sandbox git module (sandbox.git) and its public types and errors. Run git through the commands module instead, e.g. sandbox.commands.run('git clone <url> repo'). The module keeps working and will be removed in the next major version.

2.46.0

Minor Changes

  • 9d1c90d: Remove client-side API key format validation. The SDK no longer checks that the API key matches the e2b_ hex format — only that a key is present. The validateApiKey/validate_api_key option is deprecated and has no effect, and the E2B_VALIDATE_API_KEY environment variable is no longer read. The server remains the source of truth for key validity.

Patch Changes

  • 67c06e0: Point the two Code Interpreter README links at code-interpreting/analyze-data-with-ai instead of the code-interpreting section index. The index has no landing page and 307s to that article, dropping the query string on the way, so the UTM parameters were lost before the reader arrived. Linking at the resolved path keeps them.

  • 182b498: Point the README documentation links at docs.e2b.dev instead of e2b.dev/docs. The docs site moved to its own subdomain and has no /docs path prefix there, so e2b.dev/docs serves a 308 to docs.e2b.dev/ and e2b.dev/docs/code-interpreting maps to docs.e2b.dev/code-interpreting. The UTM parameters are unchanged and survived the redirect, so this removes a redirect hop rather than fixing broken attribution.

  • b802997: Fix two network.egressProxy / network["egress_proxy"] cases an untyped caller reaches.

    The JS SDK had no shape guard: buildEgressProxyBody rebuilds the body from the known fields, so an address that was missing or not a string vanished and the caller got an API error about a config they never wrote ({"egressProxy":{}}). It now raises InvalidArgumentError naming the option, the way the Python SDK already did:

    // InvalidArgumentError: network egressProxy must be an object with a string
    // 'address' (e.g. 'proxy.example.com:1080').
    await Sandbox.create({
      network: { egressProxy: 'proxy.example.com:1080' as never },
    })
    

    A null / None username or password is now treated as absent instead of being serialized as a JSON null the API rejects — reading a credential out of an unset environment variable is how a caller lands there, and it means the proxy takes no credentials:

    await Sandbox.create({
      network: {
        egressProxy: {
          address: 'proxy.example.com:1080',
          // Unset in the environment; the proxy takes no credentials.
          username: process.env.PROXY_USER,
        },
      },
    })
    
    Sandbox.create(
        network={
            "egress_proxy": {
                "address": "proxy.example.com:1080",
                "username": os.environ.get("PROXY_USER"),
            },
        },
    )
    

2.45.1

Patch Changes

  • b17b726: Stop the shared retrying transports from mirroring streamed request bodies in memory. pyqwest's retry middleware keeps a non-bytes body replayable by copying it as it is sent, so a streamed files.write of a file-like object or volume.write_file grew a full in-RAM mirror and peak memory scaled with file size. The transports now declare pyqwest's RetryMode.UNBUFFERED: a streamed body is replayed only while nothing has been read from it, which is all the SDK's connect-only retry policy needs — a ConnectionError is raised only before the request was written. Connect retries are unchanged, and bytes bodies (unary RPC payloads, in-memory writes) were already replayable without a copy.

2.45.0

Minor Changes

  • 8787dfe: Add sorting and new filters to Sandbox.list. The order option ('asc' / 'desc', default 'desc') sorts sandboxes by start time across the whole paginated dataset, and the query now supports startedAfter / started_after (inclusive lower bound on start time) and template (exact template ID or alias) filters, all applied server-side before pagination. The CLI e2b sandbox list command exposes these via --order, --started-after, and --template.

2.44.0

Minor Changes

  • 5759f17: Add an E2B client that binds a connection config once and exposes the resource surfaces off it, so a single process can talk to several API keys, domains or deployments. The classes it exposes are per-client subclasses of the real Sandbox/Volume/Template/Secret classes, so they behave exactly like the top-level ones — per-call options still win over the client's options, which win over the environment variables. The named top-level exports are unchanged and keep reading the environment.

    Nothing existing changes: Template is now the TemplateBase class made callable as a factory, so Template(...), the statics and instanceof keep working, and the default export is still Sandbox.

    import { E2B } from 'e2b'
    
    const { Sandbox, Volume, Template, Secret } = new E2B({
      apiKey: 'e2b_***',
      domain: 'e2b.dev',
    })
    
    const sandbox = await Sandbox.create()
    const volume = await Volume.create('my-volume')
    const exists = await Template.exists('my-template')
    await Template.build(Template().fromPythonImage('3'), 'my-env')
    await Secret.create('openai-api-key', 'sk-***')
    
    from e2b import E2B
    
    client = E2B(api_key="e2b_***", domain="e2b.dev")
    Sandbox, Volume, Template = client.Sandbox, client.Volume, client.Template
    Secret = client.Secret
    
    sandbox = Sandbox.create()
    volume = Volume.create("my-volume")
    exists = Template.exists("my-template")
    secret = Secret.create("openai-api-key", "sk-***")
    
    # Async variants are exposed too.
    AsyncSandbox = client.AsyncSandbox
    async_sandbox = await AsyncSandbox.create()
    

2.43.0

Minor Changes

  • f89f8c3: Add Secrets Management to the SDK. The Secret class (and AsyncSecret in Python) now manages E2B secrets: create and update store secret values (write-only — no read surface returns them), getInfo / get_info and the paginated list read metadata, exists and destroy are idempotent existence and lifecycle helpers, and fill formats the ${e2b.secrets.name} placeholder that the runtime resolves to the secret's current value.

Patch Changes

  • 2be6c12: Internal refactor: the template API operations resolve their connection config through a class-level hook, so a TemplateBase subclass can carry bound connection options. No behavior change — Template / AsyncTemplate keep reading config from per-call options and environment variables. In the Python SDK the terminal template operations (build, build_in_background, get_build_status, exists, alias_exists, assign_tags, remove_tags, get_tags) became classmethods, with signatures unchanged for callers.
  • 05aa03c: Add typed not-found errors for volumes: VolumeNotFoundError / VolumeNotFoundException (thrown when a volume is not found) and VolumePathNotFoundError / VolumePathNotFoundException (thrown when a path inside a volume is not found). All subclass the existing NotFoundError / NotFoundException, so existing catches keep working.

2.42.0

Minor Changes

  • 7af41e9: Refresh the MCP server types from the current MCP gateway catalog: 49 servers are new (n8n, neo4j, okta, temporal, proxmox, zscaler, the AWS Labs family, ...), 61 titles and 10 descriptions were rewritten, and 4 servers changed their options (awsDiagram, context7, neo4jCypher, onlyofficeDocspace).

    Six servers the catalog no longer publishes are gone from McpServer: postgres, root, tembo, flexprice, triplewhale, cdataConnectcloud. awsDiagram and context7 now require an option (outputDir and apiKey), so awsDiagram: {} and context7: {} stop type-checking, and onlyofficeDocspace is down to baseUrl and docspaceApiKey. The removals also narrow McpServerName, so Template().addMcpServer('postgres') stops compiling. The config is still passed to the gateway as written, so a dropped server can be kept by casting past the type — whether it starts is up to the gateway.

    import { Sandbox } from 'e2b'
    
    const sandbox = await Sandbox.create({
      mcp: {
        n8n: {
          apiKey: process.env.N8N_API_KEY!,
          apiUrl: 'https://n8n.example.com/api/v1',
        },
      },
    })
    

Patch Changes

  • 15bd48b: Omit autoPause from the create-sandbox request when no timeout lifecycle is configured, and omit autoPauseMemory unless keepMemory / keep_memory was chosen. Sending the SDK's local defaults for those fields was indistinguishable from an explicit choice, so the API could not tell "no preference" from a client choice and own its defaults. Explicit values are still always sent:

    import { Sandbox } from 'e2b'
    
    // No timeout lifecycle: autoPause is omitted, the API applies its default.
    await Sandbox.create()
    
    // Explicit action: autoPause: false / autoPause: true, as before.
    await Sandbox.create({ lifecycle: { onTimeout: 'kill' } })
    await Sandbox.create({ lifecycle: { onTimeout: 'pause' } })
    
    // Snapshot kind is only sent when keepMemory is set.
    await Sandbox.create({
      lifecycle: { onTimeout: { action: 'pause', keepMemory: false } },
    })
    
    from e2b import Sandbox
    
    # No timeout lifecycle: auto_pause is omitted, the API applies its default.
    Sandbox.create()
    
    # Explicit action: autoPause: false / autoPause: true, as before.
    Sandbox.create(lifecycle={"on_timeout": "kill"})
    Sandbox.create(lifecycle={"on_timeout": "pause"})
    
    # Snapshot kind is only sent when keep_memory is set.
    Sandbox.create(
        lifecycle={"on_timeout": {"action": "pause", "keep_memory": False}}
    )
    
  • 5367693: Omit autoResume from the POST /sandboxes request when lifecycle.autoResume / lifecycle["auto_resume"] is not configured, instead of sending the SDK's local default as { "autoResume": { "enabled": false } }. The API can now tell an unset preference from an explicit opt-out and own the default itself. Explicit values are unchanged on the wire.

    import { Sandbox } from 'e2b'
    
    // autoResume is left out of the request entirely — the API's default applies
    await Sandbox.create({ lifecycle: { onTimeout: 'pause' } })
    
    // an explicit choice is still sent as before
    await Sandbox.create({ lifecycle: { onTimeout: 'pause', autoResume: true } })
    
    from e2b import Sandbox
    
    # auto_resume is left out of the request entirely — the API's default applies
    Sandbox.create(lifecycle={"on_timeout": "pause"})
    
    # an explicit choice is still sent as before
    Sandbox.create(lifecycle={"on_timeout": "pause", "auto_resume": True})
    
  • 666241d: Run every persistent HTTP stack in the SDK on one shared pyqwest connection pool instead of four: the control-plane REST API, the envd HTTP API, the envd RPC clients, and the volume content API now all draw from e2b.api.client_sync/client_async, keyed on the three knobs that are fixed when a pyqwest transport is built — proxy, idle read bound, and HTTP version. reqwest pools per host internally, so one pool serves the API host and every per-sandbox host without interference — and since envd RPC and the envd HTTP API hit the same host, an active sandbox now needs a single HTTP/2 connection instead of one per stack. Streamed downloads keep a pool of their own, the only one carrying the idle read_timeout: reqwest's read timer runs during body send and TTFB, so on a shared pool it would cut off long uploads. No signature changes — get_transport and get_envd_transport keep the http2 parameter restored in 2.39.1, and the two are now the same pool per key rather than two.

  • 2daced6: Tag the package homepage and README links with UTM parameters (utm_source=npm/pypi) so registry traffic to e2b.dev is attributed correctly. No functional change.

2.41.0

Minor Changes

  • 6824cdf: Add network.egressProxy / network["egress_proxy"] for routing a sandbox's outbound TCP through a SOCKS5 proxy you operate ("bring your own proxy"). Tunneling happens on the host after the allowOut / denyOut lists are evaluated, so nothing runs inside the sandbox and code running there can neither see the proxy nor route around it. UDP-based traffic — DNS and QUIC/HTTP3 — is not tunneled.

    import { Sandbox } from 'e2b'
    
    const sandbox = await Sandbox.create({
      network: {
        egressProxy: {
          address: 'proxy.example.com:1080',
          username: 'proxy-user',
          password: 'proxy-password',
        },
      },
    })
    
    from e2b import Sandbox
    
    sandbox = Sandbox.create(
        network={
            "egress_proxy": {
                "address": "proxy.example.com:1080",
                "username": "proxy-user",
                "password": "proxy-password",
            },
        },
    )
    

    It combines with the rest of the network configuration — here everything except api.example.com is denied, and the traffic that is allowed goes through your proxy:

    await Sandbox.create({
      network: {
        allowOut: ['api.example.com'],
        denyOut: ({ allTraffic }) => [allTraffic],
        egressProxy: { address: 'proxy.example.com:1080' },
      },
    })
    
    Sandbox.create(
        network={
            "allow_out": ["api.example.com"],
            "deny_out": lambda ctx: [ctx.all_traffic],
            "egress_proxy": {"address": "proxy.example.com:1080"},
        },
    )
    

    updateNetwork / update_network sets or replaces the proxy on a sandbox that is already running, with no restart. The update replaces the whole configuration instead of merging into it, so an update that leaves the proxy out stops tunneling — repeat it in every update that should keep it.

    // Start tunneling on the running sandbox
    await sandbox.updateNetwork({
      allowOut: ['api.example.com'],
      denyOut: ({ allTraffic }) => [allTraffic],
      egressProxy: { address: 'proxy.example.com:1080' },
    })
    
    // Stop tunneling: an update without egressProxy clears it
    await sandbox.updateNetwork({})
    
    # Start tunneling on the running sandbox
    sandbox.update_network({
        "allow_out": ["api.example.com"],
        "deny_out": lambda ctx: [ctx.all_traffic],
        "egress_proxy": {"address": "proxy.example.com:1080"},
    })
    
    # Stop tunneling: an update without egress_proxy clears it
    sandbox.update_network({})
    

    getInfo / get_info reports the proxy the sandbox's egress is currently tunneled through. The password is never returned, so the returned SandboxEgressProxyInfo does not have the field at all:

    const info = await sandbox.getInfo()
    console.log(info.network?.egressProxy)
    // { address: 'proxy.example.com:1080', username: 'proxy-user' }
    
    info = sandbox.get_info()
    print((info.network or {}).get("egress_proxy"))
    # {'address': 'proxy.example.com:1080', 'username': 'proxy-user'}
    

    Egress fails closed: when the proxy is unreachable or does not speak SOCKS5, outbound connections fail rather than falling back to a direct connection. The address is validated server-side when the sandbox is created — a rejected create leaves nothing behind. Available on E2B Cloud and in BYOC deployments; a sandbox that names a proxy on a deployment built from the open source e2b-dev/infra repository is rejected as unsupported.

Patch Changes

  • 02ba746: Raise the h2 floor to >=4.4.1 so it can no longer resolve to a version affected by CVE-2026-71554, where a duplicate Host header is forwarded to the consuming application and becomes a request smuggling primitive once HTTP/2 is downgraded to HTTP/1.1.

  • e2eebd5: Fix URL encoding of namespaced template names and aliases in the Python SDK.

    The endpoints that take a template ID also accept a template name, and names may be namespaced (e.g. namespace/name). The SDK interpolated them into the request path without encoding, so a call like Template.exists("namespace/name") hit /templates/aliases/namespace/name instead of /templates/aliases/namespace%2Fname — the slash split the route rather than staying inside one path segment. Every method that takes a template ID or name, an alias, or a snapshot ID in the path — Template.exists / alias_exists, get_tags, the build/upload/status calls, and Sandbox.delete_snapshot (whose snapshot IDs are namespace/name:tag) — now percent-encodes the value, matching the JavaScript SDK (which already encodes path parameters via encodeURIComponent).

    from e2b import Template, Sandbox
    
    # Namespaced templates now resolve correctly
    Template.exists("my-team/my-template")
    Template.get_tags("my-team/my-template")
    
    # Namespaced snapshots can now be deleted
    Sandbox.delete_snapshot("my-team/my-snapshot:default")
    

2.40.0

Minor Changes

  • 6248b12: Remove the deprecated accessToken / access_token option and its E2B_ACCESS_TOKEN environment fallback. E2B access tokens are no longer accepted for API authentication, so the SDKs no longer resolve one or send it as an Authorization: Bearer header — requests authenticate with the API key alone.

    If you were relying on the option to send a bearer token to a custom deployment, pass the header directly, which is what the deprecation notice already pointed to:

    // Before
    const sandbox = await Sandbox.create({ accessToken: token })
    
    // After
    const sandbox = await Sandbox.create({
      apiHeaders: { Authorization: `Bearer ${token}` },
    })
    
    # Before
    config = ConnectionConfig(access_token=token)
    
    # After
    config = ConnectionConfig(api_headers={"Authorization": f"Bearer {token}"})
    

    Note that Sandbox.envd_access_token / traffic_access_token are unrelated per-sandbox tokens and are unaffected.

2.39.1

Patch Changes

  • 0d507cd: Restore the http2 parameter on get_transport and get_envd_transport, which the pyqwest migration dropped in 2.38.0. http2=False again returns a transport pinned to HTTP/1.1, on its own connection pool.

2.39.0

Minor Changes

  • 07eb9be: Allow a network rule's transform to be a callback, so a workload identity token from the iam option can be injected into egress requests without the SDK ever seeing its value. The callback receives placeholder strings that the egress proxy resolves per request — iam.tokens.aws is ${e2b.identity.tokens.aws} on the wire — and referencing a token that is not registered in iam.tokens fails with InvalidArgumentError / InvalidArgumentException instead of silently sending a placeholder no token will ever replace.

    updateNetwork / update_network accepts the same callbacks, but its payload carries no iam config, so token names cannot be checked there and every name resolves to its placeholder.

    Token names are validated where they are registered and again before they are interpolated: a name cannot be empty or contain {, } or control characters, since the proxy reads a placeholder up to its first } and a brace in the name would resolve a different token than the one referenced.

    import { Sandbox, Secret } from 'e2b'
    
    const sandbox = await Sandbox.create({
      iam: {
        tokens: {
          aws: Secret.iamToken({
            audience: 'sts.amazonaws.com',
            tokenType: 'JWT-SVID',
          }),
        },
      },
      network: {
        allowOut: ({ rules }) => [...rules.keys()],
        rules: {
          'api.internal.example.com': [
            {
              transform: ({ iam }) => ({
                headers: { Authorization: `Bearer ${iam.tokens.aws}` },
              }),
            },
          ],
        },
      },
    })
    
    from e2b import Sandbox, Secret
    
    sandbox = Sandbox.create(
        iam={
            "tokens": {
                "aws": Secret.iam_token(audience="sts.amazonaws.com", token_type="JWT-SVID"),
            },
        },
        network={
            "allow_out": lambda ctx: list(ctx.rules.keys()),
            "rules": {
                "api.internal.example.com": [
                    {
                        "transform": lambda ctx: {
                            "headers": {"Authorization": f"Bearer {ctx.iam.tokens['aws']}"},
                        },
                    },
                ],
            },
        },
    )
    
  • 64b25bb: Add the iam option to Sandbox.create for configuring sandbox workload identity, and a Secret class with an iamToken / iam_token method for defining the workload tokens. Passing a non-empty tokens map (name → { audience, tokenType }) enables workload identity for the sandbox:

    import { Sandbox, Secret } from 'e2b'
    
    const sandbox = await Sandbox.create({
      iam: {
        tokens: {
          aws: Secret.iamToken({
            audience: 'sts.amazonaws.com',
            tokenType: 'JWT-SVID',
          }),
        },
      },
    })
    
    from e2b import Sandbox, Secret
    
    sandbox = Sandbox.create(
        iam={
            "tokens": {
                "aws": Secret.iam_token(audience="sts.amazonaws.com", token_type="JWT-SVID"),
            },
        },
    )
    

    Plain { audience, tokenType } objects ({"audience": ..., "token_type": ...} dicts in Python) are accepted as token values too.

Patch Changes

  • 11912ff: Build the envd HTTP API client once per sync Sandbox and share it across the filesystem, commands, and PTY modules, which now receive it instead of each constructing their own — matching AsyncSandbox. No behavior change: the pyqwest transport underneath is already cached process-wide per (proxy, for_streaming), so the separate clients shared one connection pool either way. Filesystem still builds the streaming sibling client whose transport carries the idle read timeout, in both flavors.

2.38.0

Minor Changes

  • b048369: Move the envd HTTP API client (sandbox file transfers, health checks) onto pyqwest via its httpx-compatible transport adapter. envd RPC already runs on pyqwest through connectrpc, so all sandbox traffic now shares one HTTP stack built from the same transport pieces (with separate connection pools per use).

    The per-thread (sync) and per-loop (async) envd httpx clients are gone: the pyqwest transports are thread-safe and loop-independent, so a single client per module serves all threads and event loops.

    Timeout semantics through the adapter:

    • Streamed downloads (files.read(format="stream")): a request_timeout set explicitly for the call is the deadline for the whole transfer — by default the transfer is unbounded in total, as before. A stalled stream is reclaimed by a 60-second idle read timeout that resets on every chunk. stream_idle_timeout keeps working on the async client (applied per read); the sync client cannot interrupt a blocking read, so it relies on the transport-wide idle bound and now ignores the parameter.
    • Uploads: a buffered upload is bounded by request_timeout as a whole-request deadline, and a streamed (file-like) upload carries no client-side timeout (a stalled one is bounded server-side by envd's idle read timeout) — both matching the JS SDK.
    • Non-streamed reads (files.read() as text or bytes) and buffered uploads are bounded by request_timeout for the whole transfer (default 60 seconds), where the previous transport bounded each socket operation and left total duration unbounded. Reading or writing a file too large to transfer inside the deadline now raises httpx.ReadTimeout — pass a larger request_timeout (or 0 to disable), or use format="stream"/file-like data, for large transfers.

    E2B_MAX_CONNECTIONS is no longer read: it configured httpx's global connection cap, and the last transport that took one is gone (reqwest has no counterpart — it does not cap concurrent connections). E2B_KEEPALIVE_EXPIRY and E2B_MAX_KEEPALIVE_CONNECTIONS keep tuning the pools.

  • a874ced: Move the REST API client (sandbox lifecycle, listing, templates, volumes control plane) onto pyqwest (Rust reqwest/hyper) via its httpx-compatible transport adapter, replacing the httpx-native HTTPTransport/AsyncHTTPTransport. The generated httpx client API is unchanged — only the transport underneath is swapped — so logging event hooks, headers, and redirect handling (follow_redirects, response.history) behave as before.

    One timeout semantics change: through the adapter, request_timeout is a deadline for the whole API call, where the previous transports applied it to each phase (connect, read, write) separately — a slow request could exceed it in total. For the REST API's small JSON exchanges this tightening is what request_timeout reads as promising; 0 still disables it.

    Because pyqwest transports are thread-safe and loop-independent (I/O runs on a Rust runtime), the API connection pool is now shared process-wide per proxy, instead of one pool per thread (sync) or per event loop (async), and ApiClient no longer maintains per-thread/per-loop httpx client caches — a single httpx client serves all threads and event loops. Connection-establishment failures are retried with backoff (E2B_CONNECTION_RETRIES, default 3), matching the connect-only retries of the previous transports. Timeouts keep raising httpx.ReadTimeout (an httpx.TimeoutException), as before, whether they fire while waiting for the response head or while reading the response body, and connection, network, and protocol failures keep raising their httpx counterparts (httpx.ConnectError, httpx.ReadError, httpx.RemoteProtocolError).

    proxy for API calls takes a URL string (e.g. proxy="http://user:pass@localhost:8030", scheme http, https, socks5, or socks5h), an httpx.URL, or an httpx.Proxy — including its credentials (sent as Proxy-Authorization) and any headers configured for the proxy. The one httpx.Proxy option pyqwest cannot express, a per-proxy ssl_context, raises InvalidArgumentException rather than being silently dropped.

    Low-level HTTP logs stay available: where enabling the httpcore logger used to show connection-level detail, pyqwest logs one line per request on the pyqwest.access logger and request lifecycle records on pyqwest, both at DEBUG and off unless enabled:

    import logging
    
    logging.basicConfig()
    logging.getLogger("pyqwest.access").setLevel(logging.DEBUG)
    # DEBUG pyqwest.access - HTTP Request: POST https://api.e2b.app/sandboxes "HTTP/2 201 Created"
    

    The SDK's own logger option is unchanged and independent of these.

    envd traffic is not affected: RPC (commands, PTY, filesystem watch) already runs on pyqwest via connectrpc, and the envd HTTP API (file transfers, health checks) keeps its httpx transports.

  • b3a7c9f: Move template build-context uploads (to S3 presigned URLs) onto pyqwest via its httpx-compatible transport adapter. Content-Length framing for the streamed archive body is preserved (S3 rejects chunked transfer encoding), and redirects stay with the httpx client instead of being followed inside the transport. The 1-hour upload timeout now bounds the entire upload rather than each socket operation, and verify_ssl=False on the client is no longer honored for uploads (pyqwest has no insecure-TLS option).

  • 458c2c4: Move the volume content client (Volume/AsyncVolume file operations) onto pyqwest via its httpx-compatible transport adapter, the same stack the REST API client uses. The connection pool is shared process-wide per proxy instead of one pool per thread (sync) or per event loop (async), and connection-establishment failures are retried with backoff (E2B_CONNECTION_RETRIES, default 3), as before.

    For streamed volume reads (Volume.read_file(format="stream")), a stalled stream is by default bounded by a transport-wide idle read timeout of 60 seconds that resets on every chunk (still surfaced as httpx.ReadTimeout; matches the JS SDK's default stream idle timeout). AsyncVolume.read_file keeps honoring an explicit stream_idle_timeout per read (including 0 to disable); the sync client ignores it — it cannot interrupt a blocking read. Passing request_timeout to a streamed read now bounds the whole transfer rather than individual socket operations.

    The same whole-transfer semantics apply to non-streamed operations: read_file(format="text"/"bytes") and uploads are bounded by request_timeout as a total deadline (default 1 hour for file content operations), where the previous transports bounded each socket operation and left total duration unbounded. Pass a larger request_timeout (or 0 to disable) for very large transfers on slow links.

Patch Changes

  • cab27aa: Kill newly created sandboxes when MCP gateway startup fails. The failure now surfaces as SandboxError (JS) / SandboxException (Python) with a Failed to start MCP gateway: <stderr> message instead of a bare command exit error.

2.37.1

Patch Changes

  • 88f41f3: Align ANSI stripping of template build log messages across both SDKs. The Python SDK's strip_ansi_escape_codes now ports the JS SDK's stripAnsi regex: OSC sequences (hyperlinks, window titles) are matched non-greedily up to the first string terminator — including sequences spanning newlines — and CSI sequences are stripped without requiring a terminator. Both implementations additionally strip the remaining ECMA-48 string controls (DCS/Sixel, SOS, PM, APC) through their string terminator so control payloads no longer leak into cleaned logs.
  • 998e560: Relax the Python SDK's wcmatch requirement from >=10.1,<11 to >=10.1,<12 so e2b can be installed alongside packages that already require wcmatch>=11 (for example deepagents>=0.7.0), which previously failed to resolve. The SDK only calls glob.glob() with GLOBSTAR | DOTMATCH for template context matching; wcmatch 11.0's single breaking change affects translate() callers using extended-glob capture groups, so it is a no-op here. The template glob test suite passes against 10.1, 10.2.1 and 11.0.

2.37.0

Minor Changes

  • 2821fb0: Route volume content requests to a team's custom (BYOC) cluster. When a team is connected to a custom cluster, the volume create and get endpoints now return that cluster's domain, and the SDK uses it as the destination for volume content requests instead of the default api.<E2B_DOMAIN> host. Teams on the default cluster are unaffected and keep their configured domain.

2.36.0

Minor Changes

  • 1504fbc: Add fromFedoraImage, fromAlpineImage, and fromArchImage base-image helpers to the Template builder (from_fedora_image, from_alpine_image, from_arch_image in the Python SDK), alongside the existing fromUbuntuImage/fromDebianImage/etc. Templates can now start from Fedora, Alpine, and Arch base images (the orchestrator identifies the distro from /etc/os-release). Fedora and Alpine default to pinned tags (fedora:44, alpine:3.24) so builds stay reproducible; Arch defaults to latest because it is a rolling release and provisioning runs pacman -Syu regardless.

Patch Changes

  • 6733f36: Align the Python SDK's from_fedora_image and from_alpine_image defaults with the JS SDK: fedora:44 and alpine:3.24, replacing fedora:42 (end-of-life, so its repositories leave the normal mirror network and provisioning can fail) and alpine:3.22. Callers that omit the variant now get the same base image in both SDKs, and both tags are the ones the orchestrator's distro build tests cover. Also corrects the JS TemplateFromImage type docs, which still named the old defaults.
  • 45d2679: Regenerate e2b/sandbox/mcp.py with datamodel-code-generator 0.64.0: the MCP server option types now use builtin generics (list[str], dict[str, Any]) and are closed TypedDicts, mirroring the spec's additionalProperties: false. Raises the typing-extensions floor to >=4.10.0, the first release accepting PEP 728's closed.
  • ee0ad25: Update snapshot docstrings to use project terminology instead of team (e.g. "my-project/my-snapshot", project slug)