Skip to content

Releases

sdk-v1.11.0

Choose a tag to compare

Features

  • api: add Claude Code managed settings to the Admin API
  • api: add the sandbox_egress_restricted error to Admin API plugin methods

Performance Improvements

  • client: speed up streaming of large events

Chores

  • docs: explain how the input-token rate limit types differ
  • docs: name Claude Security in the usage and cost report descriptions
  • docs: note that a Google Cloud external key's provider and key_name can't be changed

sdk-v1.10.0

Choose a tag to compare

Features

  • api: add beta support for the billing_policy request parameter
  • api: add contract_spend_limit_reached error code to messages
  • api: add error codes for linked billing refusals
  • api: add the user_profile_billing_not_authorized error code
  • api: add webhook events for external key failure and recovery

Bug Fixes

  • tools: keep the session tool runner's maxIdleMs countdown running when a session thread terminates

    If the runner fails to read the session's event history after a reconnect, a termination that then arrives on the event stream still cancels the countdown, until the next idle or successful read.

  • tools: stop EnvironmentWorker from running tools for a terminated session

    For a terminated session, the worker now force-stops the work item right after it retrieves the session, before it downloads skills or memory stores. It no longer throws SessionMemoryError for such a session. The API reports an archived session as terminated.

  • tools: stop the session tool runner from running the tool calls of a terminated session thread

    The runner also stops waiting for a confirmation of these calls and no longer yields them, including a denied call (confirmation: 'deny') and a call to a tool that it does not own. A call that the runner starts before it reads the session thread's session.thread_status_terminated event still runs to the end, and the runner yields it with posted: false when the API refuses the result.

Chores

  • api: rename error data models, deprecating the old names

  • docs: clarify which AWS KMS keys a workspace can use on Claude Platform on AWS

  • docs: clarify which credential work get, update and list accept

  • docs: clarify which uploads a plugin pin holds back

  • docs: note that a plugin scan's reason also covers errored scans

    content_scan.reason on an errored scan now says why there is no verdict. It is not a security finding, so code that treats any reason as a concern should also check status.

sdk-v1.9.0

Choose a tag to compare

Features

  • api: add source and error codes to the RBAC Role read endpoints
  • api: add Update Bundle, which sets a Claude Tag Bundle's instructions, and return instructions on Bundle

Bug Fixes

  • helpers: keep $schema out of the transformed schema description

  • streaming: reject characters that aren't JSON in streamed tool input

    The lenient parser for streamed tool input used to skip characters and words that aren't JSON, such as a comment, and to join numbers such as 12 3, so it could return a value that the text doesn't contain. It now rejects such text, and a stream with such tool input fails with an error, as it already did for a bad escape or a missing comma. Valid and truncated JSON parse as before.

  • tools: bound the grep tool's search at 60s where ripgrep is missing, and keep its regex off the main thread

  • tools: compare directory names exactly when confining file paths on Windows

  • tools: fully resolve file tool paths that pass through a dangling symlink

  • tools: ignore ripgrep config files in the agent toolset grep tool

  • tools: leave symlinks that point outside /memories out of listings

  • tools: normalize file modes of extracted skill files

    Extracted files now get 0644, or 0755 if any execute bit is set, in 0700 directories, whatever the archive recorded. The system tar must accept --no-same-owner and --no-same-permissions.

  • tools: pass over a tar or unzip inside the directories the agent toolset's file tools can write

    setupSkills takes tar, unzip and a helper that tar starts, such as gzip, only from the PATH entries outside workdir and allowedRoots, and passes over a link that leads inside them. A host whose only tar or unzip is inside those directories can no longer download skills until one is installed outside. Only when no PATH entry is outside them, as with a workdir of /, is all of PATH still searched. A tar or unzip that fails now rejects with an AnthropicError.

  • tools: pass over an rg inside the directories the agent toolset's file tools can write

    grep passes over an rg that resolves to somewhere inside workdir or allowedRoots. It uses the next rg on PATH, or the built-in search when there is none. With a workdir that holds every rg on the host, such as /, every search is a built-in search. rg, and the tar and unzip that extract skills, no longer get the ANTHROPIC_* variables of the process, like the shell of bash.

  • tools: read memory files only through the path that was validated

  • tools: stop looking for rg, tar and unzip in the current directory

  • tools: stop the grep tool's built-in search at a NUL byte in a file

    The built-in search of grep reads no further into a file once it meets a NUL byte, much as ripgrep does. A very large binary file with no line break used to end the process with an uncaught RangeError.

Chores

  • api: deprecate old beta header values

    The API still accepts these values, but a later release removes them from the SDKs. Delete prompt-caching-2024-07-31, pdfs-2024-09-25, token-counting-2024-11-01, token-efficient-tools-2025-02-19, extended-cache-ttl-2025-04-11, cache-diagnosis-2026-04-07 and ce-user-management-2026-07-13 from your requests. files-api-2025-04-14 and skills-2025-10-02 select the beta response shapes, so follow the Files API and Skills migration guides before you delete them.

    In C#, the AnthropicBeta enum's members have new numbers, so code compiled against an earlier SDK version sends a different beta value, with no error, until you rebuild it.

  • api: stop sending an old beta header on batch requests

    beta.messages.batches no longer sends message-batches-2024-09-24, which the API doesn't need. If your client sets a default anthropic-beta header, these requests now send it, where in some SDKs the old value replaced it. The API still accepts the old value. If you set it yourself, delete it: a later release removes it from the SDKs.

  • docs: add an example value for the workspace name

  • docs: clarify when the auto permission policy denies a tool call

  • docs: clarify which workspaces an external key can be attached to

  • docs: recommend multiagent_20261001 over the coordinator type

  • docs: say when organization_id is required on plugin reads

vertex-sdk-v1.1.3

Choose a tag to compare

Bug Fixes

  • vertex: send a ; in the model as %3B
  • vertex: throw an AnthropicError for a project ID, region or model such as ;x, .;x or ..;x

sdk-v1.8.0

Choose a tag to compare

Features

  • api: add Claude Tag metrics to the analytics summaries and users reports
  • api: add reply_mode to Claude Tag Slack channel settings

Chores

  • docs: clarify display_name for RBAC Roles created by Anthropic
  • docs: clarify plugin delete and skill name conflict descriptions
  • docs: clarify when to omit user_profile_id on external keys
  • docs: note that Claude Tag settings accept self-hosted runner pools

Documentation

  • examples: show multiagent when creating a Managed Agents agent

google-cloud-sdk-v1.1.4

Choose a tag to compare

Bug Fixes

  • google-cloud: throw an AnthropicError for a project, location or workspace ID such as ;x, .;x or ..;x

bedrock-sdk-v1.1.2

Choose a tag to compare

Bug Fixes

  • bedrock: throw an AnthropicError for a model such as ;x, .;x or ..;x

sdk-v1.7.0

Choose a tag to compare

Features

  • api: accept user_profile.program_status_updated on organization webhook endpoints
  • api: add claude-fable-5-2 and claude-mythos-5-2
  • api: add parent organization filtering and grouping to the usage and cost reports
  • api: add the sandbox_egress_restricted error to Admin API write methods
  • api: add the user_profile.program_status_updated webhook event

Bug Fixes

  • client: include the body of a failed response in debug logs
  • helpers: keep const and enum in structured output schemas

Chores

  • api: remove some error detail types and update descriptions
  • docs: say when a work heartbeat does not compare its value
  • internal: add this repository's review rules

Documentation

  • client: list error code types in api.md
  • client: say where the default logger writes and how to log to stderr
  • helpers: import from zod/v4 in the examples, for Zod 3.25 users

Refactors

  • client: declare the name of every AnthropicError subclass the same way

sdk-v1.6.0

Choose a tag to compare

Features

  • api: add error code for ramp limit errors
  • api: add error codes for External Key validation failures
  • api: add support for signing organization members out everywhere

Bug Fixes

  • client: set the name of each error to its class name
  • client: set the name of WorkloadIdentityError

Chores

  • docs: describe the type and message of API errors
  • internal: generate the OpenTelemetry span attribute tables
  • internal: tell contributors when a runtime type check belongs

Documentation

  • add the rules the SDKs share to CLAUDE.md
  • client: describe the type, message, code and details of API errors
  • client: list each method's errors without their descriptions

vertex-sdk-v1.1.2

Choose a tag to compare

Bug Fixes

  • vertex: percent-encode the project ID, region and model in the request path, so a / in the model can't move the request outside /publishers/anthropic/models/, and throw an AnthropicError for an empty model or a project ID or region of . or ..