Skip to main content

Claude Agent SDK

Changelog

Visit WebsiteGitHub

Anthropic Agent SDK, showing TypeScript releases

This page shows at most the 60 most recent releases from official sources. It does not claim to be a complete version history.

Latest Version:v0.3.296
Shown Releases:76
Source:GitHub Releases
RSS

Latest Releases (76)

v0.3.296New

v0.3.296

October 9, 2026

  • •Added claude_code_version to the initialize control response so clients know the CLI version before the first turn
  • •Added autoCompactWindow to AgentDefinition, so a subagent can auto-compact earlier than the main conversation's window
  • •Fixed the sandbox option discarding an inline settings.sandbox block: the two now merge, the option's values win, and deny and credential lists from both sides are combined
  • •Changed the default limit on MCP tool descriptions sent up front and on MCP server instructions, in-process SDK servers included, from 2,048 to 4,096 characters
  • •Changed permission answers that arrive after a restart to be parsed like live ones: over 4,096 updatedPermissions entries count as a denial, and a malformed list is ignored as a whole
  • •Updated to parity with Claude Code v2.1.296
v0.3.295New

v0.3.295

October 8, 2026

  • •Added overageEnabled to SDKRateLimitInfo: on usage-limit warnings, whether the account has extra usage turned on
  • •Fixed an MCP server added with mcp_set_servers sometimes keeping an outdated tool list when another server in the same request connected later
  • •Fixed assistant text blocks losing their citations in streamed responses
  • •Fixed Claude being told "disabled by the user" about an in-process MCP server that was switched off with toggleMcpServer()
  • •Fixed rate_limit_event with status allowed_warning omitting overageStatus, overageResetsAt and overageDisabledReason, which allowed and rejected events already carry
  • •Changed MCP tool descriptions the model loads through tool search, including those from in-process SDK servers, to be cut at 16,384 characters instead of 2,048
  • •Changed a permission answer whose updatedPermissions holds over 4,096 updates, rules and directories to count as a denial
  • •Changed tool_result_meta[].non_execution_kind to be absent when a mod withheld the result of a tool that ran without error; it was permission-rule
  • •Changed pasted_content on user messages to be ignored when it holds over 1,000 entries and blocks in all
  • •Changed how the SDK passes option values to Claude Code: the values of its named options are now sent in the same argument as their flag (--flag=value)
  • •Changed tool_use_result for MCP Apps tools: structuredContent or _meta over 8,388,608 JSON characters is left off, and _meta.ui with no ui:// resource no longer lifts the ordinary caps
  • •Changed tool_use_result for MCP tools: when a long result is saved to a file, or cut because it could not be, the text kept past 500,000 characters is dropped, with a note of how much
  • •Changed tool_use_result for MCP tools: a result _meta over 1,048,576 JSON characters is left off, except for SDK-server tools; so is text blocks' _meta that totals over 65,536
  • •Updated to parity with Claude Code v2.1.295
v0.3.294New

v0.3.294

October 8, 2026

  • •Updated to parity with Claude Code v2.1.294
v0.3.293New

v0.3.293

October 7, 2026

  • •Added an optional subagent_type to background_tasks_changed task entries, so hosts can name each subagent's type without pairing with task_started
  • •Updated to parity with Claude Code v2.1.293
v0.3.292New

v0.3.292

October 6, 2026

  • •Added agent_id to the assistant and user messages a subagent produces; it equals the task_id on that subagent's task events and stays the same when the subagent is resumed
  • •Added parent_task_id to task_started events and background_tasks_changed entries, naming the subagent task that launched a task
  • •Added run_id to background task events and origin.runId to task notifications, so a host can tell a resumed task's runs apart
  • •Added typed sections and notes to the ListAgents tool's tool_use_result, so hosts can list agents without parsing its text
  • •Added recipient_kind to SendMessage tool_use inputs on assistant messages, and their top-level approve is now always a boolean, so hosts no longer infer them from the raw input
  • •Added two optional fields to the ReadNotifications tool output: read_at, when the tool read the session's queued notifications, and arrived_at on each notification, when it reached the session
  • •Fixed subagents with a declared auto permission mode having their tool calls decided by the auto-mode classifier instead of the canUseTool callback when auto mode is unavailable
  • •Changed background_tasks_changed for a finishing background task to arrive after that task's task_updated and task_notification
  • •Changed canUseTool options to carry suppressAlwaysAllowRule: true for connector tools that an organization set to require approval, so hosts can hide their "always allow" choice
  • •Updated to parity with Claude Code v2.1.292
v0.3.291New

v0.3.291

October 6, 2026

  • •Updated to parity with Claude Code v2.1.291
v0.3.290New

v0.3.290

October 6, 2026

  • •Added an optional offset field to the WebFetch tool input for reading on through long pages
  • •Fixed failed Claude in Chrome tool calls dropping the result's _meta: when the result has one, tool_use_result is now { content, _meta }, as for other MCP tools
  • •Fixed deny and ask rules missing toolAliases tools when written with a wildcard (Bash*), a match-all pattern (Bash(**), also on the mapped name), or an input field also in disallowedTools
  • •Fixed a user message sent again under the same uuid being replayed (--replay-user-messages) while the first copy still waited for its turn; its replay now comes when a turn takes it
  • •Fixed includePartialMessages streams leaving a message without message_stop when the stream was cut, interrupted or fell back to non-streaming
  • •Changed user_message_uuid, user_message_uuids and resume_reason to also be set on a turn a restarted worker resumes from a permission answer or from tool calls still in flight
  • •Updated to parity with Claude Code v2.1.290
v0.3.289

v0.3.289

October 4, 2026

  • •Updated to parity with Claude Code v2.1.289
v0.3.288

v0.3.288

October 2, 2026

  • •Updated to parity with Claude Code v2.1.288
v0.3.287

v0.3.287

October 1, 2026

  • •Added optional remote-session latency fields (first_text_post_queue_wait_ms, first_text_post_queued_behind) to the success result message
  • •Fixed includePartialMessages streams sending a cut-short reply's message_stop late or never, so apps could show the reply as still in progress
  • •Fixed the error for a revoked claude.ai login, which now reads "Failed to authenticate: OAuth token revoked" instead of a generic or "does not have access" message
  • •Fixed a tool call to an in-process MCP server being left waiting after toggleMcpServer() disabled the server or setMcpServers() removed it
  • •Fixed commands_changed arriving before init, or twice, at session start: the initialize response now includes commands registered at startup
  • •Changed tool_use_result for a WebFetch or WebSearch call that steps aside for a priority "now" message to { detachedToolCall: true }; the result follows in a later turn
  • •Changed tool_use_result for MCP tools: structuredContent over 1,048,576 JSON characters is left off and structuredContentOmitted: true set, except for SDK-server and MCP Apps tools
  • •Updated to parity with Claude Code v2.1.287
v0.3.286

v0.3.286

September 30, 2026

  • •Added the initialize response field sdk_mcp_manifests_parked and the system/init capabilities sdk_mcp_manifests and sdk_mcp_tools_list_changed
  • •Fixed foreground subagents sometimes not receiving the task-tracking tools listed in tools or allowedTools
  • •Fixed an SDK MCP server listing no tools when one tool's schema cannot be converted to JSON Schema; that tool is now left out with a warning naming it
  • •Fixed toggleMcpServer() not disconnecting, and failing to re-enable, an in-process MCP server created with createSdkMcpServer(), except one named claude-in-chrome
  • •Changed a person's priority now message to move running shell commands, agents and MCP calls to the background and join the running turn instead of stopping it
  • •Changed the TypeScript Agent SDK to leave an omitted permissionMode to Claude Code, so a settings defaultMode now applies and, on third-party providers or with telemetry off, the session starts in auto mode like claude -p; pass permissionMode: 'default' for manual approvals
  • •Updated to parity with Claude Code v2.1.286
v0.3.285

v0.3.285

September 29, 2026

  • •Added an optional read.title (the artifact's stored title) to the Artifact tool's structured read result
  • •Added provider_not_allowed to startup_failure_reason and allowedProviders to the Settings type
  • •Fixed sessions with CLAUDE_CODE_FORK_SUBAGENT=1: a subagent's own Agent call now runs in the foreground, so the subagent gets the child's result
  • •Fixed getSessionMessages() leaving out a message sent while Claude was working when the process stopped before the reply, or when another prompt followed it with no reply in between
  • •Fixed toggleMcpServer(name, false) leaving the connection open for a server that had not yet connected when the session started, or whose config was edited after it connected
  • •Fixed rewind_conversation leaving a backgrounded MCP tool call running after the message that started it was removed
  • •Changed Bash/PowerShell timeout to bound a run_in_background command (was ignored; default 30 min, max 2 h); the stopped task notification for a stop at that limit says why
  • •Changed getSubagentMessages() to also return the messages a subagent read while it ran, such as a message sent to it; offset and limit count these rows
  • •Updated to parity with Claude Code v2.1.285
v0.3.284

v0.3.284

September 28, 2026

  • •Added a renamed skill's directory name to its SlashCommand aliases when the name is plain and Claude Code ships no command by that name
  • •Added applied.ultracodeAvailable and applied.ultracodeRequested to getSettings(): whether this session can run Ultracode, and whether it is requested, independent of whether it is in effect
  • •Fixed forkSession() copies reading back the wrong history when cut at a progress row or fork briefing after a rewind, or when the session was compacted with some messages kept
  • •Fixed getSessionMessages() leaving out messages from other agents, sessions and channels that the CLI transcript shows
  • •Fixed { decision: 'block' } returned by Elicitation and ElicitationResult hook callbacks being ignored; it now declines the MCP elicitation
  • •Fixed query() closing stdin before a follow-up turn woken by a finished background agent, which failed that turn's hooks, canUseTool and SDK MCP calls with "Stream closed"
  • •Changed the first turn to still wait up to 2s for connecting MCP servers named in allowedTools or by an mcp_tool hook, even with CLAUDE_CODE_MCP_STARTUP_WAIT_MS set to 0
  • •Changed applyFlagSettings({ ultracode: true }) to keep the current effort level instead of switching to xhigh; an effortLevel that changes the level without ultracode turns Ultracode off
  • •Updated to parity with Claude Code v2.1.284
v0.3.283

v0.3.283

September 25, 2026

  • •Added plugin_errors to the SDKSystemMessage type (system/init), including path for a --plugin-dir entry that did not load
  • •Fixed getSessionMessages() returning, and forkSession() copying, a rewound-away branch when the newest branch ends at a meta row or a local command's rows
  • •Changed stream-json output to include warnings and notices raised during a turn as system/informational messages; it previously dropped them
  • •Changed the set_max_thinking_tokens control request: omitting max_thinking_tokens now leaves the session's thinking budget unchanged; send null to reset it to the session default
  • •Updated to parity with Claude Code v2.1.283
v0.3.282

v0.3.282

September 24, 2026

  • •Added support for strictKnownMarketplaces and blockedMarketplaces in host-supplied managedSettings: the allowlist applies only where admin policy sets none; the blocklist adds to the admin's
  • •Added @anthropic-ai/claude-agent-sdk/core, a smaller entry point for apps that bundle the SDK (query, MCP tool helpers, session mutations, resolveSettings); it uses your installed zod and MCP SDK
  • •Added prewarm() and SpareProcess.claim() (alpha): start a Claude Code process before its session is known and bind it to a folder and its per-session options later
  • •Fixed readMcpResource() relaying content _meta keys under the CLI-reserved com.anthropic/ prefix; they are now dropped, as for tool results
  • •Updated to parity with Claude Code v2.1.282
v0.3.281

v0.3.281

September 23, 2026

  • •Added optional trigger, user_message_uuid and timestamp fields to the conversation_reset message so clients can tell what reset the conversation, match a /clear to its message, and show when the reset happened
  • •Fixed permission and dialog callbacks still being invoked for requests that arrived after close()
  • •Fixed control requests issued after a query closed hanging or leaking, and permission prompts that could not be cancelled after a cancelled request was redelivered
  • •Fixed session_state_changed staying at requires_action after an overlapping permission prompt and sandbox network-access prompt were both answered
  • •Improved SDK package size and load time: sdk.mjs no longer bundles unused dependencies (1.47 MB → 0.97 MB)
  • •Improved startup time for query() sessions with in-process MCP servers (createSdkMcpServer) by running their handshake inside the SDK; initialize may now wait up to 250 ms for them
  • •Improved SDK session start-up: the CLI now answers the host's initialize request before starting its background start-up work
  • •Changed the Settings type's attribution field to boolean | {...}; TypeScript code that reads attribution.commit from returned settings needs a type narrow
  • •Updated to parity with Claude Code v2.1.281
v0.3.280

v0.3.280

September 22, 2026

  • •Added optional fireReason to the task-notification SDKMessageOrigin; a local host's declared scheduled-task fire is honored only in a process it started with CLAUDE_CODE_HOST_SCHEDULED_RUN=1
  • •Added verbatimPrompts option: prompts are delivered as written — no @path expansion, no slash-command dispatch and, on current CLIs, no ambient attachments with the prompt (Claude Code 2.1.248+)
  • •Added _meta to mcpServerStatus() tool entries, carrying a tool's MCP Apps ui metadata so a host can find its ui:// resource
  • •Added readMcpResource() (alpha) to read an MCP Apps ui:// resource from an MCP server that Claude Code connected
  • •Improved askSideQuestion(): asked while a turn is running, it now sees that turn (its prompt, replies and finished tool results so far) instead of only the last completed turn
  • •Improved unattended retry (CLAUDE_CODE_RETRY_WATCHDOG): a usage-limit wait emits rate_limit_event (rejected, resetsAt) as it begins; api_retry heartbeats continue while sub-agent work waits
  • •Changed session_state_changed events (CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=1) to report requires_action while an MCP elicitation waits on the user, as for permission prompts
  • •Changed headless sessions to cancel an MCP server's pending form question when the tool call that drew it ends
  • •Updated to parity with Claude Code v2.1.280
v0.3.278

v0.3.278

September 19, 2026

  • •Updated to parity with Claude Code v2.1.278
v0.3.277

v0.3.277

September 18, 2026

  • •Added an optional builtin field to SlashCommand, set when a command is built into Claude Code
  • •Added pasted_content to SDKUserMessage: text the user pasted rather than typed, appended after the typed prompt
  • •Added optional remote-session latency fields (first_text_post_ms, first_text_post_wall_ms, first_stream_post_queue_wait_ms, first_stream_post_queued_behind) to the success result message
  • •Added 'userSettings' as an updateSettings() source, accepting only effortLevel, which is saved for the session's current model as /effort saves it
  • •Fixed a resumed or forked session's total_cost_usd, modelUsage and get_usage totals starting at zero instead of continuing from the earlier turns (maxBudgetUsd is unchanged)
  • •Changed SDKUsageReport usage rows to always carry severity and is_active: the report relays only rows from a live server reply, and none while the usage fetch is failing
  • •Updated to parity with Claude Code v2.1.277
v0.3.276

v0.3.276

September 18, 2026

  • •Updated to parity with Claude Code v2.1.276
v0.3.275

v0.3.275

September 18, 2026

  • •Fixed a deferred tool call's result being emitted with internal keys such as toolUseResult instead of tool_use_result when the tool re-runs at the start of a resumed turn
  • •Fixed getSessionMessages() and forkSession() sometimes missing a turn's assistant message when called right after the turn's result message
  • •Fixed forkSession({ upToMessageId }) rejecting a client-supplied SDKUserMessage.uuid that is not in UUID format
  • •Fixed forkSession rejecting the id that getSessionMessages returns for a message sent while Claude was working, and a fork showing a re-run prompt twice
  • •Fixed getSessionMessages() omitting a task notification or other queued message that Claude read while running a tool; it now comes back where Claude read it
  • •Updated to parity with Claude Code v2.1.275
v0.3.274

v0.3.274

September 17, 2026

  • •Added startup_failure_reason to the error result a stream-json run writes before exiting on a known startup failure
  • •Added mcpServer: {name, source} to canUseTool options, mcp_server to tool hook inputs, and source to MCP server status rows, so hosts can key trust on source === "sdk"
  • •Added CLAUDE_CODE_MCP_STARTUP_WAIT_MS (pass via env) to bound or disable (0) the first-turn wait for connecting MCP servers
  • •Added CLAUDE_CODE_EMIT_STARTUP_TIMING=1 for stream-json hosts: the session's first system/init then carries the per-phase startup_timing breakdown that cloud sessions already emit
  • •Fixed getSessionMessages() omitting a message the user sent while Claude was running a tool; it now comes back as a user message where Claude read it
  • •Fixed missing origin: {kind: 'task-notification'} on the replayed user message when a background task finishes during a running turn (--replay-user-messages)
  • •Improved startup: the first turn no longer waits up to 2s for connecting MCP servers from settings files or plugins whose tools tool search defers; options.mcpServers servers are still awaited
  • •Changed queued background-task completions to share one model call: each still gets its own result, all but the last empty with num_turns: 0
  • •Updated to parity with Claude Code v2.1.274
v0.3.273

v0.3.273

September 15, 2026

  • •Added a usage_report sibling (SDKUsageReport: session totals, the plan's usage rows as the server sends them, extra usage) on the assistant message that delivers a headless /usage result
  • •Added reason: "worker_restart" to task_notification messages when a background task was stopped by a worker process restart
  • •Added a one-line transcript notice when the SDK host's Stop or SessionStart hook callback times out, shown once until the host answers again
  • •Fixed a Stop, SubagentStop or SessionStart hook callback that exceeds its timeout being reported as a hook failure and discarding other hooks' decisions; it now counts as no decision
  • •Fixed the browser SSE transport dropping the live slash-command list update (system/commands_changed)
  • •Updated to parity with Claude Code v2.1.273
v0.3.272

v0.3.272

September 15, 2026

  • •Updated to parity with Claude Code v2.1.272
v0.3.271

v0.3.271

September 15, 2026

  • •Added optional omitClaudeMd to AgentDefinition in the agents option, so a subagent can run without user, project and local CLAUDE.md files; managed policy files still load
  • •Fixed listSessions, getSessionMessages and getSessionInfo with dir on Windows not finding sessions for a directory on a mapped network drive or SUBST drive
  • •Fixed sessionStore resume losing the global config when it is stored under the legacy .config.json name or an OAuth-suffixed file name
  • •Removed persistent from the MonitorInput tool type
  • •Updated to parity with Claude Code v2.1.271
v0.3.270

v0.3.270

September 12, 2026

  • •Updated to parity with Claude Code v2.1.270
v0.3.269

v0.3.269

September 11, 2026

  • •Changed user_message_uuid, user_message_uuids and resume_reason to be stamped on a turn's first complete assistant message as well as its first stream event when partial messages are on
  • •Fixed result.permission_denials omitting Read, Edit and Write calls blocked by a path-scoped deny rule
  • •Fixed interrupts and permission responses being delayed while a host-started MCP server OAuth sign-in waited on a slow authorization server
  • •Fixed missing tool_use_id on task_started / task_notification when the CLI resumes a background subagent on its own; they now carry the agent's last call id
  • •Changed plan mode to route writes through canUseTool even when allowDangerouslySkipPermissions is set; the flag now only enables switching to bypassPermissions later
  • •Updated to parity with Claude Code v2.1.269
v0.3.268

v0.3.268

September 10, 2026

  • •Added result_index to result messages: the result's position in delivery order within the run, from 0
  • •Added local_command to the result message of a turn that ran a slash command without entering the model loop, carrying the command's name
  • •Added hold_on_cache_impact to the reload_plugins control request (Query.reloadPlugins({ holdOnCacheImpact: true })): holds a reload that would invalidate the session's prompt cache
  • •Added resume_reason to assistant, stream-event and result messages, set only on the automatic re-run of a turn a host restart interrupted
  • •Added kind (used, free, buffer, deferred) to each category in the get_context_usage control response, matching the /context result's context_usage rows
  • •Added optional defaultToNo and suppressAlwaysAllowRule hints to canUseTool options: the prompt should open on its decline option, or offer no persistent "always allow" choice
  • •Changed setModel() to confirm a model id the CLI doesn't know locally with the API the first time a session uses it, instead of refusing it as unrecognized
  • •Changed user_message_uuid on the automatic re-run of an interrupted turn to name that turn's last user prompt
  • •Changed the initialize success response to always include pending_permission_requests (empty when nothing is pending), so clients can tell that apart from an older CLI
  • •Changed the task-tracking tools (TaskCreate/Get/Update/List, TodoWrite) to be default tools only on Claude 3.x, Opus 4.0–4.7, Sonnet 4.0–4.6 and Haiku 4.5; elsewhere list them in tools/allowedTools
  • •Updated to parity with Claude Code v2.1.268
v0.3.267

v0.3.267

September 9, 2026

  • •Added getCcrEvent(query, message) and getSseLastSequenceNum(query) to the browser SDK's SSE transport, plus fromSequenceNum, onCatchUpTruncated and onDeliveryUpdate SSE options
  • •Changed systemPrompt recording to default on for custom prompts and appends (a mid-session prompt change takes effect at the next compaction); pass snapshot: false to keep per-request rendering
  • •Updated to parity with Claude Code v2.1.267
v0.3.266

v0.3.266

September 9, 2026

  • •Updated to parity with Claude Code v2.1.266
Showing 30 of 76