Anthropic
Claude Code CLI
2 days since its last release · 2.1.322 ·
Claude Code is an AI coding tool from Anthropic. Its latest release is 2.1.322, released on 10 Nov 2026. It shipped 22 releases in the last 30 days, and we have tracked 560 since 24 Feb 2025.
- Latest
- 2.1.322
- 10 Nov 2026
- Last 30 days
- 22
- releases
- Usual gap
- ~1d
- between releases
- Tracked
- 560
- since 24 Feb 2025
- Homepage
- code.claude.com/docs/en/overview
Releases per month
last 12 months- 21 Dec
- 29 Jan
- 29 Feb
- 23 Mar
- 29 Apr
- 27 May
- 30 Jun
- 23 Jul
- 29 Aug
- 27 Sep
- 27 Oct
- 5 Nov
Release history
560 releasesReleases 76
-
Fast mode now available on Vertex in global, US and EU regions
- Fast mode now works on Vertex in the global, us and eu regions
- Falls back to standard speed automatically if Vertex refuses fast mode
-
Claude Fable 5.2 arrives as the new default Fable model
- Claude Fable 5.2 is now the default Fable model, with 1M context at $8/$40 per Mtok and $0.20/Mtok cache reads
- Saved Fable 5.1 picks move to fable; use /model claude-fable-5-1 to switch back
- claude remote-control gains --settings to apply a settings file to every session it starts
- Security fixes for MCP OAuth secrets sent to redirected hosts and for Bash commands that skipped sandbox or permission checks
- Many fixes to Remote Control, cloud sessions, hooks and file edits on network paths, plus faster startup and install
-
MCP tasks now run in the background, and auto mode changes behind gateways
- MCP server tasks now move to the background right away and report their result to Claude when they finish
- Auto mode with telemetry or feature flags off, including behind LLM gateways, no longer falls back to the built-in classifier
- New permissions.denyInsteadOfPrompting setting denies tool calls that would wait on a permission prompt and lets the turn continue
- Added /claude-api latency-optimize, spend totals in /usage behind Claude apps gateways, and a /tasks row for finished subagents
- Many fixes, including a Bash permission bypass, a sandbox escape when creating worktrees, and MCP connection issues
-
Sign-in expiry warnings, cloud sessions without a repo, many fixes
- Cloud sessions on Anthropic-hosted environments can now start without picking a repository
- claude auth status and claude -p JSON results now show when a stored claude.ai sign-in expires, with a stderr warning 10 days ahead
- Mods can run a model fork at a set effort level, such as low for quick side jobs
- Claude Code no longer installs or updates the IDE extension when started in an IDE terminal
- Many fixes to config and login saves on shared network home directories, plus fixes for --continue, cloud sessions and permission prompts
-
Gateway cost reports, an email privacy check and dozens of fixes
- Claude apps gateway admin API adds a cost report of past spend, in total or per developer, day or IdP group
- Claude now asks before a web or shell tool call sends your account email address to the network
- Agent SDK hook callbacks can set onFailure "block" so a failing callback blocks the action instead of letting it run
- /schedule now warns when a new cloud scheduled task gets connectors you did not ask for
- Many fixes across auto mode, MCP servers, compaction, self-hosted runners, OpenTelemetry exports and the VS Code extension
-
Opus 4.8 fast mode deprecated, gateway gets pooled group budgets
- Fast mode for Opus 4.8 is deprecated and will be removed on November 30, 2026, after which requests error out; migrate to Opus 5.5 to keep the speedup
- Claude apps gateway adds pooled monthly group budgets, layered policy mode and Bedrock guardrail input tagging
- New CLAUDECODESUBAGENTEFFORT sets subagent effort level, and CLAUDECODEDISABLEULTRATHINK turns off the ultrathink keyword
- Cloud sessions add a form card for MCP server requests and Enterprise members can now manage their own network secrets
- VS Code adds gateway sign-in from the chat panel, plus many fixes across the Bash tool, auto mode, gateway model checks and Slack
-
New setting blocks file writes outside your working directories
- New permissions.blockWritesOutsideWorkingDirectories setting stops Write, Edit and NotebookEdit touching paths outside working directories in any mode
- VS Code and Claude Desktop's Code tab now summarize idle conversations over 200K tokens and use the summary on your next message if the prompt cache has expired
- Claude apps gateway upstreams can now be limited to listed IdP groups, and Bedrock upstreams can log which guardrail filter fired
- Sessions no longer fail turn after turn when an admin caps a model's effort: Claude Code lowers the effort, retries and says which level it uses
- FORCEHYPERLINK=false now turns hyperlinks off, and claude -p now prints the error and exits 1 when startup fails
-
Unrecognized cloud model names now get the shorter system prompt
- Bedrock, Vertex, Foundry and Mantle now use the shorter system prompt for unrecognized model names; set CLAUDECODESIMPLESYSTEMPROMPT=0 to keep the full one
- The Claude apps gateway can identify people by an idtoken claim other than sub, such as Microsoft Entra ID's oid, via oidc.useridclaim
- Prompt and agent hooks now run at most 24 at a time per session, and a timeout under 1 second counts as 1 second
- In fullscreen, clicking the "to send now" hint under a queued message sends it right away, and cloud routines gain an effort level setting
- Claude no longer overwrites a very large CLAUDE.md or rules file after reading only its start, among many other fixes
-
Ultrathink now runs at xhigh effort, plus a long list of fixes
- Typing ultrathink in the terminal now runs the turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API
- Standard-context sessions that go over 200K tokens now compact and continue instead of stopping with a 1M-context credits error
- Bedrock, Mantle, Vertex and Foundry now load model settings from a signed file on downloads.claude.ai; CLAUDECODEMODELCATALOG=off turns this off
- VS Code extension adds syntax coloring to code blocks and sends no usage metrics or error reports while VS Code telemetry is off
- Fixes for startup freezes, Write and Edit stalling on large files, subagents missing MCP tools, and SubagentStop hooks firing repeatedly
-
Monitor tool comes to Bedrock, Vertex and Foundry, plus gateway fixes
- Monitor tool and plugin monitors now work on Bedrock, Vertex and Foundry, and when telemetry or nonessential traffic is off
- Requests now retry without the effort beta value or mid-conversation system messages when a gateway or compatible server rejects them
- Enterprise custom roles can now control access to cloud sessions and Remote Control, and turning either on in admin settings now asks for confirmation
- Fixed compaction after a Prompt is too long error keeping too much dense text, and background cleanup deleting recent files such as pasted images
- Windows hooks no longer let a tool call through when Git Bash is missing and the hook command is only quoted strings
-
Sandbox and permission fixes, and ultrareview now needs workspace trust
- Fixes many sandbox and permission gaps, including shell startup files running outside the sandbox and Bash checks missing some cd and redirect cases
- claude ultrareview now requires workspace trust; scripts must pass --trust-workspace or set CLAUDECODEDISABLEULTRAREVIEWTRUSTCHECK=1
- sed -i now gets the standard Bash prompt and runs as written instead of a simulated edit preview, and /rewind no longer restores it
- In bypass permissions mode in the Desktop app and VS Code, dangerous rm and rmdir commands are now denied at once with a rewrite hint
- Clearer disk-full errors, faster git commands on Windows and quicker Edit and Write in the VS Code chat panel
-
New API version migration command and safety-event telemetry
- Added /claude-api api-version to move Claude API code to API version 2026-10-21
- New claudecode.safetyevent OpenTelemetry event for each tool call stopped, prompted or sent back by permission rules, auto mode, hooks or the sandbox
- Code Review adds @claude review stop (also off and pause) to end an always-review subscription on a pull request
- Self-hosted runner checkout hooks now stop after 30 minutes by default; raise --checkout-hook-timeout-min or set it to 0
- Many MCP fixes, including HTTP tool calls no longer sent twice when a connection closes mid-call
-
Fixes every request failing on Vertex, Bedrock and gateways
- Fixed every request failing on Vertex, Bedrock and other providers when tool search is rejected; Claude Code now retries without it
- New sandbox.allowNestedUserNamespaces setting lets sandboxed Linux commands create user namespaces, though this bypasses the sandbox's file protections
- allowedMcpServers no longer filters managed-mcp.json stdio servers that use ${VAR} only in env; use deniedMcpServers to keep one off
- /usage warnings are now one-line fixes that name their source, and pressing f applies them
- Unattended retry mode can now notice a usage limit that was lifted before its reset
-
Guided auto mode setup, Code Review levels and a resume usage warning
- New /auto-mode-setup reads your project and recent sessions to teach auto mode about your environment, now also on Bedrock, Vertex and Foundry
- Code Review adds Quick, Standard and Deep review levels, plus an option to review every repository in a GitHub organization
- Pro and Max plans now warn when resuming an inactive conversation would use a noticeable share of the 5-hour usage limit
- Mods gain decider: "person" and e.isReadOnly in tool.check, and settings.read hooks can make an organization setting a ceiling
- Fixed claude --worktree removing a worktree and its branch without asking at exit, which could lose commits
-
Fork subagents on by default everywhere, plus a long list of fixes
- Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app and VS Code; CLAUDECODEFORKSUBAGENT=0 turns them off
- Desktop app and VS Code now always run the main conversation's subagents in the background, as the terminal does
- New /config row to turn idle compaction off or on, and CLAUDECODEAUTOMODEMODEL to pick the client-side auto mode classifier model
- Standalone Claude Design (claude.ai/design) closes by December 14, 2026, with a notice added to /design-sync and /design-login
- Faster startup with many plugins enabled, and dozens of fixes across MCP, plugins, cloud sessions, VS Code and Claude Tag in Slack
-
LaTeX math in VS Code, vim bracket matching and many fixes
- VS Code now renders LaTeX math in Claude's responses, both inline and as blocks
- Vim mode adds the % motion to jump to the matching bracket, alone or with d/c/y
- Switching to a model with a smaller context window now has the previous model summarize a long conversation first
- OpenTelemetry tooldecision events now include classifier verdict, category and permission mode
- Many fixes, including the fable alias falling back to the previous Fable model and Read no longer treating binary files as text
-
New mod hooks and APIs, /advisor on AWS, and a long list of fixes
- Mods gain virtualText prompt decorations, $.ui.view to show an agent's transcript, a prompt.change hook, right-to-left Text and agent.spawn rewrites
- /advisor arrives on Claude Platform on AWS, off until you pick an advisor model
- Claude in Slack supports Slack Connect channels, private approval cards for connector actions and voice clip transcripts
- Hook updatedInput is now validated like Claude's own input, and remote-control's no-op --sandbox flags now exit with an error
- Many fixes for worktrees, reftable repos, session resume after compaction, Remote Control, cloud sessions and the VS Code extension
-
Idle compaction keeps long conversations from reprocessing after cache expiry
- Conversations of 200K+ tokens left idle for about 54 minutes are now compacted before the one-hour prompt cache expires
- Idle compaction's token minimum can be changed with CLAUDECODEIDLECOMPACTMINTOKENS, or turned off in settings.json
- New CLAUDECODEDISABLEFILEMENTIONS sends Claude the paths of @ file mentions instead of the files' contents
- MCP allow rules and agent tools lists now reach only the server they name exactly, and some MCP proxies no longer follow redirects
- Many fixes for sandbox escapes, mods, Remote Control, cloud sessions, MCP connections, Windows and text rendering
-
Gateway policies for Desktop's Code tab, plus subagent and Read tool options
- Claude apps gateway policies gain a code key that applies CLI settings to Claude Desktop's Code tab and turns on its gateway mode
- Subagents can set autoCompactWindow to auto-compact earlier than the main conversation
- Read tool gets an allowlarge option to read a large text file in one call when context has room
- Sonnet 5.5 cache reads now priced at $0.10 per million tokens in /cost, the status line and budgets
- Default limit for MCP tool descriptions and server instructions raised from 2,048 to 4,096 characters
-
Hooks can block actions on failure, and terminals can show live status
- Command and HTTP hooks can set onFailure: "block", so a hook that fails to start, times out or exits oddly blocks the action instead of letting it through
- Terminals that support Program Status Protocol (OSC 7501) can show whether Claude Code is working, waiting on you, or done
- Claude apps gateway upstreams accept an optional models list, a stream start timeout on cloud upstreams, and upstream request IDs in audit logs
- VS Code adds a chat row for files Claude sends you, which opens the file in the editor when clicked
- Many fixes for MCP reconnects, plugins and mods, terminal freezes on long output, and background sessions
-
Claude Haiku 5.5 becomes the default Haiku model, plus many fixes
- Added Claude Haiku 5.5, now the default Haiku model on the Anthropic API, with 1M context at $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)
- Fixed Claude sometimes retracting or redoing finished work after a context compaction
- Fixed a memory leak in HTTP MCP connections and messages lost when moving a session to the background
- Status line payloads for subagents now include agentType, and mods can register tools that skip tool search
- Reverted two earlier changes: the auto mode denial message wording and the cloud session wake-up fix from 2.1.290
-
Install plugins straight from a marketplace, set sub-agent effort
- claude plugin install now takes --marketplace to add a marketplace if needed and install a plugin from it in one step
- The Agent tool gains an effort parameter so Claude can run a sub-agent at the effort level you ask for
- Local MCP servers now negotiate protocol version 2026-07-28 by default; MCPPROTOCOLNEGOTIATION=legacy opts out
- Mods get prompt caching in $.model.complete and a new prompt.autocomplete event for custom autocomplete rows
- Many security, sandbox, permission, plugin and cloud session fixes, plus faster rendering of long lists
-
Refilling WebSearch budget, Managed Agents setup and a big batch of fixes
- WebSearch budget in interactive sessions now refills at 100 calls per hour instead of stopping after 200 calls, and the rate can be configured
- New /claude-api managed-agents-onboard sets up Managed Agents from a page URL or a Console quickstart template such as deep-researcher
- Claude Tag in Slack adds fast mode: mention Claude with !fast to switch a thread to fast mode, and !fast off to switch back
- claude attach and claude logs now accept part of a session name in place of the id
- Many fixes to permission checks, sandboxing, scheduled tasks, the agents view, /ultrareview uploads and plugin hooks
-
Permission rule fixes and sturdier plugin and mod rendering
- Fixed Bash deny and ask rules being skipped under sandbox auto-allow when a command had an environment variable prefix or a variable assignment before it
- Fixed Read deny rules not applying to files reached through a symlink in the IDE, and nested compound-command rules being overridden by a mod's approval
- Plugins can now use agent.spawn for teammates, get one agent id across hook events, and see idle and waiting states in $.agent.list()
- A plugin or mod that fails while drawing now fails on its own instead of ending the session or taking down everything around it
- Fixed the terminal and published artifact pages freezing on short code blocks with many unclosed script tags
-
Recover cleared prompts, set code-review finding limits, and many fixes
- Pressing Up on an empty prompt brings back a draft cleared with Ctrl+C, including pasted text and images
- /code-review accepts --max-findings to report more or fewer findings, and remembers the choice
- Cloud sessions get a built-in gh api when the image has no GitHub CLI, and MCP servers can ask you to re-authenticate for more OAuth scope
- Mid-response API timeouts no longer fail the turn, and long conversations auto-compact instead of failing with "Prompt is too long"
- Background command time limit now applies only to unattended sessions such as -p, SDK, CI and cloud
-
Claude Mods let plugins change deeper behavior, plus a watchdog side agent
- Added Claude Mods, letting plugins modify Claude Code's deeper behavior
- Added You should know, a built-in mod where a side agent flags things you or Claude might miss
- Opus 4.7+ and Fable now default to a 1M context window on Bedrock, Vertex, Foundry and the Claude apps gateway
- MCP servers on the 2025-11-25 protocol can now show URL prompts, for example to sign in
- VS Code can move a running command or sub-agent to the background, alongside many screen reader and remote session fixes
-
VS Code gets bookmarks, plus permission prompt counts and many fixes
- VS Code extension adds a Bookmarks side panel for saving Claude's responses, and option previews on question cards
- Stacked permission prompts now show a count such as 2 of 5, and more prompts now match the look of file edit prompts
- --bare now connects only the MCP servers named on the command line, sends no system reminders and starts no background tasks
- When the API refuses the model your default or alias points to, Claude Code now retries once on the previous model of the same tier
- Many fixes for resumed sessions losing turns, secret redaction in logs, Remote Control, MCP connectors and subagents
-
Open the desktop app from the CLI, plus plugin configuration commands
- New claude --desktop opens the Claude desktop app on the current directory or on a resumed session
- New claude plugin configure command, and install-time settings for bundled MCP servers, so plugins can be set up without the /plugin menu
- New allowedProviders managed setting lets admins limit which API providers a machine may use
- Background Bash and PowerShell commands now stop after a time limit (default 30 minutes, max 2 hours)
- Sessions using a custom ANTHROPICBASEURL now get the 1M context window on models that support it
-
Sonnet 5.5 becomes the default Sonnet, and sessions start in auto mode
- Added Claude Sonnet 5.5 as the default Sonnet model on the Anthropic API, with 1M context at $2/$10 per Mtok and $0.20/Mtok cache reads
- Terminal and VS Code sessions now start in auto mode when no permission mode is configured; permissions.defaultMode still overrides it
- Ultracode is now its own toggle in /effort, no longer forces xhigh effort and stays on at any effort level
- Added /mcp reconnect all to retry every failed MCP server at once, and dollar amounts for gateway spend limits in /usage and the status line
- Fixed damaged response streams, "Prompt is too long" errors that persisted after compacting, and many rendering, vim mode and VS Code issues
-
New admin controls to allow or block models, plus a prompt audit
- New deniedModels managed setting blocks specific models, and availableModelsMatch "exact" keeps new model versions blocked until listed
- New /doctor prompt-audit checks CLAUDE.md files, skills, agents and commands for prompting patterns written for older models
- Sessions on third-party providers or with telemetry off now start in auto mode when no permission mode is set
- Lists across /mcp, /tasks, /help, /plugin and other pickers now support page keys, the mouse wheel and clicks
- Many fixes for MCP servers, plugin commands, vim mode, VS Code sessions and cloud sessions, plus faster startup and first replies
-
New prose-width setting, telemetry settings lockdown and many session fixes
- Added a maxProseWidth setting that caps prose width in wide terminals while tables and code blocks keep full width
- Project and local settings now ignore OpenTelemetry variables that turn on export, set the endpoint or capture content, with a startup notice
- Fixed resumed and continued sessions losing earlier extended thinking and requests failing with 400 or redactedthinking errors
- Cloud sessions can attach a repository from a different GitHub owner and show Claude GitHub App status in Settings
- Auto mode now uses the server-side classifier by default on direct Anthropic API connections when telemetry is off
-
Bedrock role and guardrail options for the apps gateway, plus many fixes
- Claude apps gateway Bedrock upstreams can assume an IAM role through STS and apply an Amazon Bedrock guardrail to every request
- Send now (ctrl+enter) moves running tools to the background instead of cancelling the turn
- New "attribution": false setting hides all commit and PR attribution
- MCP servers can open browser-based flows through URL-mode elicitation, and claude plugin validate now checks MCP server entries
- Many fixes for session resume, prompt caching, proxy stream handling, vim mode and dialogs, plus a prompt for rm -rf on command-substitution targets
-
Claude Opus 5.5 arrives as the default Opus model, with 1M context
- Added Claude Opus 5.5 as the default Opus model, with 1M context at $4/$20 per Mtok and $0.20/Mtok cache reads
- Pro and Team Standard plans now default to Opus instead of Sonnet, matching Max, Team Premium and Enterprise
- VS Code gains Status, Sandbox, Claude in Chrome and Export dialogs, plus typed /plan and /skills commands
- Writes through symlinks are now judged by where they actually land, so acceptEdits and auto mode no longer approve writes outside the project
- A stray y or n no longer confirms or closes dialogs; Enter and Esc do, and you can rebind y/n in keybindings.json
-
Auto mode now defaults to the server-side classifier at no extra cost
- Auto mode for API, Enterprise, Bedrock, Vertex, Foundry and gateway users now defaults to the server-side classifier, with no charge for its overhead
- Setting CLAUDECODEAUTOMODESERVER=0 opts out on Bedrock, Vertex, Foundry and gateways
- A warning now appears when auto mode falls back to a billed classifier
- /status has a new Auto mode server row showing whether the session's classifier runs on the server
-
Reads AGENTS.md when a project has no CLAUDE.md
- Projects without a CLAUDE.md now use AGENTS.md for instructions, configurable under Project instructions in /config (not yet on Bedrock, Vertex or Foundry)
- The deprecated TaskOutput tool is removed, so Claude reads background task output with Read and taskOutputMaxChars no longer has any effect
- Invisible Unicode formatting characters are stripped from prompts, and the cleaned prompt is shown for review before sending
- VS Code adds a Copy response button, a Sign out option, running tasks in the agent map and cost and token usage for API-key and cloud-provider users
- Many fixes for crashes, hangs in claude -p and SDK sessions, unexpected logouts, plugin installs and prompt-cache misses
-
Syncs claude.ai skills and plugins, adds a send-now key
- Skills and plugins enabled on your claude.ai account now sync to terminal sessions signed in with it, with settings to opt out
- New send-now key (ctrl+enter) interrupts the current turn and sends all queued messages at once
- VS Code proposed-change diffs get accept and reject buttons under each change, and saved memories can be edited in the Memory dialog
- Plugins from npm sources are now fetched without running install scripts and are integrity-verified
- Many fixes for resume crashes, fullscreen freezes, sandboxed Bash exit codes and tokens leaking in plugin URLs
-
Critical memory warnings, MCP reliability fixes and many bug fixes
- Shows a visible warning when memory usage is critical, with steps to free memory or restart safely
- Fixes sessions endlessly retrying "unexpected tooluseid" errors; corrupted transcripts now self-heal or end with a /rewind hint
- Fixes several MCP connection, timeout and refresh problems, and adds CLAUDECODEMCPSTARTUPWAITMS to bound startup waits
- VS Code adds continue-after-reload, Memory and Instructions entries in the Customize menu, and better screen reader support
- Cloud session diff view gains a Compare against branch picker for diffing changes against any branch
-
Fork Remote Control sessions from the app, plus permission fixes
- Sessions started with Remote Control can now be forked from the Claude app and run as a background session on your computer
- Fixed permission gaps where unanalyzable Bash commands skipped prompts and a subshell could hide a dangerous rm in bypass mode
- New opt-in request headers for LLM gateways via CLAUDECODEGATEWAYHINTHEADERS=1, and a notice when an MCP server disconnects for good
- Auto mode on Bedrock, Vertex and Foundry now uses the local classifier by default
- Auto-compact no longer fires at about half the real context window in sessions using the advisor tool
-
Fast mode in Remote sessions and per-command network access in auto mode
- Fast mode now works in Claude Code Remote sessions on cloud and self-hosted runners, where your organization allows it
- Auto mode with sandboxing lets Bash, PowerShell and Monitor commands open only the hosts each one needs, and refuses all others
- Subagents can skip user, project and local CLAUDE.md files with the new omitClaudeMd option
- Monitor watches now always have a deadline of at most 30 minutes, replacing the no-timeout persistent option
- Dozens of fixes to Bash permission checks, MCP OAuth, session resume, fast mode and the VS Code extension
-
Plugin eval suites, output-style switching and VS Code agent map
- New claude plugin eval command runs a plugin's eval suite and produces scored, reproducible JSON and HTML reports
- New /output-style command lists and switches output styles, including over Remote Control and in cloud and headless sessions
- VS Code adds an agent map of sub-agents plus dialogs for managing hooks and permission rules
- Fixed Edit deny rules and write-path checks not applying to files written by Bash tee commands
- Many fixes for prompt-cache reuse, terminal key handling and /goal runs stalling after errors
-
Fixes third-party endpoint errors and adds gateway pricing sync
- Fixed every turn failing with HTTP 400 on third-party Anthropic-compatible endpoints since 2.1.265
- Claude apps gateway can push pricing to signed-in clients so /cost and telemetry match the spend meter, plus new network safety warnings
- Task-tracking tools are now offered only on older models unless CLAUDECODEENABLETODOTOOLS=1 is set
- Permission rules now apply on symlinked paths, and secrets in plugin and MCP errors are no longer shown
- WebFetch now times out after 300 seconds, and --continue/--resume show the conversation immediately
-
New maxEffortLevel setting caps effort, plus many prompt-cache fixes
- New maxEffortLevel setting caps the effort level on every provider, including Bedrock, Vertex and Foundry, set globally or per model
- New --system-prompt-snapshot off option renders the system prompt fresh on every request, for iterating on prompt text
- Many fixes for prompt-cache misses and dropped extended thinking when resuming sessions, switching models or when MCP tools change
- claude remote-control no longer exits and drops attached sessions when its server credential expires after about 30 days
- VS Code extension fixes, including a 100% CPU hang, broken WSL2 screenshot pastes and diff blocks that ignored the color theme
-
Plugin folders load from one directory, plus many resume and cache fixes
- --plugin-dir can now point at a folder of plugins, loading each child with a manifest and picking up ones added or removed while running
- VS Code adds automatic archiving of inactive sessions, set to 14 days by default
- Tool results saved to disk are now capped at 1 GB, and the preview says when a file was truncated
- Fixed resumed subagents and agent teammates breaking prompt-cache reuse, and made resuming long sessions faster
- MCP servers set as http that only speak legacy HTTP+SSE now fall back to SSE and connect
-
New /skill-doctor command, bigger inline output limits and many fixes
- Added /skill-doctor to show which loaded skills go unused and how much context they cost, so you can prune them
- New bashOutputMaxChars and taskOutputMaxChars settings let Claude receive up to 128K characters of command output inline
- /status and claude doctor now explain why your organization's policy could not be loaded
- Prompt word-editing keys now match Bash, and the keybindingFlavor setting no longer has any effect
- VS Code can now add and remove MCP servers from a dialog, and many Remote Control and VS Code bugs are fixed
-
Live diff panel in fullscreen mode, plus prompt-cache miss explanations
- New diff panel beside the conversation in fullscreen mode shows uncommitted changes as Claude edits, toggled with /diff
- /cost and the status line now show a likely cause when the prompt cache misses
- Fixed permission rules with parentheses in paths being ignored, which left read-only folders writable
- Fable 5.1 now appears in the /model picker, and its prompt caching now covers context added after tool results
- Background commands started by subagents no longer stop after one hour, and auto-compact on 1M-context models now waits until close to the limit
-
Admins can now push MCP servers to every user in their organization
- New managedMcpServers setting lets organizations provide HTTP/SSE MCP servers to all users
- New --permission-prompts none flag auto-denies prompts on unattended headless hosts
- allowedMcpServers now governs only user-added servers; use deniedMcpServers to block managed ones
- GitLab merge requests from glab commands now show as MR !N, and plugin validate gains --json output
- Fixes concurrent sessions overwriting each other's settings and gaps in Bash Read() deny rules
-
Claude Fable 5.1 becomes the default Fable model, plus auto mode safeguards
- Added Claude Fable 5.1 as the default Fable model, with 1M context at $10/$50 per Mtok and $0.25/Mtok cache reads
- Auto mode no longer auto-approves cloud credential fetches, egress evasion or cross-tenant reach, and asks before reading files outside working directories
- New timeFormat and timeZone settings for turn-end clock and transcript timestamps, and a session-only effort option in /effort
- bypassPermissions in project .claude settings is now ignored, and several permission bypasses in Bash, plugins and auto mode are fixed
- VS Code adds a model pill with effort picker, output style selection and archiving in place of session deletion
-
Model-switch hooks, prompt-cache stats in /cost and many security fixes
- New PreModelSwitch and PostModelSwitch hook events can block, confirm or annotate a model switch
- /cost now shows per-session prompt-cache stats, and /usage adds a spend limit bar for Claude apps gateway users
- Remote Control clients can now watch a foreground subagent's tool calls and results live
- Security fixes for symlink handling in file tools, plugin path traversal and Bash arithmetic auto-approval
- /effort now saves a default per model, and CLAUDECODESUBAGENTMODEL no longer overrides models set per agent
-
New restricted mode locks down tools, settings and permissions
- New --restricted flag removes command-running tools and WebFetch, keeps file tools in the working directory and ignores local settings files
- Cross-session messaging now works on Bedrock, Vertex and Foundry, and when telemetry is disabled
- New /usage-credits lets some Enterprise members ask their admin for a higher usage limit
- Workflow tool description shrinks from about 5.7k to 1k tokens, with its reference moved into a bundled skill
- Fixes hourly prompt-cache misses in long sessions and Desktop sessions disappearing after 30 days
-
Claude can now draft feedback reports, plus a cost-optimizing API skill
- New SendFeedback tool lets Claude draft a feedback report for you to review and send from /feedback, and the feedbackDrafts setting turns it off
- New /claude-api cost-optimize profiles a project's Claude API spend and works through cost levers one measured change at a time
- Bash permission prompts now offer a one-keystroke option to approve and switch to auto mode
- Sonnet 5 now auto-compacts at about 967K tokens on the full 1M context window, up from about 934K
- Fixes for sub-agents failing on model 404s, sessions wedged by huge hook output, and fast arrow-key plus Enter selecting the wrong row
-
Auto mode rules in /permissions, smarter /cd and many fixes
- New Auto mode tab in /permissions for viewing and editing auto mode classifier rules
- /cd now applies the new directory's settings, hooks, MCP servers, skills and agents right away
- Non-interactive sessions automatically continue responses cut off by server errors, connection loss or stalls
- Subagents that hit their maxTurns limit now return partial output with a hint to continue via SendMessage
- Dozens of fixes for fullscreen scrolling, background sessions, plugins, MCP tools and permissions
-
Loop usage breakdown, custom model picker and keyless Console login
- /usage now breaks down each /loop task by run count, total tokens, tokens per run and last run, so runaway loops are easy to spot
- New modelPicker setting builds an ordered, labeled /model list that can add to or replace the built-in models, including Vertex and Bedrock ids
- /login can now sign in with an Anthropic Console account without an API key, for organizations that don't allow API keys
- New settings set prompt cache lifetime separately for the main conversation and for subagents, and admins can set contracted model pricing for cost figures
- Native binary download drops from about 340 MB to 75 MB on Linux x64, with 40–70 MB less memory per session and faster startup
-
Cost estimates include data-residency premium, plus many fixes
- Cost estimates in /cost, the status line and --max-budget-usd now include the 1.1x US-only-inference premium for data-residency workspaces
- New /claude-api upgrade command migrates Python projects from anthropic 0.x to 1.x
- Cross-session messaging with SendMessage and ListAgents now works on Windows
- Fullscreen renderer now offered on Bedrock, Vertex and Foundry, and new installs there start in fullscreen
- Fixed Bedrock streaming behind proxies silently doubling billed API calls, plus dozens of resume, fullscreen and input fixes
-
Readline-style Ctrl+W, plugin auth headers and many Remote Control fixes
- New keybindingFlavor setting: "readline" makes Ctrl+W delete back to the previous whitespace, as in Bash
- Plugin marketplaces can use headersHelper to mint HTTP headers such as short-lived tokens, and plugin install/update now asks for confirmation
- Self-hosted runners gain graceful shutdown on SIGTERM and support for egress proxies that need a fresh Proxy-Authorization header
- Fixed unbounded memory growth in long sessions and made many Remote Control fixes for reconnects, model picks and lost messages
- Ctrl+L and Cmd+K in fullscreen now only repaint, and the double-press /clear shortcut is removed
-
New built-in Concise output style and prompt caching fix for gateways
- Added a built-in Concise output style that leads with results and skips preamble and narration, selectable under Output style in /config
- Fixed prompt caching for sessions using an LLM gateway or custom base URL
-
New default-model env var and idle notices between sessions
- New ANTHROPICDEFAULTMODEL environment variable sets the starting model, while a /model pick still overrides it and persists across restarts
- SendMessage can now ask another local session to send a one-time notice when it next goes idle (macOS and Linux)
- macOS sandbox wildcard read-deny rules such as /.env now take precedence inside allowed regions and can't be bypassed by renaming files
- VS Code adds screen reader support for the transcript, with live announcements and per-turn heading navigation
- Auto mode now reviews Monitor commands like Bash commands, and many fullscreen, /model picker and SendMessage bugs are fixed
-
Optional spellcheck in the prompt input, plus many fixes
- New optional spellcheck setting underlines misspelled words in the prompt as you type, using installed aspell, hunspell or ispell
- Fixed Shift+Tab in the permission prompt's comment field granting session-wide edit permission instead of closing the field
- Permission dialogs now match what a grant would cover and withhold "don't ask again" when contents can't be fully shown
- Lower memory and CPU use while cloud sessions like /ultrareview or /autofix-pr run in the background
- Fixed whole prompt cache being invalidated when a language server disconnected or reconnected mid-session
-
Sessions resume automatically after usage limits reset, plus GitLab MR badge
- Sessions now continue automatically when a claude.ai usage limit resets, which you can turn off in /config
- Repos with a GitLab remote and an authenticated glab CLI show a merge request badge in the footer and statusline
- /permissions, /add-dir and several other dialogs can now be opened while Claude is working
- Loading the built-in claude-api skill now costs about 25k tokens of context, down from over 200k
- Security hardening against Windows NT-namespace paths, plus many fixes for Remote Control, permission prompts and rendering
-
GitLab merge request support, Bash memory limits and todo tools off on newer models
- GitLab merge request URLs now work with the --worktree flag and the claude agents view, where MRs show as !N
- Todo and task-tracking tools are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer models; CLAUDECODEENABLETODOTOOLS=1 restores them
- Opt-in memory cgroup limits for Bash tool commands on Linux stop a runaway build from stalling the session
- Opt-in apps gateway setting forwards the signed-in user's identity as headers so a proxy can attribute spend per user
- Fixes a Windows path validation bypass that could leak NTLM credentials, plus fixes for idle CPU use, Notification hooks and cloud sessions
-
Subagent forking on by default, plus @-mentions to message other sessions
- Forked subagents now on by default, inheriting the full conversation and prompt cache; agent spawns in interactive sessions run in the background
- Type @ to mention another Claude session by name and message it directly via SendMessage
- GitLab support for plugin marketplaces and secret redaction for GitLab token families
- Fixed several permission bypasses in PowerShell, Git Bash symlinks and nested git repository trust
- Many Remote Control reliability fixes, including longer reconnects and clearer session takeover messages
-
Plugin marketplace command sources, VS Code session groups and many fixes
- Plugin marketplaces can use a local command source, re-resolved each session and applied without a restart
- VS Code sidebar adds session groups, and the /btw side-question panel can now be resized
- Gateway streaming sends keepalive pings during long thinking pauses to prevent Vertex and Bedrock idle disconnects
- /commit-push-pr no longer auto-approves git/gh commands with dangerous flags like --force, --amend or --no-verify
- Fixes long streamed responses partly vanishing or printing twice, plus several crashes on narrow terminals and Windows paths
-
Write tool can overwrite unread files, plus many bug fixes
- Newer models can now use the Write tool to overwrite an existing file they haven't read this session, matching the Edit tool's rules
- Skills synced from claude.ai no longer shadow local commands or MCP prompts, and on your machine they don't run ! commands or expand @ files
- Fixed interactive sessions that could stop redrawing after a rare layout error, and session cleanup deleting a project's memory folder contents
- Expired or missing Vertex AI credentials now fail within seconds instead of retrying for minutes
- Cross-session messages now show the sender and body inline, and compaction shows a retry countdown and stall hint
-
SendMessage can start chats with Remote Control sessions on other machines
- SendMessage can now start a conversation with your Remote Control sessions on other machines by name, not only reply to them
- Usage warning supports gateway spend limits, naming the cap, its reset time and the operator's message
- claude agents now shows a workspace trust prompt for untrusted directories, like claude does
- Photos attached from the Claude app in Remote Control are now shown to Claude directly
- Fixes for headless OAuth token loss, MCP OAuth 401 errors on macOS and Remote Control resume after compaction
-
Run web, mobile and desktop sessions on your own machines
- New self-hosted runner command turns your own machines or containers into environments for web, mobile and desktop sessions on Team and Enterprise plans
- Sessions can now message each other across your machines, with agent discovery on macOS and Linux
- Plugins can be installed from a zip archive over HTTPS without git or npm, with optional SHA-256 pinning
- New sandbox credential-masking options, including JWT-aware masking and AWS SigV4 re-signing
- The 200-subagent-per-session limit is gone, and Remote Control gets clearer connection and compaction status
-
Security fixes for permission bypasses and /review merged into /code-review
- Fixed several permission bypasses, including crafted Bash commands and invisible Unicode hiding parts of a command from approval prompts
- Fixed workflow scripts escaping the sandbox via dynamic import(), and agent bypassPermissions ignoring the org disable policy
- /review is now an alias of /code-review, which reuses your last effort level when none is given
- Managed settings now accept owner wildcard entries (owner/) to allow or block all marketplace repos under a GitHub org
- CLAUDECODEDISABLE1MCONTEXT now holds every Claude model with a native 1M window to 200K
-
Worktree isolation fixes, safer auto mode and ultraplan removed
- Worktree-isolated sessions and their subagents can no longer run destructive git commands against the main checkout
- Ultraplan feature removed
- Repo-local settings can no longer turn on Remote Control auto-start; enable it at user scope via /config
- Auto mode now checks SendMessage messages to other agent sessions with the permission classifier before sending
- Fixes for HTTPS proxy startup hangs, /usage MCP attribution, false connection errors and hook bypasses in background tasks
-
VS Code Focus view and masked sandbox credentials on Linux
- VS Code gets a Focus view that hides tool activity behind an expandable per-turn summary, toggled with Ctrl+Alt+F
- Sandboxed commands on Linux and WSL can read masked copies of credential files, with the real value substituted only on egress
- Background sessions now commit and push to preserve work, open draft PRs only when needed and report where the work lives
- Plugins installed from /plugin activate immediately when safe, and /fork now creates its own worktree
- Fixes Bash and PowerShell permission-check bypasses, plus many MCP, Vim mode and Windows fixes
-
Opus 5 becomes the default Opus model, with 1M context
- Claude Opus 5 added as the default Opus model, with 1M context and fast mode at $10/$50 per Mtok
- Subagents can now spawn nested subagents up to depth 3 by default, up from 1
- Opus 4.7 removed from fast mode; /fast now applies to Opus 5 and Opus 4.8
- New sandbox.network.strictAllowlist setting blocks non-allowlisted hosts for sandboxed commands without prompting
- Dynamic workflows now default to a medium size guideline, configurable in /config or via the workflowSizeGuideline setting
-
Code review now runs in the background, plus many fixes
- /code-review now runs as a background subagent, so review work no longer fills your conversation
- Skills with context: fork now run in the background by default; opt out per skill with background: false
- Auto mode lets its classifier judge risky rm, background & and Windows path checks instead of opening permission dialogs
- MCP connection failures now show HTTP status and error text in claude mcp list and /mcp
- Fixed the left arrow key discarding the conversation with no undo, and Windows paths being corrupted into CJK characters
-
Emoji autocomplete in prompts and new limits on subagent fan-out
- Subagents are capped at 20 running at once by default, and subagents no longer start nested subagents unless you raise the depth limit
- Type an emoji shortcode like :heart: in the prompt to insert the emoji, with suggestions as you type; it can be turned off in settings
- Warnings now appear when transcripts fail to save, such as when the disk is full, so they are no longer lost silently
- The max budget flag now stops background subagents once the spending cap is reached
- Fixes a memory leak from truncated MCP tool outputs, Windows auto-update failures and auto-compact for Opus 4.8 on Bedrock
-
Sandbox setting to skip filesystem isolation, plus a long-session speedup
- New sandbox.filesystem.disabled setting skips filesystem isolation but keeps network egress control
- Fixed long sessions stalling for seconds and resuming slowly as message processing slowed with each turn
- Fixed auto mode denying commands with HTTP 401 errors after the OAuth token expired or rotated mid-session
- Fixed worktree-isolated subagents pointing git at the shared checkout, and resumed background agents losing their prompt and tool restrictions
- Tighter Bash and PowerShell permission checks, and /context now warns when the conversation exceeds the context window
-
Verify and code-review skills now run only when you invoke them
- Claude no longer runs the /verify and /code-review skills on its own
- Type /verify or /code-review yourself whenever you want them to run
-
Claude can now end abusive chats, plus a round of permission-check fixes
- New EndConversation tool lets Claude end sessions with highly abusive users or jailbreak attempts, as on claude.ai
- Closed several Bash and PowerShell permission-check gaps, and commands over 10,000 characters now always ask for approval
- docker commands with daemon-redirect flags and file commands using -m or -f now need permission first
- Long-running tool calls now send a periodic progress heartbeat instead of going silent
- Many fixes for background sessions, Windows PowerShell, MCP, plugins and OpenTelemetry
-
/fork now spins off a background session; new caps stop runaway loops
- /fork now copies your conversation into a new background session while you keep working; the old in-session subagent is now /subtask
- New per-session caps on WebSearch calls and subagent spawns (default 200 each) stop runaway loops, tunable via environment variables
- MCP tool calls running longer than 2 minutes now move to the background automatically so the session stays usable
- /resume in the agent view opens a picker of past sessions, including deleted ones, and resumes your pick in the background
- Fixed plan mode running file-modifying Bash commands without a permission prompt, plus many other fixes
-
Subagent text in stream-json output, plus a long list of fixes
- New --forward-subagent-text flag adds subagent text and thinking to stream-json output
- Always-allow permission rules now save at the repository root, so approvals carry across worktrees and sessions
- Fixed a prompt-caching regression on Bedrock, Vertex, Mantle and Foundry that billed system context as fresh input tokens
- Background agents now report the status of agents still running and wait for real completion instead of making up results
- A PreToolUse hook's ask decision is no longer overridden by auto mode for unsandboxed Bash
-
Live timers on long tool calls, plus a big batch of agent and worktree fixes
- Collapsed tool summaries now show a live elapsed-time counter, so long-running calls visibly tick instead of looking stuck
- Worktree-isolated subagents can no longer run git-mutating commands against the main repo checkout
- Auto mode's permission classifier now defaults to Sonnet 5 for external sessions, pinned for the session
- New startup warning for Write, NotebookEdit and Glob permission rules, which should use Edit or Read instead
- Many fixes for claude attach, background agents, plugin MCP servers, hook timeouts and plan approvals
Smaller updates 24
-
No release notes published.
-
Bug fixes and reliability improvements
-
Fixes for cloud session login and misleading prompt-length errors
-
Bug fixes and reliability improvements
-
Instruction-style prompt and agent hooks now block and stop as intended
-
Fixes for lost permission answers and lost messages on quit
-
Fixes all requests failing when routed through a proxy or gateway
-
Bug fixes and reliability improvements
-
Fixes read-only git commands unexpectedly asking for permission
-
Fixes gateway sign-in errors for proxy and API-key setups
-
Bug fixes and reliability improvements
-
Fixes launch failure on macOS Monterey and remote session errors
-
Bug fixes for Bash commands, permissions and Remote Control
-
Bug fixes and reliability improvements
-
Fixes startup crash on Linux distributions with glibc 2.44
-
No release notes published.
-
Bug fixes and reliability improvements
-
Bug fixes and reliability improvements
-
Fixes MCP OAuth sign-in for servers with pre-registered clients
-
Fixes wrong usage-credit prompts for Max users and Bash failures in CI
-
Bug fixes and reliability improvements
-
Bug fixes and reliability improvements
-
No release notes published.
-
Fixes /model and other dialogs in background agent sessions
Showing the newest 100 of 560.