Anthropic Release Notes

• Workbench is now Playground in the Claude Console. Playground supports every Messages API parameter and includes templates that demonstrate API features such as code execution and web search. It shows the full SDK request and the API response for each run, to help you understand the API and build with it. For more, see the Claude Help Center or try it at platform.claude.com/playground.

• Added an optional spellcheck setting that underlines misspelled words in the prompt input as you type, using your installed aspell, hunspell, or ispell
• Fixed whole-prompt-cache invalidation when a language server disconnected or reconnected mid-session
• Fixed nested markdown list items misaligning at depth 3+ and added a hanging indent to wrapped list items in the terminal UI
• Fixed prompt input highlights (slash commands, keywords, mentions) appearing shifted by one or more characters in some multi-line prompts
• Fixed Shift+Tab inside the permission prompt's comment field approving the edit and granting session-wide edit permission instead of closing the field
• Fixed the Agent tool advertising a general-purpose default in sessions where that agent is unavailable: an omitted subagent_type there now gets a clear error listing the available agents
• Fixed notebook cell delete/replace approval dialogs silently omitting the existing cell content when the notebook or cell could not be read; the dialog now says why
• Fixed slash commands run while Claude is responding showing HTML entities instead of the actual characters
• Fixed the prompt footer not showing the "Update installed" restart notice after a background auto-update
• Fixed the expanded task list (ctrl+t) always starting collapsed when resuming or relaunching into a session that still has open tasks
• Improved memory and CPU usage while cloud sessions such as /ultrareview or /autofix-pr run in the background — their event streams are no longer re-scanned and re-rendered on every update
• Improved permission dialogs: display text and "don't ask again" options now always match what a grant would cover, and "don't ask again" is withheld when contents cannot be fully displayed
• Improved the embedded grep in native macOS/Linux builds: pathological patterns now fail fast instead of exhausting memory, and -m N with -A/-C prints correct context
• Improved the context-limit error to say when auto-compact is off and point to /config to re-enable it
• Vim mode: NORMAL mode and cursor position are now preserved when toggling the detailed transcript (ctrl+o) or closing a panel
• Dialogs: arrow keys and Enter pressed in quick succession now select the option you navigated to instead of the previously highlighted one
• SendMessage now refuses messages too large for cross-session delivery up front instead of silently dropping them
• Remote Control: claude rc now applies the same enterprise-gateway availability check as interactive startup
• [VSCode] Fixed focus jumping between open Claude tabs on its own when a window with several Claude panels is restored or reloaded

Added

• Added the optional CLAUDE_CODE_PROJECT_DIR_NAME environment variable: hosts that give each session its own config directory can choose a short name for the per-project transcript directory
• Added the selection:clear keybinding action, so a key can be bound to clear an in-app text selection; also works in the agents view
• Added a GitLab merge request badge to the footer and statusline: repos with a GitLab remote and an authenticated glab CLI show MR !N with draft/pending/green states

Claude Code

• Claude Code now continues your session automatically when a claude.ai usage limit resets; turn it off in /config ("Continue automatically at usage limit")
• Claude is now told to use your account email only to identify you, and not to send it to unrelated services unless you ask

Security

• Security: remote file reads, session restore, CLAUDE.md includes, workflow scripts and file uploads now reject Windows NT-namespace (\??\) paths, hardening the remaining pre-approval file accesses against the NTLM credential-leak vector

Fixed

• Fixed auto mode in very long sessions repeatedly re-checking and denying sandboxed commands' network access after the conversation had been compacted
• Fixed session-scoped permission answers (including denies) being dropped when answering background subagent tool permission prompts
• Fixed a crash when an API response on the non-streaming fallback path (typically via third-party gateways) contained a thinking block missing its thinking field or a text block missing its text field
• Fixed markdown rendering becoming extremely slow for some messages containing unusual Unicode sequences
• Fixed SendMessage rejecting a recipient copied from ListAgents when the session name is at the 200-character cap or emoji-heavy
• Fixed repository detection mis-reading the host of git remotes with unusual userinfo, producing links and repo-specific behavior for the wrong host
• Fixed MCP diagnostics printing resolved secrets: scope-conflict warnings now show the configured ${VAR} form, and connection-failure details show only the server origin
• Fixed strictKnownMarketplaces allowlists accepting SCP-style git marketplace sources whose host differs from the one git would actually connect to
• Fixed modal text such as the /login OAuth URL losing characters when copied in fullscreen
• Fixed a --- horizontal rule in rendered markdown running into the line after it
• Fixed consecutive shell commands splitting into multiple "Ran 1 shell command" rows when todo/task updates were interleaved between them
• Fixed dialogs like /permissions opened while a ! shell command was running being dismissed when the command finished
• Fixed a queued ! shell command being sent to the model as plain text after pressing up-arrow to edit the queued input
• Fixed queued messages reappearing in the prompt history while still queued, Esc while selecting a queued message no longer interrupts the turn, and ! mode no longer sticks after a mid-turn submit
• Fixed accepting the "Try the new fullscreen renderer?" prompt restarting the session without its permission mode (e.g. --dangerously-skip-permissions), tool allow/deny rules, model or effort flags
• Fixed /tui dropping launch --allowed-tools/--disallowed-tools rules when it restarts; it now declines to switch, with the reason, when the session has restrictions a restart can't carry over
• Fixed trust prompts omitting the repository-wide scope warning when the directory was first seen before the repository existed there
• Fixed a case where an IDE diff tab closing during a permission re-prompt could answer the new prompt with the previous input
• Fixed: files sent to the user during Remote Control sessions hosted by Claude Code Desktop or VS Code now upload, so they open on phone and web instead of showing an empty card
• Fixed: after /login while CLAUDE_CODE_OAUTH_TOKEN is set, the stale-token reminder no longer leaks into Claude's automatically resumed turn — it now appears only to you
• Fixed: permission previews now relay only to channel servers admitted by the inbound trust gate, and a server's explicit permission-capability opt-out is honored
• Fixed: credential masking on relayed permission previews can no longer hide commands, paths, or destinations from the approver; oversized private-key blocks now redact under full-strength redaction
• Fixed: provider API tokens that mask on permission previews now mask even when directly followed by shell delimiters
• Fixed Claude Desktop inter-session messages being silently dropped by the recipient session when cross-session messaging read as disabled, which left the sender's query "thinking" for many minutes
• Remote Control: signing this computer in to a different claude.ai account or organization now stops the running session within seconds and says why, instead of a misleading HTTP 404 hours later
• Remote Control sessions started from Claude Code Desktop or VS Code now keep phones and claude.ai/code updated on the session's permission mode (and claude.ai/code on the model) as they change
• Remote Control: effort picks made on a phone or on claude.ai/code now apply to terminal- and Desktop/VS Code-hosted sessions, and the session publishes its effort level to connected clients
• SendMessage and ListAgents now say when your account's session list was too long to check completely, instead of treating unseen sessions as absent
• Expired Anthropic profile credential now points you at /login when a claude.ai login would take precedence
• Improved the transcript: your own prompts now render markdown (highlighted code blocks, inline code, lists) the same way replies do
• Improved the "API returned an empty or malformed response" error to say what came back (content type, body kind, size, request ID) and why the original streaming request failed
• Improved auto-generated session titles to read as short, specific names (e.g. "Login button bug") rather than sentences restating your request (e.g. "Fix the login button on mobile")
• Reduced the context cost of loading the built-in claude-api skill from ~200k+ tokens to ~25k by loading reference docs on demand
• /permissions can now be opened while Claude is working — rule changes apply to the rest of the current turn
• /add-dir <path> can now be used while Claude is working; /add-dir, /autocompact, /theme, /help, /config and /advisor dialogs open mid-turn in the fullscreen TUI
• /goal now clears itself with a notice when a turn dies on an unrecoverable error (e.g. revoked auth, an exhausted credit balance, or a context overflow) instead of staying armed
• /goal: when background tasks keep a goal waiting for 30+ minutes, Claude now checks in on them instead of waiting indefinitely (set CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0 to opt out)
• claude setup-token now rejects unexpected extra arguments instead of silently ignoring them
• Changed Esc in fullscreen mode to no longer clear a mouse text selection: it interrupts or dismisses as usual and the selection stays highlighted
• Removed the redundant "Allowed by auto mode classifier" line that auto mode showed under every Agent tool call
• Removed the "Default teammate model" setting from /config; agent-team teammates now use the leader's model unless the spawn names one
• Dimmed the elapsed-time counter on the running tool header so it no longer competes with the bold counts
• Background task notifications delivered between turns are now sent to the model inside <system-reminder> tags, matching mid-turn delivery
• Mantle: skip the admin-pin availability probe at startup when a main-loop model is already picked
• Windows: startup no longer stalls on repeated rename retries when ~/.claude.json is read-only

• Added GitLab merge request URL support to the --worktree flag and the claude agents view (where MRs display as !N)
• Added an opt-in forward_user_identity apps gateway setting on Anthropic upstreams that sends the signed-in user's identity as headers, so a proxy behind the gateway can attribute spend per user
• Added opt-in memory cgroup support for Bash tool commands on Linux (CLAUDE_CODE_TOOL_MEMORY_LIMIT) so a runaway build can't stall the session
• Added CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS environment variable to configure the WebFetch session URL cache TTL (default unchanged: 15 minutes)
• Fixed cloud sessions occasionally being marked as lost when the environment shut down while Claude was waiting on a permission prompt
• Fixed MCP v2 connections endlessly reopening the subscriptions/listen stream against servers that terminate long-held streams on a fixed timeout (e.g. serverless hosts)
• Fixed Notification hooks not firing for permission prompts when running under Claude Desktop or VS Code
• Fixed idle sessions on Linux sometimes keeping one CPU core at 100% when sandboxing is enabled
• Fixed bundled skill aliases like /checkup and /review reporting "Unknown command" in -p mode or with plugins/MCP loaded when a user or project skill shadows the bundled skill
• Fixed skill/command argument substitution to prevent argument values from being re-expanded as template markers
• Fixed Windows paths spelled with the NT \??\ device prefix bypassing UNC path validation, closing an NTLM credential-leak vector
• Improved claude self-hosted-runner session start time: the session branch is now created without rewriting the working tree, and two server round trips no longer block the agent's launch
• Improved apps gateway error forwarding: 400/413 errors from Vertex, Foundry, and Claude Platform on AWS upstreams now carry the upstream's own message; fixes a bug with auto-compact on apps gateway
• Improved claude plugin validate to check a bare .claude/skills directory, reporting SKILL.md files whose frontmatter fails to parse
• Improved screen reader mode: the /effort selector renders as a numbered list with a typed-number prompt, and hint and dialog text is no longer clipped
• Improved print mode diagnostics: a [claude-code:unrecognized_model] line is written to stderr when a request goes out for a model ID Claude Code doesn't recognize; map it with modelOverrides to silence
• Changed the GitHub app setup tip to no longer appear in repositories whose origin remote is on gitlab.com or bitbucket.org; the enterprise marketplace tip now covers non-GitHub internal git hosts
• Todo/task-tracking tools (TaskCreate/Get/Update/List, TodoWrite) are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to bring them back
• Windows: fixed auto mode repeatedly stopping for manual approval on ordinary cd <dir> && <command> > file Bash commands (a 2.1.232 regression)
• Reverted the 2.1.232 Bash permission changes for Cygwin-style symlinks on Windows and for input redirections (< file); a narrower version will return in a later release

Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act.

In this article, we share answers to some of the questions we’ve received about how our chosen watermarking method works, whether it affects Claude’s outputs, and why we’re making this change. To summarize:

• We use a method of watermarking that does not have any practical impact on the quality or content of Claude’s outputs;
• The difference between watermarked and un-watermarked text will not be distinguishable to readers;
• Nothing is added to the text and there are no hidden characters;
• Watermarking doesn’t require extra tokens, and will not be more expensive;
• Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat;
• Watermarking won’t be specific to Claude. As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.

What is watermarking?

Large language models like Claude work by generating one word at a time. Each time the model decides on the next word, it chooses among a list of potential candidates, ultimately selecting the most sensible or likely based on the preceding text. Take the sentence “The weather today was cold and…”. The next word is very unlikely to be “sugary.” But it is quite likely to be “overcast” or “grey.” Under most circumstances, it doesn’t matter much to the reader which of these latter two words the model ultimately chooses—the meaning of the sentence is largely the same either way. In cases like this, the choice is settled by a random number.

Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it. When watermarking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermaking uses the key and a few words that come before to settle what word the model should pick. That is, the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.

Importantly, it isn’t that the model will now always be biased toward overcast or grey. Just as with non-watermarked text, overcast might be selected in one sentence, grey in the next, depending on the words that came before. And it’s not the case that the watermarking method pushes Claude to choose a word it wouldn’t have considered anyway (for instance, it wouldn’t make Claude pick a word like “nubilous”—an obscure1 synonym for overcast or grey that Claude almost certainly wouldn’t use under normal circumstances).

How does watermarking affect Claude’s outputs?

Watermarking does not impact the quality of Claude’s output. To a reader, a watermarked response is indistinguishable from an unwatermarked one (in this way, AI watermarks differ substantially from their namesakes on banknotes, other physical objects, and some digital documents, which are visible to the naked eye).

In internal testing, we’ve seen no impact of watermarking on the content, level of creativity, or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique we use, Google DeepMind tested this impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model. And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.

A useful analogy is to imagine you’re playing a game like Monopoly. On each turn, each player moves a random number of spaces around the board according to the roll of a die. Suppose that, instead of rolling the die to get this randomness, we decided to use a book of the digits of pi.2 We start from a randomly-chosen digit (say, the 1,012,845th after the decimal place, which happens to be a 6), and from that point on each player simply uses the next digit in the sequence as their next “roll”.

For all intents and purposes, the moves are still random: it makes no difference to the players—or to the outcome of the game—whether the randomness comes from pi or from dice rolls each time. But if we could see the sequence of all the moves after the game (and we knew the value of pi), we could work out whether this was a game that likely used pi to determine its moves. The game that used pi is, in a sense, “watermarked”.

It’s the same for Claude-generated text. Watermarking doesn’t change the meaning or experience for the person reading it, but if you wanted to check after the fact whether the text was likely generated by Claude, the watermark allows you to do so.

Which specific method of watermarking do you use?

Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024. It belongs to a family of approaches that go back to a proposal by Scott Aaronson in 2022, all of which share the same design principle that we described above—the watermark only changes the source of the randomness used to pick among words.

There are limitations to the effectiveness of watermarking. Using our key, one can only answer the question “What is the likelihood this was partly written by Claude?” It doesn’t confirm whether the text was human-written, and it can’t tell whether the text was written by a different AI (even if that other AI uses watermarking, it would have a different key; it might also use a different watermarking method altogether). Detecting a watermarking also doesn’t work well on small samples, where there are fewer word choices and thus less information to go on. As a passage increases in length, confidence about Claude’s involvement increases too.

Watermarking is sparser on factual passages where there are fewer choices that can be made without decreasing the accuracy of the text. For example, take the sentence “Isaac Newton’s most famous work was called Principia…”. It really matters whether the next word is “Mathematica” (it’s the only right answer), so the watermark would have nothing to act on. The same is true for proofreading. If you hand Claude a piece of writing and ask it to edit only the grammar and punctuation and nothing else, the watermark can only live in the handful of corrections, which might be too few to register.

What about cases where Claude has proofread or edited human text?

The watermark only applies to words Claude chooses. When Claude proofreads text written by a person, what it gives back has generally only been lightly edited; because nearly all the words are the person’s, there’s very little (if anything) for the watermark to attach to. Depending on the length of the text and how heavily Claude has edited it, those changes might not be enough to make Claude’s involvement detectable. The more Claude writes, the more decisions it has to make, and the more space there is for a watermark.

What about code?

As we noted above, AI watermarking takes advantage of decisions where either choice of a word would be equally good. Where an exact output is required—where there isn’t a choice, and something would be factually wrong or a piece of code would break if a different term was chosen—the watermark isn’t applied.

For example, once the model has written “2 + 2 =”, there is a very clear best choice for the next token (if the model is completing the sum, there isn’t an answer that’s equally as good as “4”; if it’s talking about George Orwell’s Nineteen Eighty-Four, there isn’t an answer that’s equally as good as “5”). The “nudge” of the watermark wouldn’t be applied here. For the same reason, code—which in very many cases has to be exact—has generally less watermarking than some other forms of text.

Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code. But by definition, it will have a negligible effect on the actual code produced.

What does this mean for users?

Does this slow the model down, or make it more expensive?

No. Watermarking has a negligible impact on the speed of models, and because it produces no extra tokens, the model is the same price to serve and use.

Can a watermark be traced back to me or my organization?

No. The watermarking applies to Claude and its outputs. It doesn’t identify anything to do with individual users. There’s nothing in the watermark, or its key, that would allow anyone to recover any information about the user, their organization, or their chats with Claude.

Why are you watermarking Claude’s outputs?

We’re implementing watermarking to comply with the EU AI Act. Anthropic, along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. This requires AI system providers to use methods of “marking” AI-generated text. We’re applying watermarking globally at launch because we don't yet have a durable way to scope it by region. However, we will continue to evaluate different approaches, and will share updates when we have them.

Other questions

How do I check if a piece of text was written by Claude?

We will soon be offering a watermark detection API. We’re in the process of working out the details of its implementation.

What about images and other files?

When Claude produces a file of a supported type (such as a .png, .jpg, or .svg), it will attach a content credential in the form of a small, cryptographically signed note in the file’s metadata, saying that the file was made or processed with Claude. This is an open industry standard called C2PA—the same used by camera manufacturers and in photo-editing software to record where an image came from. Any C2PA-aware tool can read it; we’ll be providing our own where you can drop a file and check.

This metadata label is very different from a watermark. Nothing in the file changes—it is not embedded or hidden. As with text, the credential only says Claude was involved in producing the file; it doesn’t include any identifying information.

Can’t someone just edit the text to get around the watermarking?

To some extent, yes. Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will. In the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated.

What does a watermark actually prove?

A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish “Claude wrote this” from “Claude heavily edited this.”

Do watermarks apply to translations?

Yes. A translation produced by Claude carries a watermark, because in this case every word is chosen by Claude.

What about older Claude models?

The EU law includes a transition period for Anthropic models launched before August 2, 2026, and we’re working to add watermarking for those models as well. This will be rolled out over the coming months.

How does this differ from AI detection software, like Pangram?

AI detection software uses a different method, because the companies that provide it don’t have our key. Among other things, those services look at aspects of the text like the subtle (and not-so-subtle) “tells” that often appear in AI’s phrasing. For example, AI models appear to be fond of the construction “this isn’t [X], it’s [Y]”, and use the word “quietly” a lot more than you might expect. Picking up on these patterns is fundamentally different from checking for a watermark.

Does this change who owns a given output, or who is legally responsible for it?

No. A watermark only helps test whether Claude might have produced or processed the content. It doesn’t say anything about ownership or authorship, and doesn’t change a user’s rights under our terms. We only apply the watermark when Claude was involved in processing the content or file.

• Subagent forking is now on by default: a subagent_type: "fork" subagent inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions now run in the background by default
• Type @ in the prompt to mention another Claude session by name; Claude then uses SendMessage to reach that session directly
• SendMessage now delivers to a bare name that exactly matches one live session, instead of asking to confirm with a ref first
• Interactive sessions on one machine now keep unique names: starting or renaming a session to a name another live session already uses gives it a name-word-word variant and tells you
• Added /config rows for "Dialog expiry" and "Messages from your other sessions" (cross-session inbound accept/hold/refuse)
• Added secret redaction for GitLab token families (glrt-, gloas-, glptt-, glagent-, glimt-, glsoat-, glcbt-, glft-, glffct-) and full redaction of routable glpat-/gldt- tokens; the glab CLI config store gets the same sandbox and credential-path protection as gh
• Added GitLab support to plugin marketplaces: bare gitlab.com repo URLs (including nested subgroups) now clone like github.com URLs, and clone auth-failure hints name your actual git host
• Settings: additionalMarketplaces and allowedMarketplaces are now accepted as friendlier aliases for extraKnownMarketplaces and strictKnownMarketplaces
• Enterprise policy: a url-typed blockedMarketplaces entry for a bare repo URL keeps blocking that URL when the CLI classifies it as a git clone
• Gateway: the desktop: overlay now accepts every released Desktop setting (was 11 hand-listed keys), validated at boot against Desktop's own schema; unknown or invalid keys fail boot
• Gateway: empty managed.policies[].match.groups/admin.admin_groups entries and malformed email_domain values (empty, or containing @, whitespace, or commas) now fail at boot instead of silently matching no one or granting admin access
• Fable 5 is offered as an advisor in /advisor again for organizations with Fable access, with usage-credits consent set up through /model fable
• Fixed a PowerShell permission bypass where variable-writing parameters could silently overwrite $PSDefaultParameterValues and redirect later commands' file access
• Fixed a Windows permission bypass where Git Bash followed Cygwin-style symlinks that path validation saw as regular files; writes through them now require permission approval
• Fixed nested git repositories inheriting trust from a parent directory; each repository now requires its own trust confirmation
• Fixed MCP connections hanging for the full 30-second connect timeout when a server fails to answer or sends a malformed reply to the protocol-version probe
• Fixed Remote Control sessions hosted by a bridge inside a cloud session inheriting that session's transcript or credentials
• Fixed Remote Control sessions started from Claude Desktop or an IDE appearing as a new claude.ai session each time the local session was resumed; they now reattach to the existing one
• Fixed Remote Control sessions appearing unreachable to newly attached clients while idle
• Fixed Remote Control bridge sessions not restoring conversation history when the session worker restarts
• Remote Control: resuming a conversation whose session was deleted from claude.ai or the app now starts a replacement instead of failing with a message about your login (regressed in v2.1.227)
• Fixed Cloud gateway /login exiting silently or leaving an unresponsive terminal after "Press Enter to continue" when managed settings failed to load; the reason is now shown
• Fixed voice mode on native builds getting stuck on "listening…" when the voice service rejected the connection; the rejection is now shown immediately
• Fixed plugin and settings panels not moving the terminal cursor to the focused row, so screen readers and magnifiers can follow arrow-key navigation
• Fixed the transcript preview sitting flush against the input area when attaching to a starting background session; it now leaves the same one-line gap as the live layout, so the transcript no longer shifts when the session takes over
• Improved footer PR badge links to be clickable hyperlinks even when terminal support can't be detected (e.g. over ssh/tmux); set FORCE_HYPERLINK=0 to opt out
• Changed the login-expiry warning to appear 3 days before expiry instead of 5
• Capped the frontend-design plugin suggestion tip at 3 lifetime impressions instead of repeating indefinitely
• Added a cap on concurrently-running subagents (default 20, override with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS) so one message can't fan out unbounded background agents
• Changed subagents to no longer spawn nested subagents by default; set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to allow deeper nesting
• Fixed --max-budget-usd not stopping background subagents: once the cap is reached, new spawns are denied and running background agents are halted

Claude Tag lets you add Claude to a Slack channel, where it works alongside your team. Claude responds when you @-mention it, or proactively when it thinks it can be helpful.

Before, Claude only saw one message at a time, so it made decisions to act proactively based on what was in front of it, but not the wider context of what was around it.

Now, Claude uses context from across the channel, as well as its memory and the standing instructions you have given it, to determine when to contribute to the conversation.

As a result, Claude is now roughly 30% better at determining when, and when not to, proactively respond.

This update comes at no additional cost today. While holding more context does increase Claude Tag’s usage, the additional context Claude Tag holds does not count toward usage or spend limits on any plan.

From passive responder to active participant

Previously a lightweight classifier decided when Claude should act. It looked at each new message on its own and made one yes-or-no call.

For example, here are two engineers chasing the same bug from opposite ends. Neither has a free hour to run it down, and neither message asks for anything.

Priya has a theory. Devon has the evidence. Neither message is for Claude, and neither asks for anything.

Read one at a time, neither message is for Claude, so the classifier correctly does nothing, twice. Read together, there's an obvious piece of work sitting there. One engineer has a theory, the other has the evidence for it, and nobody has time to check.

With the classifier removed, Claude uses context across the channel to make one of four moves:

• Reply inline, when the answer is short, verifiable, and something the channel doesn't already know.
• Start deeper work in a thread, when a message deserves real time.
• Route the message to work it has in flight, when it adds to a workstream Claude already has open.
• Say nothing, when nothing is called for.

Here's the same conversation with Claude Tag using additional context. Claude picks the second move, even without being @-mentioned. It sees Priya's hypothesis and Devon's evidence, opens a thread with the investigation already running, and pulls both engineers in. It acts within the boundaries of the permissions, tools, and scope you have configured.

Same thread, two minutes later. Claude reads the two messages together and starts the work. No @-mention.

The conversations aren't walled off from each other. So when Devon posts an update, it lands in the right workstream. When two investigations turn out to be the same bug, that connection gets made.

Claude now looks at all messages to understand the full context of the channel, to more accurately determine if it should participate in a conversation unprompted.

How Claude decides when not to speak

An annoying agent is worse than an unhelpful one. We built Claude Tag to speak up only when it's useful, and in most channels, on most messages, that means saying nothing.

We do this by grading Claude’s channel-by-channel choices against a rubric based on principles like how useful the comment is, how confident Claude is in the response, and whether there is a person better suited to respond.

Claude also knows when to stop paying attention, similar to how people navigate Slack. It follows a few channels closely while paying less attention to others until someone tags it in. In a channel where, message after message, Claude keeps concluding it has nothing to add, it goes to sleep. A @-mention wakes it instantly.

You can also steer its response behavior in plain language: "Never respond here unless someone tags you," or "Feel free to jump in on anything about the deploy pipeline."

And if you'd rather Claude only spoke in a channel when someone tags it, any member can switch ‘Respond automatically’ off.

The first reply is faster

The additional context also allows Claude to respond more quickly. It acknowledges you in seconds instead of operating silently while it starts up. The work itself takes as long as it always did; what's gone is the silent first minute when you couldn't tell whether it heard you.

Live today

This update is now available across Claude Tag, available for Claude Teams and Enterprise customers. You can get started here. Claude now acts as a more effective collaborator, one that can follow the conversation, decide for itself when to act, and when to stay out of the way.

Add Claude to one channel and watch what it adds to your conversations. Learn more about Claude Tag.

Release notes

• Documented claude remote-control --continue for resuming the most recent Remote Control session
• Added server-supplied Claude Code hook support for self-hosted runner sessions, matching managed-environment behavior
• Added SSE keepalive pings to gateway streaming responses during long thinking pauses, preventing idle-timeout disconnects on Vertex and Bedrock upstreams
• Added plugin marketplace command sources: a local command (e.g. an IDE) prints the plugin directory, which is re-resolved each session and applied without a restart; mode: "link" uses it in place
• ListAgents now marks disconnected Remote Control sessions as offline and labels your cloud sessions as cloud
• Fixed long responses partly disappearing while streaming and being printed twice in the terminal
• Fixed a crash to the error screen (including on --resume of the affected session) when a tool call had a non-string glob, file_path, or command value
• Fixed a RangeError crash when a progress bar or markdown table rendered in a very narrow terminal window (could also crash claude --continue/--resume at startup)
• Fixed a crash on Windows when a tool call or message referenced a file by an extended-length (\\?\) or UNC path
• Fixed auto mode failing on every tool call for users who disable the attribution header via CLAUDE_CODE_ATTRIBUTION_HEADER (direct Anthropic API connections)
• Fixed /model rejecting Sonnet/Opus 1M for claude.ai subscribers using a custom ANTHROPIC_BASE_URL gateway
• Fixed MCP OAuth with strict authorization servers by using 127.0.0.1 instead of localhost in the redirect URI
• Fixed Remote Control clients showing a stuck working spinner after a slash command typed in the laptop terminal
• Fixed the Claude Code Review workflow generated by /install-github-app completing without posting its review on the pull request
• Fixed multi-second UI stalls after editing a file with thousands of IDE diagnostics while the IDE extension is connected
• Fixed one-shot claude plugin commands leaving a stray liveness file that could prevent cleanup of outdated plugin versions
• Fixed dynamic workflows inside CPU-limited containers using the host machine's core count instead of the container's CPU limit
• Fixed a file-watcher handle leak after atomic file replacements, and an uncaught error on Windows when the scheduled-tasks watcher failed on a network or virtual filesystem
• Fixed usage-limit guidance suggesting unavailable slash commands in SDK and remote sessions
• Fixed the consent message for interactive --advisor fable launches, which told you to run /model fable in an interactive session that had just exited
• Improved fullscreen streaming: long sessions stay responsive because the whole conversation is no longer re-normalized on every update
• Improved the managed settings approval dialog: shows endpoint URLs, uses clearer wording for telemetry-only changes, skips routine OpenTelemetry options, and requires approval for server-managed sandbox binary overrides (sandbox.bwrapPath, sandbox.socatPath, sandbox.ripgrep)
• /feedback and /bug now open immediately when invoked while Claude is responding, instead of waiting for the turn to finish
• /plugin install plugin@marketplace now refreshes the marketplace first, so newly published plugins install without a manual marketplace update
• /code-review at high, xhigh, and max effort now runs in a background agent like the other levels
• Pasted and clipboard images are read without blocking the event loop
• Remote Control now keeps reconnecting for about 30 minutes after a network blip and no longer drops after a few blips spread across an hour
• Remote Control: resuming a conversation no longer silently takes Remote Control away from another Claude Code on the same machine that still has it; run /remote-control there to move it
• Updated agent panel: completed subagents hide immediately with a /tasks footer hint, and the "↓ N more" overflow indicator moved left for visibility
• Remote Control: the terminal now says whether a session was taken over by another device, ended from another app, or deleted, and stops suggesting a reconnect that would undo it
• Bash input redirections (< file) are now permission-checked like their argument spellings on all platforms
• Shortened the message shown when resuming a completed background agent
• Cowork sessions no longer inline external @-imports from user-scope memory files
• Hardened the auto-generated cross-session messaging socket directory on shared /tmp: a pre-planted symlink or another user's directory is now refused instead of used
• Hardened the Linux filesystem sandbox against a protected-path bypass
• Changed sandbox.ripgrep to be honored only from user, managed, and --settings settings; project settings can no longer override the sandbox's ripgrep binary
• Removed the startup tip suggesting you create custom subagents, and the matching nudge in the /powerup tour

Claude in Chrome

The Claude in Chrome side panel is now a Claude Cowork session. Conversations are saved to your history, your skills and connectors work in the browser, and a task you start in a tab can be finished on the Claude desktop, web, and mobile apps. It’s available on Max and Team plans today, and is rolling out to Pro users over the coming weeks.

Claude in Chrome is a browser extension that lets Claude see the page you're on and take actions in it, including clicking links, typing text, navigating between pages, and filling out forms, using your existing logins.

Many of the tools you use every day connect directly to Claude, but others don't, such as internal dashboards, legacy systems, and vendor portals. With Claude in Chrome, Claude can work in these apps through the browser.

Until now, a session in the side panel was separate from those in the Claude apps, so context and conversations didn't carry between them. Now, the side panel runs the same Claude Cowork session you use on desktop, web, and mobile for longer, multi-step work. Because sessions live with your account rather than a single device, you can start work in a browser and pick it up later somewhere else.

As an example, say you're putting together a budget spreadsheet and need to pull in invoices from several vendor portals. Now, you can ask Claude in Chrome to collect the amounts and dates, and it will open the tabs, read each invoice, and build the spreadsheet. Then, you can pick the session up in the desktop app to add files from your computer, or import last month's budget and ask what's changed, allowing you to maintain context across surfaces as you work.

Understanding the risks

Claude in Chrome carries the same risks as any AI agent that acts in a browser, chiefly prompt injection. Malicious actors hide instructions in web content, such as a web page, an email, or a document. These instructions may not be visible to you, but they can redirect Claude to take actions you never intended.

Since the pilot, we’ve added a check on Claude’s own actions. Use “automatically approve” and Claude works through a task without stopping for permission at every step. Before anything consequential, like submitting a form, sending a message, or downloading a file, a separate check reviews the action against what you originally asked for and blocks anything that doesn’t match. That creates fewer interruptions while maintaining oversight.

Claude still asks before certain irreversible or costly actions, like making a purchase or sharing personal data. While these measures meaningfully reduce the risk, they cannot eliminate it. Prompt injection is a moving target, so we keep hunting for new attacks and building what we learn into each model we release. We recommend starting on sites you trust, and our safety guide has more best practices.

Getting started

To start using Claude in Chrome, install it from the Chrome Web Store, sign in, and open the side panel. The new side panel is available on Max and Team plans today, and is rolling out to Pro users over the coming weeks. On Enterprise plans, Claude in Chrome is off by default. Admins can turn it on and limit it to approved domains. See the admin setup guide.

You’ll still need to use the Claude desktop app to work with files on your computer or with other applications. Claude in Chrome doesn’t run on other Chromium browsers or on mobile yet.

• Fixed interactive sessions that could stop redrawing entirely, while the process kept running, after a rare internal layout error
• Fixed git / Git Bash not being found on Windows when Claude Code is launched from a parent folder of the git installation
• Fixed /tui reverting the session to an earlier model when /model had been changed since the last response
• Fixed cross-session messaging sometimes starting without an inbox in the first session after install or upgrade
• Fixed Remote Control /resume while connected leaking the resumed conversation's title or history into the connected session
• Fixed claude self-hosted-runner sessions failing on every fresh runner when the checkout hook fails for a repository the session doesn't push to; that repository is now skipped with a warning
• Fixed self-hosted runners ending sessions in the gap between a background task finishing and the follow-up turn starting
• Fixed session cleanup deleting contents inside a project's memory folder
• Fixed background plugin-cache cleanup deleting a plugin's cache when its only version is a symlinked development checkout
• Fixed a settings-merge issue where a marketplace entry redefined in a higher-precedence settings tier could inherit another tier's custom headers; marketplace entries now merge as whole entries
• Fixed the deferred-tools reminder occasionally being sent to the model twice after a skill invocation
• Hardened skills synced from claude.ai: they no longer shadow local commands or MCP prompts, their descriptions are sanitized and labeled, and on your machine their bodies don't run ! commands or expand @ files
• Improved cross-session messages: the sender and body now display inline instead of a collapsed line, and messages to Remote Control sessions on other machines show your Remote Control session name as the sender
• Improved Vertex AI credential handling: expired or missing Google Cloud credentials now fail within seconds instead of retrying for minutes
• Improved compaction progress: the retry countdown and stall hint now appear during compaction instead of only a progress bar
• Updated terminal title busy-spinner glyphs to reduce tab-bar jitter on some terminals
• Changed the Write tool so newer models can overwrite an existing file they haven't read this session, matching the Edit tool's rules; older models still require the read first
• Removed the outdated note about auto mode sessions costing slightly more from the first-use notice for Pro, Max, and Team plans

• The Compliance API now returns transcripts of Cowork and Claude Code sessions that run on your users' machines, in beta for Claude Enterprise organizations. GET /v1/compliance/apps/sessions/local lists sessions across your organization, GET /v1/compliance/apps/sessions/local/{session_id} retrieves one session's metadata, and GET /v1/compliance/apps/sessions/local/{session_id}/messages returns its transcript, all with your existing Compliance Access Key and the read:compliance_user_data scope. See Sessions on users' machines.

• We've added the anthropic-workspace-id response header to the Claude API. It carries the wrkspc_-prefixed ID of the workspace that the request's API key or access token resolved to, including your organization's Default Workspace. See Identify the workspace behind an API response.

How it works

Claude's Compliance API now covers Cowork across the desktop app, web, and mobile, as well as Claude Code in the CLI and desktop app. Coverage is in beta for Claude Enterprise customers. Compliance and security teams can pull session content and metadata from both products through the same Compliance API interface they already use for Claude chats.

The new endpoints are additive: nothing changes about the data you already pull from the Compliance API today.

Security and compliance teams rely on the Compliance API to see how Claude is used across their organization — for audits and eDiscovery — without deploying separate logging infrastructure for each surface. Extending coverage to Cowork and Claude Code closes a gap: those sessions now show up alongside Claude chats.

The new session endpoints return a consolidated, server-hosted transcript for each Cowork and Claude Code session, so prompts, responses, and tool activity come back together in a single session record.

Each session record carries two kinds of data:

• Session content: prompts and responses, tool calls content (web and MCP), and skills and artifacts content captured as transcript text.
• Session metadata: verified user ID and email address, organization ID, session and per-message IDs, and timestamps.

This beta doesn't include Claude Code on the web, Claude Code accessed through the Claude Platform, or sessions run on Amazon Bedrock, Google Cloud's Vertex AI, or Microsoft Foundry.

Organizations already exporting OpenTelemetry data can keep it running: the Compliance API can work alongside it with no infrastructure required on your side.

Getting started

Coverage for Cowork and Claude Code is available today and included with the Compliance API using your existing Compliance Access Key – there’s no separate integration to build. If it's already enabled for your organization, query the new session endpoints directly. If not, review the Compliance API documentation to enable it.

• Fixed feature flags being evaluated without the user's subscription tier when a session started with an expired login token, which could wrongly prompt Max plan users to enable usage credits for Fable
• Fixed every Bash command failing under claude-code-action with allowed_non_write_users on GitHub-hosted runners
• Fixed /tui bringing back a conversation that had been rewound to before its first message
• Improved slash-command menu: blue now marks only the selected row, matched characters are bolded instead of recolored, and emoji or accented names keep their glyphs
• Improved performance: fewer event-loop stalls on file-not-found suggestions and at-mention size checks