You can now load claude.ai connectors while signed in with an API key by setting CLAUDE_CODE_I_ACCEPT_CROSS_ORG_DATA_RISK. Claude Code shows a warning when you do this. On macOS, setting CLAUDE_CODE_NATIVE_APP_BUNDLE lets the native updater install Claude Code as a signed claude.app bundle. A new takesMcpToolHooksOnConnectors setting lets MCP tool hooks run on connector servers. In-process SDK MCP servers marked as modern now connect by skipping the setup handshake. This release has 480 entries, and 330 of them are usable today.
There are 13 entries in the build that are not switched on yet. Giving claude.ai connector tools a local file path, so Claude Code uploads the file itself, is off by default. A rewind for pending turns sits next to the Esc take-back handling, but its actions do nothing yet. Claude Code can tell the model which Claude app your message came from, such as iOS, Android, desktop or web. Reusing earlier messages to keep the prompt cache intact is controlled by CLAUDE_CODE_NORMALIZE_PREFIX_CACHE. Several new server-controlled switches arrive, including ones for Projects write warnings and Chrome tool recognition.
/autofix-pr now always starts a cloud session. Plugins that call ui.backdrop will break, because that method is gone. Many SDK session-control methods are removed, including the sign-in, MCP server and plugin reload ones. Claude in Chrome no longer offers shortcuts_list or shortcuts_execute for running browser shortcuts. Auto-compact now counts some failures separately, so it gives up less quickly. Refusals caused by an old credential no longer trigger retries.
Written by our agent from the shipped bundle, not by Anthropic.
The earlier-subagents list now falls back to on without a server value, and a new CLAUDE_CODE_CRYPTIC_HORIZON variable takes over one of its checks
Sessions & agentsCloud sessions can skip rereading the transcript on resume via CLAUDE_CODE_CCR_SKIP_HYDRATE_ON_CURRENT_TIP, plus four new stream timing switches
Sessions & agentsAdmins can set pooled_conflict_rule to priority_list with groups in pooled_priority_order, and a list without any group is now rejected
ElsewherePlugins now set and clear a backdrop through theme.set and theme.clear, and the old ui.backdrop method is removed
ExtensionsBehind tengu_mossy_melody, Windows splits Bash commands too long for one argument into pieces, and disableWindowsBashCommandSplit lets admins turn it off
Settings & config
/autofix-pr always starts a cloud sessionExtensionsWant the reasoning? Read walks the 57 entries that probably matter to you, each one opening to what changed and why.
Read this release → Every row →14 more of these are in What probably matters to you, on page 1.
A claude -p run in auto mode now prints an error and exits with code 1 if auto mode could not check some actions, so they did not run
When one MCP server supplies many tools, a listing of them can now show one short line per server with a count
Unclear It is not clear where this shortened listing appears, so a reader cannot tell whether they will see it.
/model picker can hide part of a long list behind a More models row#When the /model picker list is long, some models are now hidden behind a More models row that expands the full list
Unclear The list size that triggers the collapse is not stated, so it is unclear who will see it.
When you leave a worktree, Claude Code now keeps it if its connector-downloads folder holds files, and tells you to move them first
Text from WebFetch no longer includes the paragraph telling the model to treat fetched pages as untrusted data
Unclear The warning may have moved into the "Fetched" line rather than been removed.
Allow, deny and ask rules for shell commands now compare commands through a shared normalising step instead of simple lowercasing
Unclear Which commands are now matched or allowed differently is not known.
A new startup notice says your claude.ai login ended and usage now bills to an organization, and points to /login
After a plan is approved, Claude is told you may have edited the plan document and must read it first, treating removed steps as intentional
Moving a local session to the cloud now prepares a summary first, and you can type your next prompt while it finishes
Results reported through the SDK now have a field naming which of your own auto mode rules decided a tool call
The fleet view no longer refuses to pin a session running in another terminal, and shows "esc then enter to open and answer" for some replies
Wildcard Bash permission rules now get an extra check, so a compound or unusual command no longer matches them
Unclear Which wildcard rules get the extra check is not stated.
Certain env keys in settings that name a use case are now dropped, and Claude Code logs a warning for each one
Unclear Which environment variable names fall into the ignored set is not stated.
The fleet view no longer refuses to attach to or group a session that is running in another terminal
A PermissionRequest hook's allow is now honoured where no prompt can be shown, and a hook-rewritten call that cannot be re-prompted is denied with a reason
Unclear It is not clear whether other checks still stop hooks from approving calls meant for a person only.
Auto mode now says when it refuses actions it cannot check, and a session can end because it moved region with your organization
project_info when a Project's files fail to list#If a session is attached to a Project whose files could not be listed, Claude is told to call project_info before answering about them
Claude's guidance for published pages now says to list available connector tools and show a feature as unavailable when its tool is missing or blocked
The message for Cyber Verification Program members now asks them to send feedback with /feedback instead of linking to a feedback page
Claude Code no longer restarts an out-of-date background session that is using an LLM gateway, so it is not cut off from that gateway
When no browsers are connected, Claude in Chrome can now show its setup card, behind a switch that is off unless turned on remotely
When an attached computer's Claude Code cannot talk to your session, the notice now says whether it is older or newer than the session
Queued commands and teammate messages are now passed to the model as written, and queued items can be marked as coming from a connector
Claude Code can now abort and clean up a file's first backup if its parent folder changes partway through, behind a remote switch
Cowork's refusal now says an action doesn't run where no one can answer an approval card, and more actions skip the target check
Behind a remote setting, Claude Code can hold back an Esc interrupt until PostToolUse hooks finish, for some tools
Unclear Whether this is switched on by default is not settled.
A crashed background session that runs through an LLM gateway is no longer restarted automatically, even when its saved state reads success
If approving a remote tool call comes back with edited input, Claude Code now refuses the call instead of running the edited version
Claude Code can now resume a session from a resume-head.jsonl side file, falling back to the full transcript, and CLAUDE_CODE_DISABLE_PRECOMPACT_SKIP turns this off
Unclear It is not known whether the mode that enables the faster resume is on by default.
The fleet view's output, needs and detail panels now measure wrapped lines, cut over-long text with a dim marker and drop colour codes unless an option is on
Claude Code can check MCP servers supplied by plugins against claude.ai connectors with permission limits, controlled by two remote switches
Auto mode skips its server check when the conversation is too large, retries a shell command check, and rewords one sub-agent refusal
Unclear What auto mode does with an action when the conversation is too large for the server check is not stated.
With CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST set, Claude Code says auto mode needs a safety check that isn't available there
Claude Code now closes its Chrome connection if it was opened for another account, and will not connect when nobody is signed in
The bundled claude-api skill adds reference docs on API sign-in and on making requests faster, including a small timing script
Claude Code now refuses to send files to a connector unless a person approves the request at Claude Code's own terminal prompt
Unclear It is not stated which tools count as sending files to a connector.
For some sessions, the fleet view's reply box no longer acts on a non-empty draft, unless the draft is the /stop command
Unclear Which sessions this applies to is not clear.
When saving documents to a folder, Claude Code now leaves alone any existing file that has no record of an earlier save, and asks for another out_dir
Unclear It is not clear whether documents that other people may have written are still kept from being saved to files.
When read_db saves documents as JSON files, it can now replace an existing file in some cases and skips others with existing_file
Unclear Which conditions allow an existing file to be replaced rather than skipped is not known.
A synced folder copy of a plugin is no longer loaded when its record allows it to give way and the official marketplace version is installed
Behind tengu_sorted_thimble, SKILL.md or .skill attachments that claude.ai would refuse to save are not sent, with an error saying why
Unclear It is not clear when this check is switched on.
When a remote session is resumed, reading the transcript ahead of time can now also be turned off remotely, and uses a new read path
The three-error disconnect exemption for some HTTP MCP servers no longer depends on a switch and applies whenever the URL check matches
Unclear It is not clear which server addresses pass the URL check, so which MCP servers are affected.
--permission-prompt-tool now uses the default handler when a request carries a client tool-use id and the tool has no input schema
Unclear It is not clear which requests carry a client-supplied tool-use id, so when the default handler takes over.
The rules for when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS sets the beta switch aside have changed, including when ANTHROPIC_UNIX_SOCKET matters
Unclear Which cases now apply, and so whether setting CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS behaves differently for you, is not clear.
When you update a plugin, Claude Code now installs any missing dependencies and adds a note about them to the update message
Disk checks now tell low inodes apart from low disk space, while a command runs and when its large output is saved
Commands wrapped in timeout, and the which command with known flags, can now pass the Bash read-only check without a permission prompt
Unclear It is not clear what condition turns the unwrapping on, so whether it is active by default.
If an agent tries to send a report after one was already delivered, it is now told nothing was sent
Some connector tools for site and app access now refuse a call without a text URL or reason, and explain how to fix it
The tip for raising effort and one item in the shortcut hints now show only when an extra condition is also met
Unclear What the extra condition is, and so who still sees these hints, is not known.
When no browser is connected, switch_browser now returns a specific not-connected error instead of a generic result
When the server briefly turns away writes to an artifact after repeated refusals, Claude Code now tells the model why
When a plugin is renamed, Claude Code no longer rewrites settings sources that are held and records that case
When a cloud machine has not reconnected, Claude is now told to follow what the failed call's error says instead of telling you and pausing
Remote tool calls can now be refused with a message saying the session no longer has a live link to that machine
Unclear It is not stated which feature runs tools on another machine or how to reach it.
After an artifact's database is reported as undeclared, Claude Code refuses further calls to it locally for an hour
When one connected client refuses a request for an SDK MCP server in a remote session, Claude Code waits for another to answer
Auto mode now keeps a required header after server errors unless a check passes, and fails safe when no prompt can be shown
Claude Code's cleanup no longer retires background workers held on their gateway just because their Claude Code build was deleted
How long the URL check prompt waits for an answer now comes from a remote setting held within limits, instead of a fixed time
Picking the Slack channel suggestion now passes your connected MCP servers along before it runs
Unclear What picking the Slack channel suggestion now does differently for a user is not clear.
When an MCP server stands in for the built-in Artifact tools, the model is now told to call them directly and search only for missing ones
The message for a replaced cloud worker and a machine that did not reconnect tells Claude not to do that work itself or report it as done
Connectors that have a web address but no directory entry are no longer found in certain modes
Unclear Which four modes cause these connectors to be hidden.
When a site refuses WebFetch, a subagent may now be told browser tools on your computer are available, and paging notes are reworded
Claude is now told that files from an artifact's database are untrusted and must never make it change permissions, CLAUDE.md or config
A Google Drive file reference now states that it uploads the content of the local file it names to Google Drive
The setting for an SSH host built from a command now documents quoted paths, Windows paths and Claude Desktop v4.16336.0
A remote session can now adopt a colour supplied by the service, and can report whether a turn is being stopped
Lines you select from a diff or your IDE are now flagged when the file may hold data downloaded through a connector
Unclear It is not clear what the isolation safeguard changes once it is switched on.
/rename now warns when the new name might not stick#If /rename cannot save the new name to the session transcript, it now warns the name may not last and suggests running it again
Claude Code now counts images arriving from MCP tasks as a fixed amount each when estimating tokens, not by the size of their text form
Plugin install, remove and restore can now update Claude Code's in-memory plugin list instead of rereading it from disk
When Claude Code refreshes an OAuth sign-in token, it no longer sends a scope value, so the server grants its default access
If claude remote-control cannot read a reply to its setup questions, it now exits and tells you to run it in a terminal
When you edit around pasted text in the prompt, Claude Code can now keep several separate parts of one paste
Unclear What you see differently in the prompt after this change is not known.
/project now accepts https://claude.ai/code/project/<ID> addresses as well as https://claude.ai/project/<ID>
When the last message in a turn is an API error, Claude Code now runs its usual end-of-turn stop handling
Unclear What the stop handling actually does for you when a turn ends on an API error is not stated.
When resuming, Claude Code now limits how long it spends reading ahead in the transcript and refuses unsafe appends to it
When organisation policy blocks an MCP server, programs driving Claude Code now get the error code mcp_policy_blocked before any connection is attempted
Claude Code ignores project and local settings inside a connector-downloads folder and refuses to save settings there
If a tool would prompt but a later check denies it, Claude Code now calls canUseTool with decideLocation "pre-ask" and returns the denial
Unclear It is not stated whether the callback's answer at the "pre-ask" point can change the denial.
If a turn ends because of a new message or a permission check, its PostToolUse hooks stop at a time limit and the tool result is kept without their answer
In a cloud session, asking for permissions now says there is no permissions panel and that rules come from settings files
Unclear Whether the reply is triggered by /permissions or by some other command is not confirmed.
The workspace trust dialog, titled "Accessing workspace:", has been rebuilt and its hint line now follows accessibility styling
/mcp#Long-running MCP tasks now keep checking slowly when their server is removed, explain what happens on exit, and point you to /mcp to reconnect
When there is no spend limit, /usage now shows a "Spend (UTC)" heading with a short subtitle
If Claude Code has entered a worktree while the Bash sandbox is or was on, it no longer loads project plugins from .claude/skills
Claude Code's Chrome connection now checks your signed-in account, closing when the account differs and stopping when nobody is signed in
Results of background MCP tasks are now delivered like background shell results, marked as not coming from you and not counting as approval
Task progress updates now say when a task waits on a usage limit and when it resets, and MCP tool errors now carry an error code
/loop warns that timers in a cloud session may not fire#In a cloud session, /loop now notes that its timer may stop when the session's container is released, and Claude is told to say so
Unclear It is not known what turns on the longer-lasting loop option or which tool it uses.
Workflow task progress now includes a rate-limit wait and keeps sending updates on a timer while the wait lasts
When you stop a turn, waiting updates from tasks moved to the background become passive, then those tasks are stopped
Unclear It is not clear whether this behaviour is switched on.
/context report groups unloaded MCP tools by server#A shorter form of the /context report lists only loaded MCP tools, totals the rest by server and caps long tables
Unclear Which uses of /context produce the shorter form is not stated.
With permissions.denyInsteadOfPrompting, Claude Code ends the turn after too many tool calls in a row are refused
Unclear How many refusals in a row end the turn is not stated.
With the Bash sandbox on, skills in .claude/skills folders below the working folder can stay unloaded until Claude Code restarts
Unclear Whether this is switched on depends on a server-controlled setting whose default is not stated.
WebFetch now refuses claude.ai/share links with a note that the shared chat is shown in the browser and cannot be fetched
Auto mode now turns itself off with an explanation when your permission rules and autoMode settings total more than about 1 MB
Anthropic has not published official notes for v2.1.319 yet. This section updates automatically when the entry appears in the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.
No change to the system prompt since v2.1.318.
Claude Code, interactive mode
One prompt change in this release could not be quoted from the build, so no entry on this page describes it.
413 documentation changes were recorded within 24 hours either side of this release, nearest first. The closest 12 are below. They're here because they happened near this release in time. That's not a claim that this release caused the edit, or that the page documents anything in it.
The 95 literal strings found in the bundle, with the number of entries that name each one. Picking one searches for it. A name is here because this build's code mentions it, which is not the same as it working or being finished.
What's wrong with this entry?