Update log.

[PATCH] Claude Code 2.1.288 · A prompt cleared with Ctrl+C can now be brought back with Up, /code-review gains `--max-findings`, and the agents view adds Ctrl+F to find a session by name. Mid-response API timeouts no longer fail the turn, and fixes focus on resume, auto mode, plugins, permission safeguards, and cloud sessions.

[PATCH] Claude Code 2.1.287 · Claude Mods lets plugins modify deeper behavior, and You should know arrives as a built-in mod where a side agent watches your back and flags things you or Claude might miss. Opus 4.7+ and Fable now use a 1M context window by default on Bedrock, Vertex and Foundry, and MCP `alwaysLoad: false` defers all of a server's tools behind tool search. Fixes focus on screen reader mode, file delivery in Remote Control and cloud sessions, MCP, and permission safeguards.

[NEW] Customize Claude Code with mods · On October 1, 2026, Anthropic introduced mods for Claude Code: plugins made of JavaScript or TypeScript event handlers that can rewrite prompts, guard tool calls, and add panes and buttons to the interface. Mods require Claude Code v2.1.287 or later and work in the CLI and the Code tab of the Desktop app.

[PATCH] Claude Code 2.1.286 · The permission prompt now shows a count such as "2 of 5" when several requests stack up, and the "N more" rows of fullscreen lists respond to the mouse. Failed API requests now share one retry limit per model call (at most 14 requests by default), and `--bare` connects only the MCP servers named on the command line. Fixes focus on `--resume` losing turns, secret redaction in logs and transcripts, Remote Control, subagents and MCP connections.

[PATCH] Claude Code 2.1.285 · Adds `claude --desktop`, `claude plugin configure`, the `allowedProviders` managed setting and the `CLAUDE_CODE_DISABLE_WEB_FETCH` environment variable. Background Bash and PowerShell commands now stop after a time limit (default 30 min), and sessions behind a custom `ANTHROPIC_BASE_URL` use the 1M context window. Fixes focus on the Artifact tool, `/ultrareview` uploads, plugin installs, subagent permission modes and Remote Control.

[PATCH] Claude Code 2.1.284 · Claude Sonnet 5.5 (`claude-sonnet-5-5`) is added and is now the default Sonnet model on the Anthropic API. Interactive terminal and VS Code sessions start in auto mode when no permission mode is configured, on every plan and provider, and Ultracode becomes its own toggle in `/effort`. Fixes focus on damaged response streams, "Prompt is too long" after compacting, MCP connections and plugin permissions.

[NEW] Claude Sonnet 5.5 · Anthropic introduced Claude Sonnet 5.5 on September 28, 2026, the second model in the Claude 5.5 family. It's a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work. It's priced the same as Sonnet 5 at $2 per million input tokens, $10 per million output tokens, and $0.20 per million tokens for cache reads, and the model ID is claude-sonnet-5-5. It scores 70.6% on Terminal-Bench 4.0, compared to Sonnet 5's 10.3%, and two points below Opus 5.5 on GDPval-AA.

[PATCH] Claude Code 2.1.283 · New `availableModelsMatch` and `deniedModels` managed settings let organizations pin allowed models to exact versions and block specific models, and `/doctor prompt-audit` audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. Fixes focus on MCP server reliability, `installed_plugins.json` handling and vim mode, and the 2.1.282 reservation of the `claude-ai` name is reverted.

[PATCH] Claude Code 2.1.282 · A new `maxProseWidth` setting caps the width of Claude's prose in wide terminals, and a startup notice lists telemetry variables in a project's settings files that were ignored or that turned telemetry off. Fixes focus on continued or resumed sessions losing earlier extended thinking, and on every request failing with a 400 error when history holds web search results the API cannot decrypt. Project and local settings now ignore OpenTelemetry export variables, and the `anthropic-skills` and `claude-ai` namespaces are reserved for skills synced from claude.ai.

[PATCH] Claude Code 2.1.281 · `"attribution": false` in `settings.json` hides all commit and PR attribution, MCP servers on 2026-07-28 protocol connections can open a browser-based flow through URL-mode elicitation, and `claude plugin validate` now checks MCP servers. A recursive `rm` whose target is only command-substitution output, such as `rm -rf "$(pwd)"`, no longer runs unprompted in auto mode. Fixes focus on resumed sessions re-sending earlier turns in a changed form, and responses cut short or tool calls run twice behind a proxy or gateway.

[NEW] Claude Marketplace: one place to discover plugins, agents, and services · On September 23, 2026, Anthropic launched Claude Marketplace, bringing plugins and connectors, agents and products, and service partners into one place. Choose from more than 2,000 connectors and plugins, and use a portion of your committed Anthropic spend on Claude-powered software from partners.

[PATCH] Claude Code 2.1.280 · Claude Opus 5.5 (`claude-opus-5-5`) arrives as the default Opus model, and the default model on Pro and Team Standard plans changes from Sonnet to Opus. A stray `y` or `n` no longer confirms or closes dialogs, and `CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH` changes the 2,048-character cap on MCP tool descriptions. Fixes include writes through a symlinked path being judged by their in-tree spelling, auto mode retrying or denying actions over and over, and Ctrl+C or Ctrl+D pressed twice in most dialogs quitting Claude Code.

[NEW] Claude Opus 5.5 · Anthropic introduced Claude Opus 5.5 on September 22, 2026, the first model in the new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. Input and output tokens are $4 and $20 per million, 20% less than Opus 5, and cache reads are $0.20 per million tokens, 60% less. The model ID is claude-opus-5-5, and it generates output more than 30% faster than Opus 5. Five-hour usage limits are increasing on Pro, Max, Team, and seat-based Enterprise plans.

[PATCH] Claude Code 2.1.277 · Claude Code now reads AGENTS.md in a project with no CLAUDE.md, and Claude apps gateways gain `CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1` and an optional upstream `headers:` map. The deprecated TaskOutput tool is removed, and subagent results now reach the main agent under a header marking them as subagent output. Fixes include `claude -p` and Agent SDK sessions hanging with no result after an internal error, conversations failing every request with "text content blocks must be non-empty", and unexpected logouts when an older Claude Code build runs on the same machine.

[PATCH] Claude Code 2.1.275 · A new send-now key (ctrl+enter) interrupts the current turn and sends all queued messages at once, and the skills and plugins enabled on your claude.ai account now sync to terminal sessions signed in with it. `/plugin install <plugin> --marketplace <source>` offers to add the marketplace before installing. Plugins installed from an npm source no longer run the package's install scripts. Fixes include sessions failing to resume or start because of a malformed transcript entry, sandboxed Bash commands on Linux reporting exit code 0 for failed commands under zsh, and plugin messages and logs showing a password or token stored in a git, ssh or marketplace URL.

[PATCH] Claude Code 2.1.274 · A visible warning now appears when memory usage is critical, and `CLAUDE_CODE_MCP_STARTUP_WAIT_MS` bounds how long the first non-interactive turn waits for connecting MCP servers. In VS Code, a step interrupted by a window reload now continues, and the Customize menu gains Memory and Instructions entries. Bedrock, Vertex, Foundry and telemetry-disabled installs now use the v2 MCP client by default, and `"type": "sdk"` MCP entries are skipped with a warning. Fixes include sessions stuck endlessly retrying "unexpected tool_use_id" 400 errors, Streamable HTTP MCP tool calls timing out after about 5 minutes despite a longer per-server `timeout`, an active `/goal` lost when resuming a compacted session, and Bash commands that loop over or assign certain special shell variables not asking for permission.

[NEW] Projects redesigned: from folder to conversation · On September 17, 2026, Anthropic introduced a new experience for Claude projects, now available in beta in Claude Code. You describe what needs to get done and Claude manages the work: a coordinator directs threads, and each thread is a Claude Code cloud session working on its own branch and copy of the repo. Threads add to and draw from a shared memory, and a library collects the files you add and the artifacts Claude produces. It is rolling out first to select Claude Pro and Max subscribers who use cloud sessions, with updated projects across all of Claude and Team and Enterprise plans coming after that.

[PATCH] Claude Code 2.1.273 · Five request headers for LLM gateways are available behind `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1`, a notification now points at `/mcp` when automatic reconnection to an MCP server gives up, and a session started with `claude --remote-control` can be forked from the Claude app. 17 improvements and 6 changes mostly rewrite errors to name the cause: a GitHub IP allow list, SAML single sign-on, an untrusted corporate CA, or a gateway administrator to contact. Long sessions are more responsive because hook progress and sub-agent activity no longer re-process the whole conversation on every update. 35 fixes include Bash commands the permission checker cannot fully analyze skipping the prompt, skills synced from claude.ai staying available after an organization turns Skills off, and advisor-tool turns being counted at roughly twice their real context size.

[PATCH] Claude Code 2.1.271 · Fast mode now works in Claude Code Remote sessions (cloud and self-hosted runners), and per-command `allowed_domains` for Bash, PowerShell and Monitor in auto mode with sandboxing opens only the hosts a command needs. Ten additions in total include `omitClaudeMd` for subagents and mouse support in the fullscreen `/config` panel. 27 improvements and changes cover faster rendering of large diffs and long transcripts, dynamic workflows pausing at your usage limit and continuing when it resets, and Monitor watches always having a deadline. 59 fixes cover a cached organization policy being reused after switching accounts, an unreadable `managed-mcp.json` being ignored, and Bash permission checks missing files behind wildcards or unrecognized options.

[PATCH] Claude Code 2.1.269 · `claude plugin eval` runs a plugin's eval suite against Claude Code and returns scored, reproducible results as a JSON and HTML report, and `/output-style [name]` lists and switches output styles, including over Remote Control and in cloud and other headless sessions. 18 improvements and changes cover the `/diff` panel opening fully rendered in one step, transcript updates in long sessions no longer re-processing the whole conversation, and prompt suggestion filtering for Japanese, Chinese and Korean text. 66 fixes cover the prompt cache being partially invalidated on the turn after a response was cut off at the output-token limit and automatically resumed, the keyboard regression in 2.1.247, a deny or ask rule starting with `!` applying beyond the settings source that wrote it, and `Edit()` deny rules not applying to the file a Bash `tee` command writes.

[PATCH] Claude Code 2.1.268 · With `pricing:` set in the Claude apps gateway's `gateway.yaml`, signed-in Claude Code clients receive the same rates through managed settings, so `/cost` and telemetry match the spend meter, and `claude plugin install`, `uninstall`, `update`, `enable` and `disable` gain `--json`. 26 improvements and changes cover `--continue` / `--resume` showing the conversation immediately, auto mode denials naming the rule that blocked the action, `/plugin` changes taking effect when you close the menu, and the task-tracking tools now being offered only on older models. 59 fixes cover every turn failing with HTTP 400 on third-party Anthropic-compatible endpoints since 2.1.265, WebFetch hanging indefinitely on a server that never finishes its response, deny rules on symlinked directories not applying, and errors showing git tokens and secrets resolved from MCP configs.

[PATCH] Claude Code 2.1.267 · A new `maxEffortLevel` setting, top-level or per model under `modelSettings`, caps the effort level on every provider including Bedrock, Vertex and Foundry, and `--system-prompt-snapshot off` renders the system prompt fresh on every request instead of reusing the conversation's recorded prompt. 9 improvements cover the `/diff` panel, the Bash tool's description guidance, `--resume` first-render time for sessions with many Bash tool calls, and prompt input responsiveness. 41 fixes cover several cases where the tool list was rewritten mid-conversation, breaking prompt-cache reuse and dropping extended thinking, resuming a session whose transcript is over 5 MB dropping parallel tool calls and their hook output, managed allowlists admitting everything when unreadable, and `claude remote-control` exiting and dropping every attached session when its server credential expired.

[PATCH] Claude Code 2.1.265 · `--plugin-dir` can now point at a folder of plugins: each child folder with a manifest loads, and children added or removed while running are picked up. Tool results saved to disk are capped at 1 GB, with the in-conversation preview saying when a saved file was truncated, and the telemetry Claude Desktop and Cowork send through a Claude apps gateway now carries `user.email` and `user.groups`, matching terminal sessions. 12 improvements cover `--worktree` startup on large repositories, `/workflows` agent detail, and slash commands typed mid-prompt. 34 fixes cover resumed subagents changing their system prompt prefix and breaking prompt-cache reuse, resume after the previous process died while a tool was running, non-interactive sessions resetting the shell working directory at each new user message, and `http` MCP servers that only speak the legacy HTTP+SSE transport never connecting.

[PATCH] Claude Code 2.1.261 · `/skill-doctor` shows which loaded skills go unused and what they cost in context, so you can prune them, and new `bashOutputMaxChars` and `taskOutputMaxChars` settings raise how much command and background-task output Claude receives inline before it is saved to a file, up to 128K characters. `/status` and `claude doctor` gained an "Organization policy" line saying why your organization's policy could not be loaded, and the prompt's word-editing keys now match Bash, so `keybindingFlavor` no longer has any effect. 47 fixes cover typed or pasted characters landing out of order during fast input, resuming a session losing hook output around parallel tool calls, Remote Control's stale permission mode and stuck spinners, and sustained high CPU when a background agent could not be resumed.

[PATCH] Claude Code 2.1.260 · A diff panel opens beside the conversation in fullscreen mode and shows your uncommitted changes as Claude edits, toggled with `/diff`, and `/cost` and the status line's `prompt_cache` field now name a likely cause for prompt-cache misses. `/advisor` gained a text form for the desktop app, Remote Control and other headless sessions, and `/reload-plugins` now appears in the Claude Code Desktop and SDK command lists. 38 fixes cover permission rules whose path contains parentheses being dropped and leaving read-only folders writable, one uncompilable rule pattern making every file edit fail, Bedrock and AWS SSO calls failing when the corporate root CA is only in the OS certificate store, and `/rewind` reporting success when nothing was restored. The 2.1.259 change applying `Read()` deny rules to Bash arguments has been reverted.

[PATCH] Claude Code 2.1.259 · A new `managedMcpServers` managed setting lets organizations provide HTTP/SSE MCP servers to every user, and `--permission-prompts none` lets unattended headless hosts deny anything that would prompt while the active permission mode keeps deciding. `glab mr` commands are now recognized, so GitLab merge requests show as `MR !N`, and `claude plugin validate` gained `--json` for a machine-readable report. 26 fixes cover concurrent sessions reverting each other's `~/.claude.json` changes, managed settings going unenforced when they cannot be parsed, worktree-isolated sessions refusing common Bash loops and xargs pipelines, and frontmatter `model:` being ignored in interactive sessions.

[NEW] Building commerce agents with Claude · Anthropic released a blueprint for building commerce agents on Claude on September 2, 2026. It ships two agents defined once and deployable three ways — a shopping agent for customers and a merchant agent for staff, running on the Messages API, the Claude Agent SDK, or Claude Managed Agents — along with four runnable vertical examples covering retail, travel, telecom, and entertainment. A Claude Code plugin called commerce-builder scaffolds a custom agent for your own store. The code is available under Apache License 2.0 at anthropics/commerce-agents, and every example runs on fictional data with no live orders or charges.

[PATCH] Claude Code 2.1.257 · Claude Fable 5.1 (`claude-fable-5-1`) is now the default Fable model — 1M context, $10/$50 per Mtok with $0.25/Mtok cache reads. New settings arrive for the turn-end clock and transcript timestamps (`timeFormat`, `timeZone`), for forcing one model across every subagent (`CLAUDE_CODE_SUBAGENT_MODEL_FORCE`), and a Containment Escape rule that stops auto mode from auto-approving cloud metadata-credential fetches, egress evasion, and cross-tenant reach. Rendering performance and prompt input responsiveness improved, `defaultMode: "bypassPermissions"` in project settings is now ignored, and `/btw` history browsing moved to `Shift+←`/`Shift+→`. Roughly 60 bug fixes span background sessions, subagents, prompt caching, and gateway authentication.

[NEW] Claude Fable 5.1 and Claude Mythos 5.1 · Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, describing them as the world's most advanced models for coding and knowledge work. Input and output pricing is unchanged from Claude Fable 5 at $10 / $50 per MTok, but cache reads drop to $0.25 per MTok, making Fable 5.1 an estimated 25% less expensive than Fable 5 for typical workloads. The model ID is claude-fable-5-1, with a 1M token context window and 128k max output tokens. Three changes are breaking for existing Fable 5 callers, starting with forced tool use. Mythos 5.1 offers the same capabilities to Project Glasswing participants only.

[PATCH] Claude Code 2.1.252 · A bug-fix-only release. Bash commands no longer fail with `task output swap refused (tasks dir moved or linked)` on some Macs, and `always allow` now saves in a project that has no `.claude/settings.local.json` yet. Remote Control sessions hosted by Claude Desktop or VS Code no longer stall for minutes after a tool finished when the connection to claude.ai is degraded, and background task notifications with very large failure output (for example git errors on a full disk) no longer make the conversation exceed the API request size limit.

[PATCH] Claude Code 2.1.251 · New `PreModelSwitch` and `PostModelSwitch` hook events let you block, confirm, or annotate a model switch, and `SessionStart` resume hooks now receive session staleness and the estimated re-cache cost. Remote Control clients get live streaming of a foreground subagent's tool calls and results, `/cost` gains a per-session prompt-cache line (hit ratio, misses, tokens re-cached, warm/cold) with a matching `prompt_cache` object for status line scripts, and `/usage` gains a Spend limit bar. A run of permission fixes closes paths around the permission check: file tools following a symlink swapped after the check, plugin commands pointing outside the plugin directory, and Grep and Glob skipping `Read(...)` deny rules through a symlinked search path. Project `.claude/settings.json` `env` can no longer set `CLAUDE_CONFIG_DIR` or `TMPDIR`, and `CLAUDE_CODE_SUBAGENT_MODEL` now sets the default subagent model rather than overriding everything.

[PATCH] Claude Code 2.1.248 · The new `--restricted` flag (or `CLAUDE_CODE_RESTRICTED=1`) removes the built-in tools that run commands or code and `WebFetch`, keeps file tools inside the working directory, refuses `bypassPermissions`, and ignores user, project and local settings files. A prompt-cache miss that hit long sessions roughly once an hour — tool definitions being re-rendered after an OAuth token refresh, which also lost extended-thinking context — is fixed, and agent frontmatter gains `experimental.cacheTtl` for a per-agent prompt cache TTL. Claude Desktop and Cowork sessions no longer disappear after 30 days, `claude agents` stops resurrecting weeks-old background sessions or starting a second process on a conversation already open elsewhere, and `/ultrareview` and locally seeded cloud sessions no longer upload uncommitted `prod.env`-style files, `*.tfvars`, or editor swap, temp and backup copies of credential files.

[PATCH] Claude Code 2.1.247 · The new `SendFeedback` tool lets Claude draft a feedback report when something goes wrong in a session, for you to review and send from `/feedback` (turn it off with the `feedbackDrafts` setting). Bash permission prompts now carry a tip pointing to auto mode with a one-keystroke "Yes, and switch to auto mode" option, and `/claude-api cost-optimize` profiles an existing project's Claude API spend and works through cost levers (caching, token hygiene, batch, effort, model choice) one measured change at a time. Sonnet 5's default auto-compact window moves to its full 1M context, so 1M sessions now auto-compact at about 967K tokens instead of about 934K, and a hook or background agent that printed megabytes of error output can no longer wedge the session on "Prompt is too long".

[PATCH] Claude Code 2.1.246 · `/permissions` gains an Auto mode tab for viewing and editing auto mode classifier rules, and Bash allow rules with a wildcard before the subcommand (e.g. `Bash(git * main)`) now raise a startup warning, since they also match options inserted before the subcommand. After `/cd`, the new directory's project settings, hooks, `.mcp.json` servers, skills, and agents take effect right away instead of on `--resume`, and non-interactive sessions automatically continue a response cut off mid-stream. Fixes land for background sessions failing to open after 45 seconds when the starting directory had been deleted or the machine had slept, severe transcript slowdown when a diff contained a very long single line, and telemetry requests to Anthropic carrying the API key configured for a third-party gateway.

[PATCH] Claude Code 2.1.243 · A Loops breakdown lands in `/usage` with per-loop run count, total tokens, tokens per run, and last run, so runaway `/loop` tasks are easy to spot, and a new `modelPicker` setting curates the `/model` picker with an ordered, labeled list of models. `promptCacheTtl` and `subagentPromptCacheTtl` let API-key and cloud-provider users keep a 1-hour prompt cache on the main conversation while subagents stay at 5 minutes, and `/login` gains a keyless Anthropic Console sign-in for organizations that don't allow API keys. Startup time and memory usage improve, the native binary is now zstd-compressed (about 75 MB instead of 340 MB on Linux x64), and fixes land for sessions going silent when the API never starts a response and for remote MCP servers never recovering in non-interactive and SDK sessions.

[PATCH] Claude Code 2.1.239 · The headline fix is Bedrock streaming behind proxies that strip the response Content-Type header, which silently doubled billed API calls by re-running every turn non-streaming. Cost estimates (`/cost`, status line, `--max-budget-usd`) now include the 1.1× US-only-inference premium for data-residency workspaces, and a new `/claude-api upgrade` migrates Python projects from `anthropic` 0.x to 1.x. Windows gains cross-session messaging, so sessions across your machines can reach each other with `SendMessage` and `ListAgents` as on macOS and Linux, and plugins synced from claude.ai now show as `name@synced` in cloud sessions. More than 50 changes land in total, clustered around `/resume` session pickup and titles, fullscreen rendering, and prompt-input editing.

[PATCH] Claude Code 2.1.238 · A new `keybindingFlavor` setting makes Ctrl+W in the prompt delete back to the previous whitespace, as in Bash, when set to `"readline"`. Plugin marketplaces gain `headersHelper`, a command that mints HTTP headers such as a short-lived token for catalog and same-origin archive fetches, with the command shown and confirmed before it runs on install or update. Self-hosted runners gain `--defer-shutdown-max-min` for graceful SIGTERM handling and `--proxy-authorization-command` / `--proxy-authorization-file` for egress proxies that require a freshly issued `Proxy-Authorization` header on every connection. More than 30 fixes and improvements land alongside, including unbounded memory growth in long interactive sessions, output styles drifting back to the default voice mid-session, and a large cluster of Remote Control connection and messaging fixes.

[NEW] Computer use, the Skills API, and the Files API are now generally available — plus a new browser use tool · Computer use, the Skills API, and the Files API are now generally available on the Claude Platform, alongside a new browser use tool for web applications. Computer use now executes multiple actions per turn instead of one per model call and is eligible for HIPAA-regulated workloads under Anthropic's BAA. The Skills API gets a simplified upload and versioning interface, and the Files API adds automatic expiration, 5x higher rate limits, and 1 TB of storage per organization.

[PATCH] Claude Code 2.1.236 · A new `ANTHROPIC_DEFAULT_MODEL` environment variable sets the model new sessions start on, while a /model pick still overrides it and persists across restarts. Cross-session `SendMessage` gains `notify_when_idle`, an opt-in one-shot notice when another Claude Code session on this machine next goes idle (macOS and Linux). On macOS, wildcard read-deny sandbox rules such as `**/.env` now take precedence inside allowed read regions and can't be bypassed by renaming. Auto mode reviews Monitor commands the same way as Bash and uses the Claude API defaults on Bedrock, Vertex AI and Foundry, alongside close to 30 fixes including sessions breaking after their working directory was removed.

[PATCH] Claude Code 2.1.235 · An optional `spellcheck` setting underlines misspelled words in the prompt input as you type, using your installed aspell, hunspell or ispell. The rest is refinement: memory and CPU usage improve while cloud sessions such as /ultrareview or /autofix-pr run in the background, permission dialogs now always match what a grant would cover, and the embedded grep in native macOS/Linux builds fails fast on pathological patterns instead of exhausting memory. Fixes land across nested markdown list alignment, shifted prompt input highlights, Shift+Tab in the permission prompt's comment field, and notebook cell approval dialogs.

[PATCH] Claude Code 2.1.234 · Claude Code now continues your session automatically when a claude.ai usage limit resets, and you can turn it off in /config. Repos with a GitLab remote and an authenticated glab CLI get an MR !N badge in the footer and statusline, a new selection:clear keybinding action clears an in-app text selection, and CLAUDE_CODE_PROJECT_DIR_NAME lets hosts pick a short name for the per-project transcript directory. Loading the built-in claude-api skill drops from ~200k+ tokens to ~25k, and /permissions and /add-dir can now be opened while Claude is working. Over 30 fixes land across permission previews and credential masking, Remote Control sync, and queued ! shell commands.

[PATCH] Claude Code 2.1.233 · The two Bash permission changes from yesterday's 2.1.232 — permission checks on input redirections (`< file`) and on Cygwin-style symlinks on Windows — are reverted, and the auto mode regression that made ordinary `cd <dir> && <command> > file` commands stop for manual approval on Windows is fixed. GitLab support now extends to merge request URLs in the `--worktree` flag and the `claude agents` view, where MRs display as `!N`. Three opt-in settings land: `CLAUDE_CODE_TOOL_MEMORY_LIMIT` for memory cgroup limits on Bash tool commands on Linux, a `forward_user_identity` apps gateway setting that sends the signed-in user's identity as headers, and `CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS` for the WebFetch session URL cache TTL. Todo/task-tracking tools are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models.

[PATCH] Claude Code 2.1.232 · Subagent forking is now enabled by default: a `subagent_type: "fork"` subagent inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions run in the background by default. Type `@` in the prompt to mention another Claude session by name, and interactive sessions on one machine now keep unique names. Plugin marketplaces accept bare `gitlab.com` repo URLs the way they accept `github.com` URLs, and GitLab token families are now redacted. The fixes cluster around permission bypasses and Remote Control: a PowerShell bypass through `$PSDefaultParameterValues`, a Windows bypass through Cygwin-style symlinks followed by Git Bash, and nested git repositories inheriting trust from a parent directory — each repository now requires its own trust confirmation.

[PATCH] Claude Code 2.1.229 · Plugin marketplaces can now come from a `command` source: a local command prints the plugin directory, which is re-resolved each session and applied without a restart. Gateway streaming responses send SSE keepalive pings during long thinking pauses, preventing idle-timeout disconnects on Vertex and Bedrock upstreams, and self-hosted runner sessions now get server-supplied hooks, matching managed-environment behavior. Several crashes are fixed — a RangeError in very narrow terminal windows, a crash on non-string `glob`, `file_path`, or `command` values, and a Windows crash on extended-length or UNC paths — along with long responses partly disappearing while streaming and being printed twice. `/commit-push-pr` no longer auto-approves git/gh commands carrying dangerous flags like `--force`, `--amend`, or `--no-verify`.

[PATCH] Claude Code 2.1.228 · Skills synced from claude.ai are hardened: 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. Two fixes stop things from being deleted that shouldn't be — session cleanup deleting contents inside a project's memory folder, and background plugin-cache cleanup deleting a plugin's cache when its only version is a symlinked development checkout. Also fixed: interactive sessions that could stop redrawing entirely while the process kept running, git and Git Bash not being found on Windows, and self-hosted runner sessions ending early.

[PATCH] Claude Code 2.1.227 · Feature flags are no longer evaluated without the user's subscription tier when a session starts with an expired login token, which could wrongly prompt Max plan users to enable usage credits for Fable. Two more fixes land alongside it: every Bash command failing under claude-code-action with allowed_non_write_users on GitHub-hosted runners, and /tui bringing back a conversation that had been rewound to before its first message. The slash-command menu now uses blue only for the selected row, bolds matched characters instead of recoloring them, and keeps the glyphs in emoji or accented names, and there are fewer event-loop stalls on file-not-found suggestions and at-mention size checks.

[PATCH] Claude Code 2.1.225 · Claude Code's usage warning now supports gateway spend limits, so the limit-reached message names the cap, its reset time, and the operator's message (it requires the gateway on 2.1.225). claude agents gained a workspace trust prompt for untrusted directories, matching the behavior of claude, and SendMessage can start a conversation with your Remote Control sessions on other machines by name, with ListAgents showing them as name [ref]. On the fix side, a transient 401 no longer replaces a long-lived CLAUDE_CODE_OAUTH_TOKEN with a stored login's short-lived token, MCP OAuth servers on macOS no longer fail with a burst of 401 errors after a keychain read times out, and conversation history no longer breaks on Remote Control session resume after very large conversations were compacted.

[PATCH] Claude Code 2.1.224 · Self-hosted environments arrive: claude self-hosted-runner turns your own machines or containers into a place Claude Code web, mobile, and desktop sessions can run, on Team and Enterprise plans. Cross-session SendMessage lets Claude Code sessions message each other across your machines, with ListAgents to discover them (macOS and Linux), and a new archive plugin source installs plugins from a zip over HTTPS without git or npm, with optional SHA-256 pinning. Sandbox credential masking gains extract, onExtractNoMatch for structured env values, JWT-aware masking, and AWS SigV4 re-signing. Long project paths no longer resolve to another project's session directory, sandbox deny entries written with a trailing slash are no longer bypassable, sandbox violation details now reach Bash tool results, and the 200-subagent-per-session spawn cap is gone.

[NEW] Auto mode becomes the default in Claude Code for Pro, Max, and Team plans · Starting on August 14, new sessions on Pro, Max, and Team plans will run in auto mode. Instead of permission prompts, auto mode routes each tool call through a classifier targeted at blocking actions that are irreversible, destructive, or aimed outside your environment. Users can switch modes with Shift+Tab in the CLI or the mode dropdown on the desktop app, and admins can pin an org-wide default with defaultMode or turn auto mode off entirely with disableAutoMode.

[PATCH] Claude Code 2.1.223 · A Bash permission bypass where a crafted command could hide parts of itself from permission checks is fixed, along with permission prompts where tabs or invisible Unicode could hide part of the command from the approval dialog. Workflow scripts can no longer use dynamic import() to run code outside the workflow sandbox, and an agent definition's bypassPermissions mode no longer ignores the org bypass-permissions disable policy. The strictKnownMarketplaces and blockedMarketplaces managed settings now accept owner wildcard entries ("owner/*"), and cloud sessions show a /teleport hint for continuing locally with claude --teleport <session id>. CLAUDE_CODE_DISABLE_1M_CONTEXT now holds every Claude model with a native 1M window to 200K, and /review is an alias of /code-review.

[NEW] Run Claude Code sessions on your own compute with self-hosted environments · Now in public beta, self-hosted environments let you run Claude Code sessions on your own infrastructure. You operate runners in either fixed or on-demand mode, and each session runs in its own checkout so work stays isolated between developers and accounts. Repository checkouts, build artifacts, and secrets stay on infrastructure you provision, while the conversation is sent to Anthropic for inference. Available to organizations on Claude Team and Enterprise plans.

[NEW] Bringing MCP 2026-07-28 to Claude — a stateless core, standardized extensions, and hardened auth · MCP 2026-07-28, the fifth spec release of the Model Context Protocol, is live. It moves MCP from a bidirectional stateful protocol to a stateless request/response core so servers can deploy on serverless and edge infrastructure, graduates MCP Apps and Tasks into a versioned extensions framework, and aligns authorization with production OAuth 2.0 and OIDC deployments. Support is rolling out across Claude products.

[NEW] Claude Opus 5 · Anthropic introduced Claude Opus 5 on July 24, 2026, for complex agentic coding and enterprise work. It is priced at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8, with the model ID claude-opus-5, a 1M token context window, and 128k max output tokens. Adaptive thinking is now on by default, effort runs up to max, and two betas ship alongside it: mid-conversation tool changes and automatic fallbacks. It is the new default model on Claude Max and the strongest model on Claude Pro.

[NEW] Ask Claude about the Anthropic Economic Index · A new connector lets you query the Anthropic Economic Index in claude.ai in plain language. Ask questions like "Which occupations use AI the most?" to explore AI adoption, task automation, and geographic usage patterns. There's nothing to install, but you do connect it once in the claude.ai connectors settings.

[NEW] The iOS Simulator pane comes to Claude Code Desktop (public beta) · Claude Code Desktop opens your app in an iOS Simulator pane when Claude builds, runs, or checks it, streaming the device screen live next to your conversation. Watch Claude install and tap through the app to verify its own changes, or tap through it yourself. It's in public beta on macOS, available on Pro, Max, and Team plans.

[NEW] Introducing Claude for Teachers · Claude for Teachers gives verified K-12 educators in the US free access to Claude, with teaching skills for lesson planning and differentiation, a Learning Commons of academic standards across all 50 states, and connectors to nine educational platforms. Sign up by June 30, 2027, for a full year of access.

[NEW] Claude Cowork is coming to mobile and web · Claude Cowork is rolling out to mobile and web, so your sessions and files go where you go, on any device. Work continues in the background, and when Claude reaches a call only you can make, the question reaches your phone. Beta access rolls out over the next several weeks, starting with Max users.

[NEW] Introducing Claude Science · Anthropic launched Claude Science in beta on June 30, 2026, an AI workbench that integrates fragmented scientific research tools into a unified environment. Researchers can run complete workflows — from literature analysis through manuscript preparation — in one platform, with specialist agents that query and synthesize across more than 60 curated scientific databases including UniProt, PDB, Ensembl, and ChEMBL. It's available to Claude Pro, Max, Team, and Enterprise users on macOS and Linux.

[NEW] Introducing Claude Sonnet 5 · Anthropic released Claude Sonnet 5 on June 30, 2026, an agentic model that approaches Opus 4.8 performance at lower prices. It can make plans, use tools like browsers and terminals, and run autonomously, with benchmark gains over Sonnet 4.6 across coding, reasoning, tool use, and knowledge work. The model ID is claude-sonnet-5; it is the default model for Free and Pro plans and is available in Claude Code, the Claude Platform, and the Claude API.

[NEW] Claude in Microsoft Foundry is now generally available · Claude models are now generally available in Microsoft Foundry, hosted on Azure. Claude runs in your Azure environment with the authentication, billing, and governance controls your teams already use, and you can choose where inference is processed, including a US data zone for teams with data residency requirements. Claude Opus 4.8 and Claude Haiku 4.5 are available, with support for the Messages API, prompt caching, and extended thinking.

[NEW] Introducing the Claude apps gateway — centrally manage Claude Code on Amazon Bedrock and Google Cloud · The Claude apps gateway is a self-hosted control plane for managing Claude Code deployments on Amazon Bedrock and Google Cloud. It provides corporate SSO sign-in, centrally managed settings enforced on every request, and per-user cost attribution, and ships as a stateless container within the existing claude binary. It supports OIDC-based authentication, control over allowed models, and spend caps per organization, group, or user.

[NEW] Introducing Claude Tag — work with @Claude in your Slack channels · Claude Tag is a new way for teams to work with Claude. Tag @Claude in a Slack channel to delegate a task, and it completes work autonomously while keeping context across the channel's conversations. Everyone in a channel works with the same Claude, it learns from channel history and connected data, and you can hand off work asynchronously. Built on Opus 4.8 and available in beta for Claude Enterprise and Team customers.

[NEW] Claude Code now supports artifacts — turn session work into live web pages · Claude Code can now turn session work into artifacts — live, shareable web pages built from full session context. Ask for a PR walkthrough, dashboard, or incident timeline, and the open page refreshes in place when Claude Code updates it. Artifacts are private by default and viewable only by authenticated members of your org, available in beta for Claude Team and Enterprise orgs.

[NEW] Claude Design now stays on brand for daily work · Claude Design now imports your design system — from GitHub repos, design files, or raw uploads — so Claude builds with your components and validates against your system. New /design-sync and /design commands connect it to Claude Code, the editor gains rich layout controls and hundreds of stability fixes, usage limits are now shared with chat, Claude Cowork, and Claude Code, and connectors expand to Adobe, Base44, Canva, Gamma, Lovable, Miro, Replit, Vercel, and Wix.

[NEW] Workload Identity Federation is now generally available on the Claude Platform · Workload Identity Federation is generally available on the Claude Platform, letting developers authenticate with short-lived, scoped credentials from OIDC-compliant identity providers instead of static API keys. Federation rules bind external identities — AWS IAM, GCP, Azure, Kubernetes, GitHub Actions, Okta — to service accounts that issue tokens at request time across all Claude API endpoints.

[NEW] Claude Fable 5 and Claude Mythos 5 · Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Fable 5 is a Mythos-class model made safe for general use — state-of-the-art on nearly all tested benchmarks, with the `claude-fable-5` model ID. Mythos 5 is identical with cybersecurity safeguards lifted, available in limited release through Project Glasswing. Both support a 1M token context window and up to 128k output tokens by default, priced at $10/$50 per MTok. When a safety classifier declines a request, the Messages API returns stop_reason: "refusal" as an HTTP 200, and the fallbacks parameter (beta) or SDK middleware retries on another Claude model.

[NEW] What's new in Claude Managed Agents — scheduled deployments and environment variables in vaults · Two capabilities land in public beta on the Claude Platform. Scheduled deployments give an agent a cron schedule — each time the schedule fires, the agent starts a new session and completes its task, with no scheduler for you to build or host. Environment variables in vaults store API keys tied to specific approved domains: the agent never sees your key because the sandbox only holds a placeholder, and the real key is attached at the network boundary, only on requests to domains you allow. Supports the Browserbase, KERNEL, Notion, Ramp, and Sentry CLIs plus any CLI using HTTP authentication.

[NEW] Claude in Apple's Foundation Models framework · Anthropic released a Swift package that lets Apple developers use Claude through Apple's Foundation Models framework. The framework gives developers access to tap into models natively from Swift and can return typed Swift values through guided generation in as few as three lines of code. Apps can use on-device models for simple tasks, then hand off to Claude when workflows require multi-step reasoning or code generation, with streaming, tool calls, and structured responses flowing back into your SwiftUI view. Supports iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27; an Anthropic API key is required.

[NEW] Claude Opus 4.8 · Claude Opus 4.8 builds on Claude Opus 4.7 with improvements across benchmarks and is a more effective collaborator, available today for the same price. It targets long-horizon agentic coding with better long-context handling and compaction recovery, more reliable effort calibration, and better tool triggering. The model ID is claude-opus-4-8; it supports the 1M token context window by default on the Claude API, Amazon Bedrock, and Vertex AI, with 128k max output tokens.

[NEW] Dynamic workflows in Claude Code · Claude breaks a problem into subtasks and fans them out across tens to hundreds of subagents running in parallel, dynamically writing orchestration scripts that run in a single session and checking its work before anything reaches you. It targets broad work like codebase-wide bug hunts and security audits and large migrations spanning thousands of files, verifying each result before folding it into the final answer. Available as a research preview on Max, Team, and Enterprise plans.

[NEW] Claude Managed Agents — Self-hosted sandboxes and MCP tunnels · Claude Managed Agents now supports self-hosted sandboxes (public beta) and MCP tunnels (research preview). The agent orchestration loop stays on Anthropic infrastructure while tool execution runs in customer infrastructure or with managed sandbox providers — Cloudflare, Daytona, Modal, Vercel — and agents reach private-network MCP servers through a lightweight gateway with a single outbound connection, no inbound firewall rules required.

[NEW] Agent view in Claude Code · Agent view is a centralized interface inside the Claude Code CLI for managing multiple concurrent agent sessions. Now in Research Preview, it lets developers see which agents need input, are actively working, or have completed tasks — and scale up background work via peek, `/bg`, and `claude --bg`.

[NEW] Claude Platform on AWS is now generally available · The Claude Platform on AWS — Anthropic-operated, accessed via AWS IAM with CloudTrail audit logging and a single AWS invoice — is generally available. AWS customers get day-one access to all native Claude API features, including Claude Managed Agents (beta), Advisor strategy (beta), Files API (beta), Skills (beta), and the MCP connector (beta), with Claude Opus 4.7, Sonnet 4.6, and Haiku 4.5 available at launch.

[NEW] What's new in Claude Managed Agents — Dreaming, Outcomes, Multiagent, Webhooks · A month after the April 8 beta launch, Managed Agents added two new features and promoted three from research preview to public beta. Dreaming (research preview) curates memory between sessions for self-improvement, and outcomes, multiagent orchestration, memory, and webhooks all ship as public beta.

[NEW] Claude Financial Services Agents — 10 templates, Moody's MCP, and 17 data connectors · Ten ready-to-run financial services agents — pitch builder, KYC screener, month-end closer, and more — ship as plugins in Claude Cowork and Claude Code, and as cookbooks for Claude Managed Agents. A Moody's MCP app brings proprietary credit ratings and data on 600M+ companies, and 17 data partners are integrated alongside.

[NEW] Claude for Microsoft 365 — Excel, PowerPoint, and Word now GA, Outlook joins in beta · The Claude add-ins for Excel, PowerPoint, and Word are generally available on all paid plans, and Claude for Outlook joins in public beta. Once installed, context carries automatically between apps so a model built in Excel can become a PowerPoint deck or a Word memo without re-explaining inputs, assumptions, or calculation flow.

[NEW] Claude Security Enters Public Beta · Anthropic launched Claude Security in public beta — a Claude Opus 4.7-powered tool that scans codebases for vulnerabilities and generates targeted patches. Available directly at claude.ai/security and embedded in CrowdStrike, Microsoft Security, Wiz, and other platforms. Currently for Claude Enterprise customers; Claude Team and Max coming soon.

[NEW] Claude Platform on AWS Announced (Coming Soon) · AWS announced Claude Platform on AWS — a new way to use Anthropic's native Claude Platform with AWS credentials, billing, and access controls. It sits alongside Claude on Amazon Bedrock as a second path with different data-residency tradeoffs.

[NEW] Running Claude Cowork in Amazon Bedrock · AWS published the official guide to running Claude Cowork on Amazon Bedrock — reuse the same Bedrock infrastructure you built for Claude Code to bring Claude to every knowledge worker, inside your AWS perimeter.

[NEW] Claude for Word · A Microsoft Word add-in that brings Claude into your document — with native tracked changes, template-aware drafting, and consistency checks.

[NEW] Claude Design by Anthropic Labs · Collaborative visual creation with Claude — designs, interactive prototypes, presentations, and more through conversation.

[NEW] Claude Opus 4.7 · A new Opus model with significant gains in coding, high-resolution vision, and instruction following. Claude Code adds /ultrareview and the xhigh effort level.

[NEW] Routines — automated Claude Code workflows · Autonomous routines in Claude Code triggered by schedules, API calls, and GitHub events. Run code review, bug fixes, and deploy verification from the cloud — laptop closed.

[NEW] Claude Managed Agents · A new API alongside Messages that provides a pre-built agent harness and managed infrastructure — run Claude agents without building your own loop, sandbox, or tool-execution layer.

[NEW] Claude Code Remote Control · Keep a local Claude Code session running and control it from claude.ai/code or the mobile apps — with your local files, MCP servers, and tools intact.