Released: 2026-08-20
If you are used to pressing Ctrl+W in Bash to delete the previous word, you can now add {"keybindingFlavor": "readline"} to your user settings file at ~/.claude/settings.json to switch Ctrl+W to delete back to the previous whitespace, matching Bash; the default "classic" flavor remains unchanged if you don't set it.
In fullscreen mode, pressing Ctrl+L or Cmd+K used to both repaint the display and, if double-pressed, trigger /clear, wiping out the whole conversation by accident. These shortcuts now focus exclusively on repainting the display; the double-press /clear shortcut has been removed, and 1-row nvim terminal windows no longer trigger unexpected clear loops.
Plugin marketplaces now support dynamic authentication.
By configuring headersHelper on a URL marketplace or catalog entry, the system runs a custom command to mint dynamic HTTP headers, using a short-lived token to fetch catalogs and same-origin archive files. Catalog helpers only execute when installing or updating that plugin after displaying the command, and claude plugin install or update will prompt with [y/N] unless -y is passed.
On SIGTERM, claude self-hosted-runner --defer-shutdown-max-min <minutes> lets a self-hosted runner keep serving attached sessions, then park remaining work before exiting. A single slow or lost poll request also no longer removes a healthy runner and hands its sessions to another runner.
For egress proxies requiring a freshly issued Proxy-Authorization header for every connection, self-hosted runners can use claude self-hosted-runner --proxy-authorization-command or claude self-hosted-runner --proxy-authorization-file.
Long sessions no longer keep accumulating memory from subagent tool results. Results remain in memory while in the recent display window, then are released once they leave it, preventing unbounded memory growth.
Custom, project, and plugin output styles no longer drift back to the default voice mid-session.
Permission dialogs and diff previews now format text reliably across all terminal sizes. When MCP (Model Context Protocol, an open standard connecting AI to external tools) opens an elicitation prompt, URLs exceeding 4,096 characters no longer render completely blank. If a project path is wider than your terminal window, the permission prompt will no longer drop the "don't ask again" choice. Furthermore, line wrapping in permission diff views has been fixed so that lines with tabs or wide multi-code-point characters like emoji are never clipped.
Input responsiveness and terminal state cleanup are now reliable over laggy remote connections.
Previously over high-latency SSH or mosh connections where terminals map Backspace to Ctrl+H, holding down Backspace during keystroke bursts would cause inputs to be dropped; these bursts are now handled accurately. Additionally, killing a session suspended with Ctrl+Z will no longer leave your terminal with a hidden cursor or stuck in bracketed-paste mode.
Launching the tool is now noticeably snappier from the command line. On macOS, running bare claude opens the interface significantly faster so you can get straight to work. To keep the app responsive from the very first keystroke, the automatic update check now waits approximately 10 seconds after launch instead of competing with startup routines for CPU cycles.
Security boundaries around external tool configurations have tightened significantly.
If a project .mcp.json contains headersHelper, or if inline MCP (Model Context Protocol, an open standard connecting AI to external tools) servers exist in project or --add-dir agent files, that directory's trust dialog must now be accepted first, even when running non-interactively via claude -p. In addition, helpers launched from project files, plugins, or agent configurations no longer inherit credential environment variables, and user, managed, or claude.ai-scoped helpers now run strictly from the Claude config directory.
Inspecting external tools no longer wakes up dormant or disabled services. Previously, stdio MCP (Model Context Protocol, an open standard connecting AI to external tools) servers received a server/discover request before initialize, unnecessarily forcing lazy-loaded servers to boot up their backends every time a session opened. This handshake order is now resolved. In addition, claude mcp list and claude mcp get now display disabled servers directly as ⊘ Disabled rather than opening a connection to perform a health check.
Bash tool permission checks now evaluate zsh-specific conditional syntax correctly. In isolated git worktrees, commands without a redirect stop receiving false refusals telling you to remove one.
When a Bash command is killed, times out, or is interrupted, it no longer leaves /tmp/claude-*-cwd files behind. These work-directory temporary files are cleaned up even when the command does not finish normally.
Remote Control connectivity is now far more resilient across complex network environments. When a proxy refuses a connection, it now names that proxy rather than returning a generic network error. Separately, when a network edge, VPN, or proxy briefly returns HTTP 403, it tolerates the refusal and retries for up to 3 minutes; if the block persists, it names the refusing party. When a brief network hiccup delays your sign-in credential refresh, the client retries seamlessly instead of abruptly dropping out with "login expired". Signing out also concludes the session cleanly without misreporting a reconnection failure.
A model chosen from a phone or web client now updates the terminal display. A new message sent from the web or Desktop while Claude is mid-turn also remains in the transcript after the turn ends.
When you want to stop one named task in the Remote Control panel, its individual Stop button now works for a CLI (command-line interface, a way to operate programs with typed commands) hosted session too. Previously, pressing it left that type of session running; the panel's task control now reaches it.
A remote session no longer exits when it receives a user message from a client without a valid role. A role is the field that identifies a message's identity and purpose; before, an invalid value could end the whole session, while it now remains usable. Later messages can still go to the same remote session instead of requiring a new one because of a single bad field.
New remote sessions started by claude remote-control no longer inherit session-scoped environment variables from the launching shell. They are now isolated instead of being carried into the remote environment.
After a Remote Control session process crashes, it used to stay unavailable until claude remote-control was restarted. Now, the next message can reinitialize and reuse it.
Cross-session messaging now provides direct feedback when messages cannot be delivered.
When sending messages to a local session that rejects incoming chats via {"crossSessionInbound": "refuse"}, the sender now receives a clear "refused" status instead of a deceptive success report. Likewise, if a target inbox drops messages due to rate limits or a full queue, your active session is notified immediately instead of messages vanishing silently.
Multi-agent collaboration tools can now discover and communicate with peers across environments. Running ListAgents or SendMessage inside sessions hosted by claude remote-control server mode, Desktop, or IDE hosts previously failed with "Remote Control is not connected"; they can now locate and message remote peers reliably. Additionally, idle workers pre-warmed in the background to speed up your next session are no longer prematurely listed, appearing only after a task officially claims them.
Prompt suggestions and cache status warnings now reflect exact account and cache states. If you set CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=true, suggestions previously shut off prematurely when your account approached its usage limit; they now remain active until the quota is genuinely reached. Meanwhile, switching models with /model or tweaking reasoning depth with /effort will no longer display false cache-miss warnings when prompt caches had simply expired on their own schedule.
The bundled claude-api skill has been updated to align with the Managed Agents release from August 19. On self-hosted sandboxes, agents can now configure allowed domains for web searching and web fetching, giving you granular control over external site access. Sandboxes also introduce memory stores, allowing managed agents to persist and retrieve long-term state securely within their private sandboxed environments.
Added a `keybindingFlavor` setting: set it to `"readline"` to make Ctrl+W in the prompt delete back to the previous whitespace, as in Bash; the default (`"classic"`) is unchangedChanged Ctrl+L and Cmd+K in fullscreen to always just repaint — the double-press `/clear` shortcut was removed, and 1-row nvim terminals no longer trigger automatic `/clear` loopsPlugin marketplaces: `headersHelper` on a url marketplace or a catalog entry runs a command that mints HTTP headers (e.g. a short-lived token) for catalog and same-origin archive fetches
A catalog entry's `headersHelper` runs only when you install or update that plugin, after its command is shown; `claude plugin install/update` ask `[y/N]` (or pass `-y`)Added `claude self-hosted-runner --defer-shutdown-max-min <minutes>`: on SIGTERM, keep serving attached sessions, park what is left after that many minutes, then exit
Fixed self-hosted runners occasionally being removed by the server after a single slow or lost poll request, handing their healthy session to another runnerAdded `claude self-hosted-runner --proxy-authorization-command` / `--proxy-authorization-file` for egress proxies that require a freshly issued `Proxy-Authorization` header on every connectionFixed unbounded memory growth in long interactive sessions: subagent tool results are now released once they leave the recent display windowFixed custom, project, and plugin output styles drifting back to the default voice mid-sessionFixed MCP elicitation dialogs showing nothing for URLs longer than 4,096 characters, and permission prompts dropping the "don't ask again" option when the project path didn't fit the terminal width
Fixed text-wrapping in permission prompt diffs: lines containing wide multi-code-point characters (such as emoji) or tabs are no longer clippedFixed held Backspace being ignored on terminals that send Ctrl+H for Backspace when keystrokes arrive in large bursts (slow SSH/mosh links)
Fixed killing a suspended (Ctrl+Z) session sometimes leaving the terminal in bracketed-paste mode with the cursor hiddenImproved startup: bare `claude` starts sooner on macOS
Improved startup responsiveness: the automatic update check now runs about 10 seconds after launch instead of competing with startup for CPUMCP `headersHelper` in a project `.mcp.json`, and inline MCP servers in project or `--add-dir` agent files, now require that folder's trust dialog to have been accepted (also under `claude -p`)
MCP `headersHelper` from a project `.mcp.json`, plugin, or agent file runs without inherited credential env vars; user, managed and claude.ai-scope helpers now run from the Claude config dirFixed stdio MCP servers receiving a `server/discover` request before `initialize`, forcing lazy servers to start their backend on every session open
Changed `claude mcp list` and `claude mcp get` to show disabled servers as `⊘ Disabled` instead of connecting to them for a health checkFixed worktree-isolation Bash refusals telling you to remove a redirect when the command had none
Improved Bash tool permission checking for zsh-specific syntax in shell conditionalsFixed leftover `/tmp/claude-*-cwd` files when a Bash command is killed, times out, or is interruptedFixed a proxy's refusal of a connection being reported as a generic network error instead of naming the proxy
Fixed Remote Control disconnecting with "login expired" when a brief network hiccup delays renewing your sign-in; it now retries and stays connected
Fixed Remote Control reporting a failed reconnect on sign-out; signing out now ends the session with a clear message
Improved Remote Control connection resilience: brief HTTP 403 refusals from a network edge, VPN, or proxy are now tolerated for up to 3 minutes, with the refusing party named when a block persistsFixed Remote Control messages sent from the web or Desktop while Claude is mid-turn disappearing from the transcript after the turn finishes
Fixed Remote Control model picks made on a phone or web not updating the model shown in the terminalFixed per-task Stop from the Remote Control tasks panel doing nothing on CLI-hosted sessionsFixed remote sessions exiting when a client delivered a user message without a valid roleFixed Remote Control sessions started by `claude remote-control` inheriting session-scoped environment variables from the launching shellFixed a Remote Control session whose process crashed staying unavailable until `claude remote-control` was restarted; it can now be reused when you next message itCross-session messaging: sending to a session on this machine that refuses inbound messages (e.g. `crossSessionInbound: "refuse"`) now reports "refused" to the sender instead of a silent success
Cross-session messaging: a session whose inbox drops your messages (rate limit or full queue) now tells your session, instead of the messages vanishing silentlyFixed `ListAgents`/`SendMessage` reporting "Remote Control is not connected" in sessions run by `claude remote-control` (server mode) or Desktop/IDE hosts; they now list and reach Remote Control peers
Fixed `ListAgents` and `SendMessage` exposing the idle worker that the agent view pre-warms for your next background session; it now appears only once a task claims itFixed `CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=true` not keeping prompt suggestions on when your account is near, but not over, its usage limit
Fixed the `/model` and `/effort` cache-miss warning appearing when the prompt cache had already expiredUpdated the bundled `claude-api` skill for the Managed Agents Aug 19 release: web search/fetch domain settings and memory stores on self-hosted sandboxes