Microsoft
Copilot CLI
2 days since its last release · 1.0.109 ·
Copilot is an AI coding tool from Microsoft. Its latest release is 1.0.109, released on 10 Nov 2026. It shipped 12 releases in the last 30 days, and we have tracked 196 since 26 Sep 2025.
- Latest
- 1.0.109
- 10 Nov 2026
- Last 30 days
- 12
- releases
- Usual gap
- ~2d
- between releases
- Tracked
- 196
- since 26 Sep 2025
- Homepage
- github.com/github/copilot-cli
Releases per month
last 12 months- 8 Dec
- 23 Jan
- 20 Feb
- 16 Mar
- 24 Apr
- 16 May
- 11 Jun
- 10 Jul
- 5 Aug
- 7 Sep
- 14 Oct
- 4 Nov
Release history
196 releasesReleases 92
-
Claude Fable 5.2 arrives, plus CIDR rules for sandbox network access
- Claude Fable 5.2 is now available in the model picker under New models
- Sandbox host allow and deny rules now accept IPv4 and IPv6 CIDR ranges
- URL permission prompts can approve every subdomain for the current session
- Text typed after /init now guides the generated instructions file
- Custom agents added to .github/agents/ work mid-session without reloading, alongside many fixes
-
Shared tgrep code indexing on by default, plus /index commands
- Managed shared tgrep indexing is now on by default for eligible repositories, with new /index commands
- Windows sandbox adds enumerate-only directory grants via sandbox.userPolicy.filesystem.enumeratePaths
- Autopilot now waits for your answer to MCP input and consent requests instead of declining them automatically
- Headless -p JSON output now includes session start, resume and model-change events
- Many fixes across sandbox, MCP, SDK sessions, paste handling, help text and language server file watching
-
Unified /settings dashboard with global search, session @mentions
- /settings now opens a unified dashboard for all users, with global search across settings and local resources
- Opt-in session management tools and @mentions for opted-in local sessions
- New Fusion tier setting in copilot config and a Sandbox extensions setting to run extension processes inside the sandbox
- Terminals that support it now get inline images and program status, including when the CLI needs attention
- Long list of fixes, including blocking escape sequences in titles, tool output and prompts from restyling the terminal
-
New /code command opens sessions in VS Code, plus GPT-6.1 Terra model
- New /code command opens the current session in VS Code
- GPT-6.1 Terra is now a selectable model for accounts that have access to it
- /settings now opens a unified dashboard for everyone, with global search across settings and local resources
- New Fusion tier setting in copilot config and a Sandbox extensions setting to run extension processes inside the sandbox
- Opt-in session management tools and @mentions for local sessions, plus many terminal safety and Windows fixes
-
Assisted Permissions no longer require experimental mode
- Assisted Permissions are now available without turning on experimental mode
- Interactive and SDK Autopilot sessions keep waiting for attached background tasks instead of stopping at COPILOTTASKWAITTIMEOUTSECONDS
- The -p --autopilot mode still honors the task wait timeout
- Managed model fusion now starts sessions on Auto and applies fusionTier settings
- Auto falls back to its default preference when a saved preference is no longer available, instead of failing each turn
-
Click-to-copy code blocks and sandbox settings for ACP sessions
- Clicking a completed response code block can now copy its original code, when enabled
- ACP sessions now honor saved sandbox settings and the --sandbox/--no-sandbox flags, including on resume
- Legacy --remote/--no-remote startup flags are retired in favor of clearer hosting and connection flags for remote launch
- Windows shims like npm.cmd and dnx commands now get the same sandbox grants as their underlying tools
- Fixes for Gemini tool call schema errors, cloud replies cut off after reconnects, and GPT-5 default verbosity
-
Prompt mode uses Auto model selection, plus subagent approval fixes
- Prompt mode now honors managed Auto model selection for -p and piped stdin runs
- Subagents honor custom and MCP tool approvals, avoiding duplicate permission prompts
- Plugin marketplaces accept git-subdir sources and long plugin descriptions
- Replies cut off by provider overload errors now retry with backoff and show the provider's message and request ID
- Sandbox fixes: long-running commands keep running past 10 seconds and network traffic honors NOPROXY and ALLPROXY
-
Structured output for workflow subagents and Assisted Permissions guidance
- Dynamic workflow subagents get native structured output on supported models
- New assistedApprovalInstructions setting guides Assisted Permissions confirmation decisions
- Team-assigned plugins can be installed from Team-defined GitHub marketplaces
- Permission prompts now explain escalation causes and show a request-specific diagnosis command
- Many fixes for MCP sign-in, Google model image limits, model switching and Windows shell sessions
-
Sandbox gains local network access and session-only mode, plus plugin install links
- Sandbox allows localhost and private-network access by default, and /sandbox can enable sandboxing for the current session only
- Install verified Skills and plugins from GitHub install links with copilot plugin install
- Semantic code search is on by default for online sessions, and the unified Sessions browser is the default for /sessions and /resume
- New --disable-client-telemetry and --trust-folder flags, and path-specific .github/instructions files now apply automatically
- Built-in browser plugin renamed to browser-use, plus fixes for MCP server hangs, Gemini tool calls and Windows startup failures
-
Gemini 4 Argon support and easier-to-reuse /ask answers
- Adds support for the Gemini 4 Argon model
- Expand /ask answers with Ctrl+E, copy the whole answer with Ctrl+Y, and reopen the last completed answer with /ask or /btw
- Command hooks now complete reliably even when child-exit signals are lost under Node
- MCP status and /mcp commands stay synced to the active session after switching sessions in split view
- Characters typed while the input box is still loading at startup are kept instead of dropped
-
Agent-scoped hooks for custom agents, plus a long list of fixes
- Custom agents can now define their own declarative hooks
- Inline images now render by default on supported terminals; turn them off with /settings inlineImages false
- /usage shows billed AI Credits for token-billed sessions, including after resume
- Saved settings are kept when one entry in config.json is invalid, instead of being reset on every launch
- Fixes for resumed sessions, plugin hooks, custom agent discovery, MCP tools and proxy streaming
-
Sandbox settings now flag possible secrets and let you add masking hosts
- Interactive sandbox settings suggest possible environment secrets and let you add masking hosts before saving
- /add-dir now grants sandbox access to added directories for the current session
- Timeline shows whether you, Assisted Permissions, policy or unattended fallback made each permission decision
- Model IDs in /model and /config model are now case-insensitive, and canonical IDs are saved
- /user list shows the authentication source for each signed-in account, and sessions in git repositories start faster
-
Native Microsoft Entra sign-in on macOS and sandbox credential config
- Uses native Microsoft Entra broker authentication on macOS when available, falling back to the browser
- copilot config supports sandbox credential injectHosts keys, with key completion in Bash, Zsh and Fish
- --context now applies to new and resumed ACP sessions instead of silently using the default or saved context tier
- Managed plugin setup retries hourly or after policy changes rather than on every message failure
-
Claude Haiku 5.5 arrives, plus new managed policy controls for permissions
- Claude Haiku 5.5 added to model selection and --model completions
- Managed policy can disable Assisted Permissions and keep sessions in Manual Approval mode
- Assisted permissions send visible shell code to the permission judge, reducing unnecessary manual approvals
- MCP enable/disable now works before server discovery, and copilot mcp add recovers from interrupted config setup
- Warnings and update guidance shown when managed settings suppress bypass flags or request a newer CLI version
-
Command sandboxing now available to all users
- Command sandboxing is now open to everyone via /sandbox and --sandbox, and local-network allowlists include localhost and loopback hosts
- Enterprise admins can use permissions.limitTo to keep network requests within managed domain boundaries
- MCP server configuration changes now apply between turns without restarting the session
- User settings are now read only from ~/.copilot/settings.json, and user-setting keys in ~/.copilot/config.json are ignored
- Model picker now recommends GPT-6.1 Sol, GPT-6 Astra/Luna and Claude 5.5 models first
-
New config subcommands and a Ctrl+E picker for local or cloud runs
- New copilot config subcommands list, read, set and remove settings
- Ctrl+E environment picker before a conversation switches between local and cloud runs
- Choose which account to use after Microsoft Entra sign-in, and /logout now signs out those OAuth sessions
- Sandboxed shells now withhold the ambient GITHUBTOKEN unless explicitly configured
- Many sandbox, MCP, compaction and session fixes, plus faster startup with many MCP servers
-
New sandbox CA commands for managing proxy certificate trust
- New copilot sandbox ca commands check, create, trust, rotate and remove proxy CA trust, including unattended Windows setup
- /sandbox ca install is replaced by separate create and trust commands
- Complete, statically analyzable read-only shell pipelines can enter execution-evidence review; incomplete or unbound ones need explicit approval
- Sandboxed commands now run on Windows versions without filesystem enumeration support, with a warning about PowerShell's location
- Sandbox network bypass offered for Node/npm EACCES socket denials on Windows, plus fixes for session busy status and footer selection
-
GPT-6.1 Sol model support and scoped GitHub auth for MCP servers
- GPT-6.1 Sol is now available in model selection
- New --mcp-github-auth flag limits GitHub account auth to approved MCP server origins
- Path access prompts can now grant read-only access to a directory for the current session
- Auto-approval now accounts for messages typed while the agent is working
- MCP fixes: tools recover from transient discovery failures, and OAuth sign-in reuses still-valid cached tokens
-
Adds GPT-6 Sol and Luna, Claude Opus 5.5 and Claude Code rule support
- Adds GPT-6 Sol and GPT-6 Luna to the model picker when available, plus support for claude-opus-5.5
- Claude Code rule files in .claude/rules can now be used as custom instructions
- Auto suggests a routing tier you can switch with a shortcut or click, and the unsupported Fast profile is removed
- Sidebar sessions show a blue dot when a turn has finished that you have not opened yet
- PR creation now follows repository pull request templates, and many MCP, sandbox and Gemini issues are fixed
-
Fork mid-turn, terminal notifications and many MCP fixes
- Run /fork during an active turn to branch work without waiting
- Optional OSC 777 terminal notifications for direct Ghostty and WezTerm sessions
- Enterprise managed settings now apply to ACP mode, AHP hosts and --server sessions
- A custom agent's reasoning-effort now applies when the agent is selected; Enter adds a new line in freeform askuser prompts
- MCP tools recover more reliably from listing, connection and OAuth failures, and session resume no longer stalls on pending prompts
-
Combined steering prompts you can recall, plus custom worktree paths
- Consecutive steering prompts now merge into one pending message; press Up in an empty input to take it back for editing
- New worktreePathTemplate setting controls where /worktree, /move, /new and --worktree create worktrees
- User and managed startup defaults for the Auto routing tier, including strict or user-overridable organization policy
- Rubber-duck agent enabled for every model family and for low-cost-tier session models
- Many MCP fixes, including one failing server no longer removing other servers' tools and session resume no longer hanging
-
Custom agents can now opt into repository instruction files
- Custom agents can load AGENTS.md, copilot-instructions.md and CLAUDE.md by setting include-custom-instructions: true in their frontmatter
- Resumed sessions keep marketplace plugins and skills after reload, and config read or validation failures no longer drop active plugins
- Sessions can now be resumed even when transcript files contain recoverable corruption
- Autopilot now stops once task completion is accepted instead of continuing unexpectedly
- Status row shows it is waiting for background shells such as a dev server, and expanded reasoning text is no longer dimmed
-
Vim mode for everyone, reworked plugin commands and GPT-6 Astra
- Vim mode is now available to all users via /vim or the editorMode setting, with the current mode shown while you type
- Adds support for the GPT-6 Astra model
- Breaking: plugin, MCP and skill management moves to copilot plugin, copilot mcp and copilot skill, and plugins list --json now returns a flat array
- New /config sidebar, a concise transcript view, and /usage showing per-model AI Credit consumption
- Command-line parsing moved to a Rust grammar, alongside many sandbox, MCP and terminal input fixes
-
Windows taskbar session status and stricter sandbox networking
- Running sessions now appear in the Windows 11 taskbar with live hover status cards
- Custom agents can list several models to try in order, and model-policy: required keeps model changes on that list
- On macOS and Linux, sandboxed commands can no longer reach local services; enable Allow local network in /sandbox to restore localhost
- Linux sandboxing now needs slirp4netns, nsenter and iptables tools on PATH, and limits network egress to the configured proxy
- Adds claude-fable-5.1, CIMD support for MCP OAuth sign-in, HTTPS proxy mTLS certificates and a forceLoginOrgs enterprise setting
-
Plugins dashboard for everyone, crash session restore and MCP 2026-07-28
- The plugins dashboard is now the only option behind /plugin, /mcp and /skills, and /plugins has been removed
- Hooks and LSP servers can't be enabled or disabled for now, because those toggles only existed in the removed /plugins dashboard
- At startup the CLI offers to restore sessions left open by a crash or a machine restart
- Adds support for MCP 2026-07-28, Gemini 3.7 Flash and xhigh reasoning effort for Grok 4.6
- New copilot app command, login --with-token, and settings to choose the default mode and permission behavior for new sessions
-
Manage concurrent sessions, queue prompts, and sandbox changes
- Manage multiple concurrent sessions from the new Sessions tab and sidebar, and queue prompts and commands to run after the current task
- Breaking: sandbox keys renamed (allowDevToolCaches to allowDevToolAccess, gitAuth/ghAuth to sandbox.auth.) with no migration
- Combine --plan with --mode autopilot to plan first and then implement without waiting for approval
- New worktreeBaseRef setting and /worktree new; --worktree now starts from HEAD by default
- Adds the kimi-k3 model, a regrouped model picker, and /model now applies only to the current session
-
Live tool-call timers, faster session resume and git-free /rewind
- Timeline headers now show how long each tool call takes, ticking live for calls of 5 seconds or more; turn off with /settings showToolDurations
- Resuming long sessions is much faster and uses far less memory, with a 230MB transcript loading in under a second instead of about ten
- /rewind no longer needs git and restores only files Copilot changed, with a choice of conversation only or conversation plus files
- New /permissions command switches approval modes, and an experimental /new-worktree command starts a conversation in a new worktree
- Managed settings now fall back to the cache and fail open if a fetch fails; a new forceRemoteSettingsRefresh setting requires a fresh fetch at startup
-
Browser-based login becomes the default for copilot login
- Browser-based OAuth login is now the default for copilot login on local terminals; device code stays the default on remote or headless ones
- Use --web-flow or --device-code to force a login mode, or pick one in the interactive /login command
- Ctrl+G opens your editor to edit askuser freeform answers without closing the prompt
- Managed sandbox policy can be enforced through native macOS and Windows MDM settings
- Unconditional autopilot approval now disables the sandbox for the current session when bypass is allowed
-
Plugin on/off controls, grok-4.5 support and a message queue manager
- /plugins can now turn plugins, instructions, agents, LSP servers and hooks on or off
- Adds support for the grok-4.5 model
- New queue manager lets you reorder, edit, remove, repeat and immediately send queued messages
- Experimental Sessions sidebar for switching between, spawning and monitoring multiple concurrent sessions
- Enterprise admins can enforce a minimum sandbox policy, and webfetch now follows redirects and respects sandbox outbound rules
-
Plan-mode model picker, Open Plugin Spec support and Gemini 3.6 Flash
- New /model plan command picks a separate model for plan mode and reverts to the session model when you leave it
- Adds support for Open Plugin Spec v1 plugin manifests and mcp.json configuration
- Adds support for gemini-3.6-flash
- First-run splash offers to opt into the default sandbox
- Fixes for IDE reconnection, subagent timelines, /mcp wizard env values and session multiplexing dialogs
-
Plugin and skill management overhaul, keyboard-navigable Sessions sidebar
- /plugins gains update, uninstall and help commands and can manage skills, MCP servers and marketplaces, including installing skills from a file or URL
- The Sessions sidebar can now be navigated with keyboard and mouse to switch, spawn or close sessions
- New /model --session option changes the model, reasoning effort or context window for the current session only
- Sandbox now offers opt-in git and gh authentication, and macOS keychain access is off by default
- agentStop hooks that always block now end the turn after 8 consecutive blocks instead of looping forever
-
Plan mode now blocks workspace edits, plus plugin marketplace commands
- Plan mode hard-blocks built-in tools that modify the workspace, so the agent can't edit files or run mutating shell commands while planning
- /worktree now leaves uncommitted changes behind, while the new /move carries them into the new worktree
- New plugins marketplace subcommands to list, add, remove, browse and update plugin marketplaces
- Default maximum sub-agent nesting depth lowered from 6 to 4
- Adds /voice devices to choose a microphone, canvas support for extensions and a configurable Ctrl+R history limit
-
GPT-5.6 support, new /refine command and per-session sandbox flags
- Adds support for the GPT-5.6 model
- New /refine command rewrites a rough, stream-of-consciousness prompt into a clear one
- New --sandbox and --no-sandbox flags turn the OS-level shell sandbox on or off for the current session only
- Trusted repositories can pin the model, effort level and context tier, and extend deny lists, via .github/copilot/settings.json
- Plugins can be pinned to an exact commit SHA, and preToolUse hooks that exit with code 2 now deny tool calls
-
Plugins dashboard, LLM-judged auto-approval mode and sandbox updates
- New auto allow-all mode auto-approves requests that an LLM judge evaluates as acceptable; it requires experimental mode
- New /plugins dashboard to manage installed plugins, and plugin extensions now reload without restarting a session
- webfetch now follows the active sandbox network policy, and built-in file edits can bypass the sandbox when you approve it
- New /mcp list command, and the /mcp manager can now be opened while the agent is working
- Delegate PRs now target the current branch by default, with /delegate --base to pick another branch
-
Adds kimi-k2.7-code model and session switching from the agents screen
- Adds support for the kimi-k2.7-code model
- Browse, resume and switch between sessions from the agents screen
- Plan budget details now appear in the statusline and /usage for supported plans
- IDE tools stay available during brief IDE disconnects and recover automatically on reconnect
- Many fixes for sandbox paths, hooks, plugins, Windows paths and terminal rendering
-
Claude Sonnet 5 support, and disabling the sandbox now works right away
- Adds Claude Sonnet 5 as a supported model
- Disabling the sandbox for the rest of the session now takes effect immediately, so commands stop re-prompting mid-turn
- Subagent sessions now keep the parent session's tool restrictions
- Session limits must now be at least 30 AI credits
- Fixes MCP OAuth refresh and re-authentication against Microsoft Entra servers behind a tenant vanity domain
-
/pr auto drives pull requests to green, plus Claude Opus 4.8 Fast
- /pr auto now runs a self-paced loop that fixes one thing per run around CI, and /pr automerge keeps going until the PR is merged
- Adds Claude Opus 4.8 Fast and deprecates Claude Opus 4.6 Fast
- Pass a task to /worktree to name the branch and run it as the first prompt, and valid branch names are now kept exactly as typed
- The timeline now shows as a compact highlight reel for everyone, with submit times next to prompts
- Desktop notifications for attention prompts and idle sessions, and session limits now reset on /clear and use the sessionLimits key
-
New skill subcommand, CI status bar and persistent /cd directory
- New copilot skill subcommand and /skill alias list, add and remove skills from a file, URL or directory
- Opt-in status bar item shows CI check status for the current branch
- /cd now persists the working directory across resumed sessions and discovers custom agents there
- Shell history and Ctrl+R reverse search now include past shell commands in normal mode
- Open canvases resume automatically after restart, plus many fixes for MCP, Windows, tmux and permission prompts
-
MCP registry installs, an Autopilot overhaul and a pay-as-you-go budget display
- Browse and install MCP servers with /mcp registry, and MCP servers from installed plugins are now discovered automatically
- Autopilot handles permission, askuser and elicitation prompts on its own, then goes back to interactive mode once the task is complete
- Pay-as-you-go additional usage budget shown at launch, with a friendly message when the limit is reached
- New /diagnose command, inline image rendering, /security-review for all users and an experimental --worktree flag
- Option to bypass the sandbox for shell commands, an HTTP(S) proxy setting and model family aliases like opus, sonnet and gpt
-
Rewind without git, MCP deferTools option and hook matcher fixes
- Experimental: /rewind no longer needs git and restores only files Copilot changed, with a conversation-only or conversation + files choice
- New deferTools option in MCP server config keeps a server's tools always available even when tool search is enabled
- PostToolUse hook matchers like Edit|Write are now honored instead of silently dropped
- Fork-based pull requests now appear in /pr and the branch PR badge, and /diff can hide whitespace-only changes with w
- Agent mode is now tracked per session, plus clearer errors for blocked images and sign-in network failures
-
Configurable subagent models, plugin extensions and a richer diff view
- Set subagent model, reasoning effort and context tier in user settings or the /subagents picker
- Plugins can now ship extensions that install via the plugin marketplace
- Diff view adds content search, match highlighting, a file tree sidebar and inline comment editing
- Shell commands no longer run in a pseudo-terminal, so interactive input via writebash is no longer supported
- New /app command, scheduling of slash commands with /every and /after, and Kerberos proxy authentication
-
Settings dialog, /worktree command and natural-language scheduling
- New /settings dialog lets you browse and edit all user settings in one place
- New /worktree command (alias /move) creates a git worktree and switches into it, carrying uncommitted changes
- Schedule tasks in plain language with /every and /after using cron, calendar times or relative durations
- Adds support for Claude Fable 5 and auto-loads MCP servers from .github/mcp.json
- Faster grep in large monorepos plus many fixes for blank screens on resume, UTF-8 handling and crashes
-
Max reasoning effort for Anthropic models, now on every plan
- Adds a max reasoning effort level for Anthropic models, and every effort level is now available on all plans
- Ctrl+S now stashes and restores the current prompt; the slash-command picker is still available by typing /
- Create a git worktree for a pull request directly from the pull requests screen
- Security hardening: webfetch blocks private and metadata addresses, and commands setting env vars like LDPRELOAD now need approval
- On Windows, bare command names no longer run executables from the working directory, plus many rendering, MCP and LSP fixes
-
New /voice command lets you dictate prompts with local speech-to-text
- Added a /voice command for dictating prompts instead of typing them
- Dictation runs on local speech-to-text models
-
Rubber Duck and Remote JSON RPC on by default, scheduled prompts
- Rubber Duck is now enabled by default
- Remote JSON RPC is now enabled by default
- Experimental scheduled prompts with /every and /after commands
- Experimental new UI with easy access to issues, pull requests and gists
- Experimental new GitHub theme via /theme, all experimental features enabled with /experimental on
-
Shell cancel now kills whole process tree, plus HTTP/1.1 default
- Canceling a running shell command now terminates the whole process tree instead of leaving orphaned processes
- Default networking transport is now HTTP/1.1 for reliability; opt into HTTP/2 with COPILOTENABLEHTTP2=1
- preToolUse hook errors now deny the tool call instead of silently allowing execution
- New showTipsOnStartup setting, and /diff defaults to branch diff when there are no unstaged changes
- Plugin commands show progress feedback, plus many fixes for sessions, MCP servers, tmux keys and pasting
-
Free and Student users can now pick models beyond Auto
- Free and Student users can select models other than Auto in the model picker
- Diff view redesigned with continuous scrolling, sticky file and hunk headers and theme-aware colors
- MCP tools now pass both readable text and structured content to the agent instead of dropping one
- Code review agent uses the current session's model, and the reasoning effort picker only shows options the model supports
- GitHub MCP server skips tools that gh CLI can replace when gh is installed, reducing token usage
-
Claude Opus 4.8 support, focused autopilot objectives and MCP token usage
- Adds support for Claude Opus 4.8, and Claude reasoning tokens now show in session usage summaries
- New /autopilot <objective> command keeps autopilot focused on a goal, with /goal as an alias
- Free and Student plan users on token-based billing are limited to Auto model selection, with an explanation in the model picker
- Per-MCP-server token usage in /mcp and a breakdown of MCP tool tokens in /context, plus a dedicated MCP configuration screen
- New setting to block allow-all/yolo mode, and custom agents and skills are now found recursively in subdirectories
-
Context window tiers now enforced, plus /compact focus and quota bars
- Choosing the default ~200K or 1M token context tier now actually limits compaction, truncation and the token display
- Sessions resume in their saved working directory, and -C <dir> overrides it
- /compact takes optional focus instructions, and /usage shows progress bars for session and weekly quotas
- Custom agents can opt into deferred tool loading so tool search can find tools in large tool lists
- Main conversation view gets a scrollbar you can drag with the mouse, and general-purpose subagents use GPT-5.4 or GPT-5.5 when available
-
Customizable status line, session IDs and security review command
- Customizable status line in the terminal footer shows model, context window, git branch and more
- New --session-id flag resumes known sessions or starts new ones with a specific UUID
- Experimental /security-review command checks code changes for security vulnerabilities
- New /chronicle cost-tips subcommand gives personalized token usage and cost reduction tips
- Secret scanning now covers commit messages and PR descriptions, and MCP tool loading at startup is faster
-
New /rubber-duck critiques, /chronicle search and Alpine Linux support
- Add experimental /rubber-duck command that gets an independent critique of the agent's current work
- Add /chronicle search to find past session content by keyword or topic
- The CLI now runs on Alpine Linux (musl libc)
- Add /memory on|off|show, a None reasoning effort option and experimental /mcp search for installing MCP servers from the registry
- Hooks now fire for sub-agent tool calls, and postToolUse additionalContext now reaches the model
-
Model picker shows real token prices, plus a batch of fixes
- Model picker shows actual token prices instead of dot indicators for users on token-based billing
- /context now shows the correct token limit for every model instead of always showing 128k
- Instruction files with unquoted glob patterns in applyTo frontmatter are now applied correctly
- Built-in github-mcp-server is turned off automatically in Azure DevOps-only workspaces in prompt/headless mode
- Fixes for CJK and emoji input rendering, cursor placement, ACP config updates and the /ask dialog
-
Named session forks and resuming cloud agent sessions with no pushed changes
- /fork accepts an optional name, and forked sessions show where they came from in the sessions dialog
- --resume now works for Copilot cloud agent sessions where the agent hasn't pushed any changes to its branch
- j/k keys navigate up and down in the /diff view
- Copilot Max subscribers now see the correct models for their subscription tier
-
Read-only gh commands auto-approved, plus deprecation warnings and fixes
- Read-only gh CLI commands such as list, view, status and diff run without asking for confirmation
- A warning appears when the CLI version is deprecated and premium model access may be lost
- Long lines in diff view now wrap at terminal width instead of being cut off
- Sessions no longer crash mid-turn with ERRHTTP2INVALIDSESSION errors
- PowerShell starts correctly when pwsh is installed as a .NET global tool shim
-
New /autopilot and /fork commands for switching modes and splitting sessions
- New /autopilot slash command switches between interactive and autopilot modes
- New /fork command copies the current session into a new, independent session
- On Windows, falls back to Windows PowerShell when PowerShell 7+ is not installed
- OpenTelemetry output now follows GenAI semantic conventions and adds a tool execution time metric
- Starts up to about 1.5s faster on some terminals, and fixes session resume and agentStop hook bugs
-
Mid-input slash commands, multiple skills per message and hook replies
- Slash commands can now appear mid-input, and multiple skills can be invoked in a single message
- userPromptSubmitted hooks can respond directly without making a model call
- New optional prerelease argument for copilot update and /update fetches the latest prerelease build
- Shell aliases and rc file settings now work in ! commands, which work with all shell configurations
- Fixes for Free-user quota display, a Ctrl+C hang at permission prompts and startup crashes from invalid settings URLs
-
Security fix for malicious nested repos, smarter auto model routing
- Protects against remote code execution from malicious bare repositories nested inside a project
- Auto mode now uses server-side model routing for better real-time model selection
- New username toggle in the /statusline picker shows the active account in the footer
- MCP server child processes, such as those started via npx or uvx, are now fully terminated when a session ends
- Update command now shows download progress, and the resume prompt shows the correct session name when multiple sessions are active
-
Experimental Claude-powered rubber-duck agent and a new -C directory flag
- Experimental rubber-duck agent for GPT sessions, powered by Claude, available in /experimental
- New -C <directory> flag changes the working directory before starting, like git -C
- MCP server failure warnings now include stderr output and a runnable /mcp show command
- Remote session export now works with non-GitHub repositories and directories without a repo
- Fixes for the Enter key getting stuck after cancelling, false session-in-use warnings and Windows update failures
-
Faster startup, auto-installed shell completions and file attachments in prompt mode
- CLI starts faster by showing the UI right away while authentication finishes in the background
- Shell completions for bash, zsh and fish install automatically on first run and update after copilot update
- New --attachment flag in prompt mode attaches images or documents to the initial prompt
- Extensions now load in prompt mode (-p), and experimental MCP Tasks support runs task-capable tools as background agents
- Fixes for Windows crashes, plugin marketplaces added by git SSH URL, file edit reliability and @-mention completion
-
Headless OAuth for MCP servers and a rebuilt /research
- MCP servers support the clientcredentials OAuth grant type, so they can authenticate fully headless without a browser
- /research now uses an orchestrator and subagents for more thorough and reliable deep research
- Session history, file tracking and the /chronicle command are now available to all users
- Autopilot mode limits continuation messages to 5 by default, configurable with --max-autopilot-continues
- Prompt mode (-p) now runs repo hooks and workspace MCP only when opted in via environment variables
-
Background running tasks with a shortcut, plus new ACP slash commands
- Press ctrl+x then b to move the current running task or shell command to the background
- ACP sessions gain /compact, /context, /usage and /env slash commands
- ACP clients can toggle allow-all permission mode through session configuration
- The --resume session picker has a better tab layout and status display, and loads progressively
- Transient pipe errors on child process streams no longer cause crashes or false crash reports
-
Folder permission approvals now carry over between sessions by default
- Location-based permission persistence is now on by default, so approvals carry over across sessions in the same directory
- New copilot completion subcommand generates shell completion scripts for bash, zsh and fish
- Press s in the session picker to sort by relevance, last used, created or name
- /ask responses now render markdown, including tables and formatted links
- Fixes for X11 clipboard handle leaks on Linux, detached HEAD detection and skill picker display
-
Remote control toggles, keep-alive by default and double-Esc cancel
- /remote now shows current status and supports /remote on and /remote off to toggle remote control
- /keep-alive is available without experimental mode to stop the system sleeping while Copilot CLI is active
- In-flight work now needs a double Esc to cancel, preventing accidental interruptions
- preToolUse hooks with a matcher now run only for tool names that fully match the regex
- Custom agents, skills and commands from ~/.claude/ are no longer loaded, and Claude Opus 4.6 defaults to medium reasoning effort
-
Named sessions, session management, HTTP hooks and argument tab-completion
- Name sessions with --name and resume them by name with --resume=<name>; new /session delete and delete-all subcommands
- Hooks can now POST JSON payloads to a configured URL instead of running a local command
- Slash commands support tab-completion for arguments and subcommands, and Ctrl+Y accepts completion options
- New continueOnAutoMode option switches to the auto model on rate limit instead of pausing
- User settings now live in ~/.copilot/settings.json, separate from internal state in config.json
-
Earlier usage limit warnings and new slash command aliases
- Usage limit warnings now appear at 50% and 95% capacity, giving earlier notice before hitting rate limits
- New command aliases: /bug, /continue, /release-notes, /export, /reset and /upgrade
- Slash command picker suggests similar commands when you mistype one
- Resuming a remote session with --resume or --continue now keeps the --remote flag automatically
- Sub-agents in auto mode inherit the session model, and non-interactive mode waits for background agents to finish
-
Auto model selection, document attachments and usage limit warnings
- Select auto as your model to let Copilot pick the best available model for each session
- Attach supported document files to prompts for the agent to read and reason about
- Warnings appear when you approach 75% and 90% of your weekly usage limit
- New --connect flag connects directly to a remote session by ID, and --resume accepts short session ID prefixes
- Rate-limited sessions now pause queued messages and retry automatically instead of dropping them
-
New /statusline command to customize the status bar
- New /statusline command (alias /footer) chooses which items show in the status bar: directory, branch, effort, context window and quota
- Pasting images from the clipboard works again, and both Ctrl+V and Meta+V paste images on all platforms
- The --list-env flag, which logged loaded plugins, agents, skills and MCP servers in prompt mode, has been removed
- Plugin skills and commands are now found correctly when using skills.discover
- /undo explains why rewind is unavailable, for example outside a git repository or before the first commit
-
Claude Opus 4.7 support and a new --list-env flag for CI checks
- Adds support for Claude Opus 4.7
- New --list-env flag logs loaded plugins, agents, skills and MCP servers in prompt mode to help verify CI setups
- Shell commands and MCP servers now receive COPILOTAGENTSESSIONID as an environment variable
- Remote MCP server config can omit the type field, which now defaults to http
- Fixes repository owner detection from the git remote, terminal restore after a crash on Windows, and text shifting while the cursor blinks
-
Join remote control sessions from the resume picker, plus fixes
- Connect to CLI remote control sessions directly from the --resume picker
- Custom instructions and skills now reload from disk after /clear or /new
- New COPILOTDISABLETERMINALTITLE environment variable turns off terminal title updates
- Rewind picker now uses only arrow keys and Enter, dropping the 1-9 quick-select shortcut
- Fixes for submodule paths in permission prompts, false path warnings on Azure resource IDs and duplicate agent notifications
-
New /ask command for quick questions outside the conversation
- New /ask command asks a quick question without affecting conversation history
- New copilot plugin marketplace update command refreshes plugin catalogs
- Status bar now shows @files and #issues hints while typing, and a /help hint in the slash command picker
- A paused Copilot Pro trial now shows a clear message instead of a generic policy error
- Copying to the clipboard on WSL no longer adds an invisible BOM character to pasted text
-
Leaner instruction context, Remote tab fixes and many bug fixes
- Duplicate custom instruction files are no longer sent twice, and applyTo instruction files are summarized in a table to save context
- Remote tab shows Copilot coding agent tasks and supports steering without a pull request
- ACP server now binds to localhost only to prevent unintended network exposure
- Plugin hooks get PLUGINROOT-style env vars, and installing plugins from repos, URLs or local paths now shows a deprecation warning
- Fixes for stuck Escape on prompts, spurious directory access prompts, compaction session failures and Windows LSP startup
-
Remote control for CLI sessions and MCP server installs from registry
- You can now remote control your CLI sessions using --remote or /remote
- Install MCP servers from the registry with guided configuration directly in the CLI
- New /env command shows loaded instructions, MCP servers, skills, agents and plugins
- ACP clients can provide MCP servers (stdio, HTTP, SSE) when starting or loading sessions
- MCP remote connections retry automatically on transient network failures, plus fixes for /share, /add-dir, /resume and skills
-
preToolUse hooks can now modify tool input and add context
- preToolUse hooks now respect modifiedArgs/updatedInput and additionalContext fields
- Custom agent model field accepts VS Code display names and vendor suffixes, such as "Claude Sonnet 4.5" or "GPT-5.4 (copilot)"
- Terminal state is restored correctly after crashes like out-of-memory errors or segfaults
- The --remote flag is respected when the session sync prompt appears on first run in a GitHub repo
- Redesigned exit screen with the Copilot mascot and a cleaner usage summary
-
New launch flags start the CLI directly in plan or autopilot mode
- New --mode, --autopilot and --plan flags start the CLI directly in a specific agent mode
- Ctrl+L clears the terminal screen without clearing the conversation session
- /diff, /agent, /feedback, /ide and /tuikit now work while the agent is running
- Remote tab shows Copilot coding agent tasks and supports steering via the Tasks API
- Fixes a first-turn hang when the memory backend is unavailable, plus repeated beeping from BEL characters in shell output
-
MCP config now read only from .mcp.json, plus sub-agent limits
- Breaking: .vscode/mcp.json and .devcontainer/devcontainer.json are no longer MCP config sources, and a migration hint appears if only .vscode/mcp.json is found
- Custom agents can declare a skills field to load skill content into their context at startup
- New depth and concurrency limits for sub-agents prevent runaway agent spawning
- Plugins stay enabled across sessions, auto-install on startup and can show setup instructions after install
- BYOK Anthropic models now support bearer token authentication, and hooks now work correctly with them
-
New command for managing MCP servers, plus UI fixes
- New copilot mcp command for managing MCP servers
- Hooks with PascalCase event names now get VS Code-compatible snakecase payloads with session ID and ISO 8601 timestamps
- Shell sessions no longer needed now shut down automatically to reduce memory use
- Enterprise GitHub URL input in login now accepts typing and submits on Enter
- Fixes for a spinner that looked stuck, a flickering slash command picker and a timeline that went blank
-
New monitoring help topic and smoother response streaming
- New copilot help monitoring topic covers OpenTelemetry configuration with examples
- Spinner stays active until background agents and shell commands finish, and input stays available
- Azure OpenAI BYOK now defaults to the GA versionless v1 route when no API version is set
- Less UI sluggishness while responses stream live
- /yolo and --yolo now behave the same, and /yolo state persists across /restart
-
MCP server enable/disable choices now persist across sessions
- /mcp enable and /mcp disable settings now persist across sessions
- OpenTelemetry chat spans add a time-to-first-chunk attribute, and subagent spans now use the INTERNAL span kind
- Slash command timeline entries now show the command name, such as Review or Plan
- IDE auto-connect is skipped when another client is already using the session
- Fixes for plugin hook scripts on macOS and for restoring custom agents when resuming sessions
-
New Critic agent reviews plans and complex code to catch errors
- New experimental Critic agent uses a complementary model to review plans and complex implementations for Claude models
- New notification hook event fires on shell completion, permission prompts, elicitation dialogs and agent completion
- preToolUse hooks that return permissionDecision 'allow' now skip the tool approval prompt
- Session resume picker now groups sessions by branch and repository correctly on first use
-
Built-in skills ship with the CLI, plus HTTPS support for MCP OAuth
- Built-in skills now come with the CLI, starting with a guide for customizing the Copilot cloud agent's environment
- MCP OAuth flows support HTTPS redirect URIs via a self-signed certificate fallback, improving compatibility with providers like Slack
- The /resume session picker loads much faster, especially with large session histories
-
New PermissionRequest hook lets scripts approve or deny tool use
- New PermissionRequest hook lets scripts programmatically approve or deny tool permission requests
- MCP tool calls now show the tool name and a parameter summary in the timeline
- Deprecated marketplaces repository setting removed; use extraKnownMarketplaces instead
- MCP servers now load correctly after login, user switch and /mcp reload, and reconnect with valid auth when the working directory changes
- BYOK Anthropic provider now respects the configured maxOutputTokens limit
-
MCP OAuth upgrades, HTML session sharing and GPT-5.1 Codex removal
- Removes support for the gpt-5.1-codex, gpt-5.1-codex-mini and gpt-5.1-codex-max models
- New /mcp auth command with re-authentication and account switching, plus device code flow for MCP OAuth in headless and CI environments
- New /share html command exports sessions and research reports as self-contained interactive HTML files
- Ctrl+D no longer queues a message (use Ctrl+Q or Ctrl+Enter), and text selection now needs Ctrl+C or right-click to copy
- Adds postToolUseFailure hooks, a branch diff toggle in /diff and a redesigned, less verbose timeline
-
Faster startup, lower CPU use and Gemini 3 Pro Preview removed
- Support for the gemini-3-pro-preview model has been removed
- CLI starts faster thanks to parallel startup work and a V8 compile cache, and uses less CPU while streaming
- Bring Your Own Model fixes: images now reach Anthropic models and the reasoning effort setting applies correctly
- MCP fixes: Entra ID logins no longer show the consent screen every time, and OAuth works in ACP mode
- Shift+Enter inserts a newline in terminals that support the Kitty keyboard protocol
-
Rewind to any point in a conversation, plus faster startup
- /rewind and double-Esc now open a timeline picker that can roll back to any point in conversation history
- MCP servers can request LLM inference (sampling), with your approval through a new review prompt
- Faster startup from parallel terminal detection, auth and git operations, plus a V8 compile cache
- Lower CPU usage while streaming, and grep handles large files and long lines without running out of memory
- Support for the gemini-3-pro-preview model has been removed
-
Reads Claude settings files, new model picker and many MCP fixes
- Reads .claude/settings.json and .claude/settings.local.json as additional repo config sources
- Full-screen model picker lets you adjust reasoning effort with arrow keys, and the header now shows the active effort level
- /allow-all (/yolo) gains on, off and show subcommands, and /session rename can auto-generate a name
- Alt screen is now always on, and the --alt-screen flag and altscreen setting have been removed
- Fixes for MCP server loading, Windows paths and clipboard, out-of-memory crashes on heavy shell output, and stuck autopilot cancellation
-
Full monorepo support for instructions, skills, MCP servers and agents
- Custom instructions, MCP servers, skills and agents are now discovered at every directory level from the working directory up to the git root
- Organization policy for third-party MCP servers is now enforced for all users, with a warning when servers are blocked
- /clear now abandons the current session, while /new starts a fresh one and keeps the old session in the background
- /new and /clear accept an optional prompt, and /cd keeps a separate working directory per session
- Adds ~/.agents/skills/ as a personal skill directory and merges hooks from multiple extensions
-
New /undo command and experimental multiple concurrent sessions
- New /undo command undoes the last turn and reverts its file changes
- Experimental support for running multiple concurrent sessions
- SDK clients can register custom slash commands and show elicitation dialogs to the user
- Model picker now sorts models into Available, Blocked/Disabled and Upgrade tabs based on plan and policy
- Session history is no longer lost on /quit, Ctrl+C or restart, and workspace MCP servers load only after folder trust
-
Option to include gitignored files in @ file search, plus fixes
- New includegitignored config option lets @ file search include gitignored files
- Spurious ENOTCONN and EIO errors no longer show in the timeline on SSH disconnects or terminal closes
- Copying text on WSL now preserves CJK and other non-ASCII characters
- Marketplace and plugin installs from shortened URLs such as aka.ms links now work
-
Hooks in settings files, extension mode setting and MCP allowlist
- Hooks can now be defined in settings.json, settings.local.json and config.json
- New extension mode setting controls extensibility
- Experimental MCPALLOWLIST flag validates MCP servers against configured registries
- --resume now accepts a task ID as well as a session ID
- Alternate screen buffer is on by default, and idle subagents are hidden from /tasks after 2 minutes
-
Adds gpt-5.4-mini, a subagentStart hook and better theme contrast
- Adds support for the gpt-5.4-mini model, and Pro and trial users now see every model they are entitled to in the model picker
- New subagentStart hook fires when a subagent is spawned and can inject extra context into its prompt
- New "customize" mode in system message config allows section-level system prompt overrides
- Better color contrast across themes, a subtle background on user messages, and branch indicators for unstaged, staged and untracked files
- Double-Esc clears input or triggers undo, and resuming sessions created before 1.0.6 no longer fails as corrupted
-
Tool search for Claude models and plugin and hook compatibility with Claude Code
- Claude models can now discover and use tools dynamically through tool search
- Hook config files work across VS Code, Claude Code and the CLI, with PascalCase event names and nested matcher/hooks structure
- Better Open Plugins spec support, including .lsp.json, .claude-plugin/plugin.json discovery and new manifest file locations
- Extension tools now go through the permissions system, with a per-tool skipPermission option
- Many fixes for autopilot, memory usage, Windows auto-update and session crashes with sub-agents
Smaller updates 8
-
Fix for picking Copilot Fusion High or Max in long-context sessions
-
Clearer login errors and fixes for worktree switching
-
Updated model configurations
-
Fixes for Anthropic subagents and custom agent instruction links
-
Bug fixes and minor changes
-
Bug fixes for multiline prompts, skills picker and Bash sessions
-
Rate limit error now says "session rate limit"
-
Fixes prompt frame rendering on Windows and Ubuntu terminals
Showing the newest 100 of 196.