Gemini CLI
2 days since its last release · 0.68.0 ·
Gemini is an AI coding tool from Google. Its latest release is 0.68.0, released on 10 Nov 2026. It shipped 5 releases in the last 30 days, and we have tracked 169 since 25 Jun 2025.
- Latest
- 0.68.0
- 10 Nov 2026
- Last 30 days
- 5
- releases
- Usual gap
- ~4d
- between releases
- Tracked
- 169
- since 25 Jun 2025
- Homepage
- geminicli.com
Releases per month
last 12 months- 18 Dec
- 10 Jan
- 19 Feb
- 10 Mar
- 11 Apr
- 7 May
- 7 Jun
- 5 Jul
- 5 Aug
- 5 Sep
- 4 Oct
- 2 Nov
Release history
169 releasesReleases 33
-
New extension validation command, plus a wave of bug fixes
- New command validates extensions, with evaluation tests added
- Loading indicator now shows the 429 rate-limit cool-down with a countdown
- Selected model is kept across transient fallback, cancel and rewind
- Extension installs are serialized per extension, and failed release downloads are handled
- Several Windows fixes, including case-insensitive path checks and IDE process lookup
-
New models command for model discovery, plus IDE bridging into sandboxes
- New models command and --list-models flag list the models you can use
- IDE companion now works inside container sandboxes, bridged over exec stdio
- Relative file paths now resolve against the active IDE workspace root
- Non-interactive mode handles termination signals and stops tool execution after cancellation
- Fixes for display width of Thai, Lao and ambiguous-width characters, settings migration and quota error handling
-
Terminal rendering fixes and a new Gemini 4 compatibility banner
- Adds a banner noting that Gemini CLI is not optimized for Gemini 4
- Smoother terminal output with synchronized rendering, less flicker and preserved scrollback during refreshes
- MCP OAuth flows can now opt in to private network endpoints
- keybindings.json now supports a 'keys' array, and settings placeholders resolve after .env files load
- Git and shell fixes keep user-level ~/.gitconfig and disable interactive credential and SSH prompts
-
Session resume fixes, safer untrusted folders and custom OTLP headers
- Telemetry configuration now supports custom OTLP headers
- Workspace settings are now enforced as read-only in untrusted folders
- Resumed sessions no longer lose history on quick exit, duplicate tool responses or get corrupted by rapid Ctrl-C
- Fixes OAuth and verification retry loops, and re-selecting Google login now clears cached credentials
- Fixes a hang on Enter in interactive mode, and @<directory> references no longer read every file up front
-
Adds Gemini 3.8 Flash and 3.5 Flash Lite, plus terminal and auth fixes
- Adds support for the Gemini 3.8 Flash and Gemini 3.5 Flash Lite models
- OAuth refresh tokens are now kept on refresh, and deleting credentials is idempotent
- PTY and Windows ConPTY process cleanup and output handling are more reliable
- MCP tool call titles now show as structured signatures, and ACP mode sends tool call updates before permission requests
- VS Code companion keeps terminal focus when you close diff tabs, and AbortError logs no longer appear on cancel
-
Air-gapped GDC sign-in support plus a fix for quota retry hangs
- Adds support for Service Identity on air-gapped Google Distributed Cloud (GDC) after an auth library update
- Zero-quota limits now fail fast instead of hanging in a retry loop
- Fixes a path traversal vulnerability when installing skills
- The coreTools setting is migrated to tools.core
- Fixes tmux being wrongly detected as running in the background and multi-line escaped quotes in shell commands
-
Auto mode can use Gemini 3.5 Flash, plus Antigravity CLI migration help
- Auto mode now uses Gemini 3.5 Flash when the flag is enabled, and backend definitions for 3.5 Flash are respected
- New documentation and migration commands for moving to the Antigravity CLI
- Browser agent docs no longer label the feature experimental
- Fixed Vertex AI model mapping, MCP tool discovery now updates atomically, and empty resume sessions are no longer saved
- Policy files get an EBUSY fallback and recovery from TOML parse errors
-
Updated auto model routing, plus terminal and session fixes
- Default auto routing between models has been updated
- Agent-to-agent (A2A) server now exposes usage metadata
- Fixed a bypass that let blacklisted servers appear in the MCP list
- Resumed sessions no longer show internal session context in history
- Fixed Termux relaunch and resize loops, PTY resize errors and unmapped vim normal-mode keys
-
Auto modes merged into one, plus wide-ranging auth, proxy and MCP fixes
- The separate Auto modes are now a single Auto mode
- Sublime Text and Emacs Client added as supported editors
- Agent registration is now first-wins, with project agents taking priority
- Fixes for OAuth token refresh, keychain auth in non-interactive mode, NOPROXY handling and configured MCP servers in non-interactive mode
- Dynamic fallback routing for exhausted quota models, plus model aliases and thinking config for gemini-3.1 models
-
Export sessions to a file and import them back with a flag
- Sessions can now be exported to a file and imported again using a command-line flag
- The CLI interface now shows the machine hostname
- Unsigned macOS binaries are now built and attached to releases
- Fixed a chat corruption bug in the context manager and a silent hang during OAuth sign-in on headless Linux
- Explicit model choice is now respected after Flash quota runs out, and extensions can be installed from SSH repos
-
Gemma 4 models on by default, plus Auto Memory inbox and voice mode updates
- Gemma 4 models are now enabled by default via the Gemini API
- New Auto Memory inbox flow for reviewing memory updates
- Voice mode gains a microphone indicator, wave animation, privacy warning and inserts transcription at the cursor
- New /exit --delete flag, /commands list subcommand, /bug-memory command and --ignore-env flag
- Messages can be queued during compression, and the default API timeout drops to 60s with retries on timeouts
-
Real-time voice mode arrives, plus experimental Gemma 4 support
- New real-time voice mode with both cloud and local backends
- Experimental support for Gemma 4 models
- Headless mode now loads .env files securely and enforces workspace trust, and shell command validation is stricter
- Faster startup because experiments and quota are now fetched in the background
- New command-line option to set a session UUID manually, and env vars in settings.json can now be booleans or numbers
-
New gemini gemma command for easy local model setup, plus MCP resource tools
- New gemini gemma command streamlines setting up a local model
- New tools let the agent list and read MCP resources
- Ripgrep is now bundled into the standalone binary for offline search support
- Topic update narration is now on by default, plus new GitHub colorblind themes and /new as an alias for /clear
- Security fixes block overriding IDE stdio via a workspace .env file (RCE) and address shell command injection
-
New /memory inbox for reviewing and patching extracted skills
- New /memory inbox command lets you review skills extracted from your sessions, now with skill patching support
- Plan Mode now asks for confirmation before activating skills, can show plan content and silently falls back when model routing fails
- Faster startup through a lightweight parent process, plus many fixes for memory leaks, PTY exhaustion and orphaned MCP subprocesses
- Subagents now run through a single invokesubagent tool, and agent MCP server configs support an auth block
- Mouse clicks position the cursor in multi-line AskUser answers, and Ctrl+Backspace deletes words in Windows Terminal
-
Background process tools, compact tool output and persistent approvals
- New tools to monitor and inspect background processes
- Tool output is now compact by default, and loading tips are turned off by default
- Policy approvals can now persist and take context into account
- A new background memory service extracts skills, and memory is configured automatically
- Keyboard shortcut moved from Ctrl+X to Ctrl+G, plus many sandbox fixes for Linux and Windows
-
Plan mode becomes stable, with sandbox upgrades and message queuing
- Plan mode is promoted to stable, now works in untrusted folders and encourages discussion before formal plan approval
- Linux and Windows sandboxes can expand permissions dynamically, with worktree support, forbidden paths and env file secret lockdown
- Press Tab to queue a message while a response is still generating, and compact tool output reduces clutter
- Browser agent gains persistent sessions, a max actions per task setting and stricter domain restriction enforcement
- Flash Lite Preview model shown for all tiers, plus project-level memory, topic chapters, an extension update UI and a Tokyo Night theme
-
Git worktree sessions and native sandboxing on Windows and macOS
- Git worktree support lets you run isolated sessions in parallel
- Native Windows sandboxing and a stricter allowlist-based macOS sandbox, with write-protected governance files
- Plan mode now works in non-interactive mode
- Subagents get tool isolation, memory and context injection, and clearer feedback when a tool is rejected, but agents are now off by default
- Breaking: policy config entries must now specify toolName; admins can also force MCP server installs
-
Subagents switched on, plus sandboxed tools and custom keybindings
- Subagents are now enabled, with higher turn and time limits and a simpler success display
- Tools that spawn processes now run in a sandbox, with Linux support via bubblewrap and seccomp
- Keyboard shortcuts can be customized, including literal keys, extended Kitty keys and removing bindings
- Vim mode adds missing motions plus yank and paste, and safe tools can now run in parallel
- New disableAlwaysAllow setting, an --admin-policy flag and custom base URLs via environment variables
-
Plan Mode on by default, plus gVisor and LXC sandboxing
- Plan Mode is now enabled by default and no longer marked experimental
- New sandboxing options: native gVisor (runsc) support and experimental LXC containers
- Customisable footer via /footer, a new /upgrade command and a reworked thinking display
- Skills can be started with slash commands, /compact works as an alias for /compress, and /chat and /resume now behave the same way
- The --experimental-acp flag is renamed to --acp, and a fix stops long sessions from crashing by running out of memory
-
Plan mode gets feedback annotations and built-in research subagents
- Plans can now be annotated with feedback for iteration, plus a new copy subcommand and research subagents in plan mode
- Remote A2A agents now support HTTP authentication and authenticated agent card discovery
- Large MCP tool output is now truncated, and MCP OAuth support is added
- Redesigned compact header, and the context window display now shows usage instead of what remains
- ACP adds slash commands like /memory, /init, /extensions and /restore, plus a model-setting interface
-
Plan mode gets external editing and adapts to task complexity
- Plan mode can open and edit plans in an external editor, and adapts its planning workflow to how complex the task is
- Generalist agent enabled, and model steering is now enabled in the workspace
- Interactive shell autocompletion added, and extensions now load in parallel
- New G1 AI credits overage flow, a low/full error verbosity mode and an experimental Gemma router
- Extensions can use the policy engine, and orphaned processes no longer consume 100% CPU after the terminal closes
-
Gemini 3.1 Pro Preview support, plus a stricter, read-only Plan Mode
- Adds support for the Gemini 3.1 Pro Preview model, with tool descriptions and schemas tuned for Gemini 3
- Plan Mode now enforces read-only limits, switches models automatically, lets you set a custom plans folder and summarizes work after a plan runs
- Experimental browser agent and experimental direct web fetch
- Project-level policies, MCP server wildcards and tool annotation matching in the policy engine, and Always Allow now saves at workspace level
- Security hardening: deceptive Unicode is stripped from output, deceptive URLs are flagged in tool confirmations and webfetch is rate limited
-
Plan Mode upgrades, a new --policy flag and SDK groundwork
- Plan Mode gains a formal 5-phase planning workflow, skills support, project exploration without planning and per-session plan files
- New --policy flag for user-defined policies; --allowed-tools and excludeTools are deprecated in favor of the policy engine
- Tool output masking and 30-day session retention are now on by default
- Ctrl-Z suspension, multi-line answers in the ask-user tool, /commands reload and new Solarized Dark and Light themes
- Initial SDK package with support for custom skills and dynamic system instructions
-
Plan mode arrives with a /plan command, and Gemini 3 is now the default
- New experimental plan mode with a /plan command, an enterplanmode tool, plan editing and MCP server support, enabled through settings
- Preview features setting removed and the CLI now defaults to Gemini 3
- Tool outputs can be masked to save context, with outputs stored per session and restored when a session resumes
- Extension config is on by default, with a new /extensions config command and an admin allowlist for MCP servers
- New Vim motions, DevTools integration on F12, inline thinking bubbles, quota visibility and a keyboard shortcuts panel
-
Background shell commands arrive, plus Positron IDE support
- Shell commands can now run in the background
- Adds support for the Positron IDE and session resume in ACP mode
- New /prompt-suggest slash command and automatic theme switching based on terminal background
- Undo/redo keybindings change to Cmd+Z/Alt+Z and Shift+Cmd+Z/Shift+Alt+Z, and searchfilecontent is renamed grepsearch
- Security fixes: MCP servers no longer get unintended credentials, MCP OAuth now asks for consent, and folder trust is enforced
-
Plan mode arrives, plus /rewind and hooks on by default
- New Plan mode joins Shift+Tab mode cycling, with plan files saved to disk and the AskUser tool for asking clarifying questions
- New /rewind command, and the hooks system is now turned on by default
- Breaking: positional arguments now start interactive mode by default
- New /agents config command, plus commands to enable or disable MCP servers, including ones bundled in extensions
- Large pastes collapse into a placeholder you can expand, and image pasting now works on Linux with Wayland and X11
-
Agent skills on by default, plus an experimental read-only plan mode
- Agent skills are now enabled by default, with a built-in skill-creator skill, security prompts on install and warnings when skills override each other
- New experimental plan approval mode that enforces a strict read-only policy and halts on violations
- New generalist agent, the ability to enable and disable agents, and steering toward expert subagents when present
- New /introspect slash command, a rewind confirmation dialog, Esc-Esc to clear the prompt and an 'auto' alias for default model selection
- The --prompt flag is no longer deprecated, plus fixes for out-of-memory crashes in file search, shell hangs, a Windows pty crash and newline keys
-
Agent Skills switched on, with install commands and built-in skills
- Agent Skills are now enabled, with built-in skills (including a pr-creator skill) and new skills install and uninstall commands
- New /agents command lists available subagents, and subagents can be defined in settings.json or shipped as extensions
- A clihelp agent is now on by default to answer questions about the CLI
- Hooks gain enable-all and disable-all controls, and hooks can now explicitly stop or block execution
- Tab switches focus between shell and input, terminal tab titles show CLI status, and Sublime Text is now detected as an IDE
-
Agent Skills arrive, plus remote agents and stronger hook controls
- Agent Skills: reusable skills that the agent activates on its own, managed with a new gemini skills command and /skills reload
- Support for remote agents and multi-agent TOML files, with remote agents added to the agent registry
- Hooks can stop execution, block actions, modify tool inputs and inject context at session start, plus a hooks.enabled setting
- Folder trust now defaults to untrusted, and the Policy Engine replaces the old confirmation settings
- Opt-in saving of the /model choice across sessions, and experimental extension install and uninstall subcommands
-
Gemini 3 Flash arrives, plus /auth logout and HTTP/SSE MCP servers
- Gemini 3 Flash model is now available in the CLI
- New /auth logout command clears saved credentials and auth state
- Uses the official ACP SDK and supports HTTP/SSE-based MCP servers
- Windows clipboard image support, with an Alt+V paste workaround
- Hooks can have friendly names and descriptions, and hook command expansion is sanitized to prevent injection
-
Persistent "Always Allow" policies and multi-image drag and drop
- "Always Allow" tool approvals now persist, with granular control for shell commands and MCP tools
- Drag and drop several image files into the prompt at once
- /stats now shows quota for unused models, with cleaner cached-token stats
- Settings files are now validated against the settings schema
- Hooks can use transcriptpath, and the A2A server gains an /init command
-
Hooks system arrives with a management panel and extension support
- Hooks now cover session lifecycle and compression, with a commands panel to enable, disable and migrate hooks, plus support for extension hooks
- Adds support for MCP Resources and dynamic MCP tool updates, and remote MCP servers are now configured with a single url field
- Settings now have fuzzy search, and new commands list and update per-extension settings, including user-scoped ones
- /copy now works in remote sessions via OSC52, and sessions are described by a generated one-line summary
- Fixes shell output truncation and bloat, improves retry handling for network and API errors, and fixes word navigation for CJK characters
-
Hooks can now run on tool calls and model requests and responses
- Hooks now integrate with tool execution and with model requests and responses
- MCP server instructions are now always used, and Google credential headers are added to MCP requests
- New configurable inactivity timeout for shell commands, and /stats now shows how much of your usage limit remains
- Simple slash commands now run as soon as you press Enter, and /clear keeps your up-arrow input history
- Gemini 3 now explains its actions before calling tools, and a built-in holiday theme has been added
Smaller updates 67
-
Fixes for safer file writes, Ctrl+C cancellation and Windows input
-
Fixes for auth loops, memory use and non-interactive plan runs
-
Security hardening for sandbox and prompt injection, plus fixes
-
Security hardening for sandboxes, extensions, file paths and MCP auth
-
Security fixes for MCP OAuth and workspace trust
-
Sandbox isolation fix for macOS, plus history and retry fixes
-
Smarter retries on capacity errors, plus a batch of bug fixes
-
New release with no change details published
-
Security hardening and fixes for quota errors and session handling
-
Patch release with a cherry-picked fix and version bump cleanup
-
Security hardening and bug fixes for auth, sessions and history
-
Patch release with a single cherry-picked fix
-
Security fixes for the A2A server, plus fixes for agent loops and credentials
-
Fixes for file edits, plan mode paths and account error messages
-
Security fixes for file path handling, sandboxing and memory imports
-
Release pipeline fixes plus tool registry discovery
-
Bug fixes, including PTY resize crash hardening
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Maintenance update with no detailed release notes
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with no detailed release notes
-
Patch release with a single cherry-picked fix
-
Minor patch release with no detailed release notes
-
Minor point release with no detailed changes listed
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Maintenance patch with no detailed release notes
-
Maintenance release with no detailed release notes
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Bug fix patch release
-
Patch release with a single cherry-picked bug fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Maintenance release with no detailed notes
-
Patch release with a single backported fix
-
No release notes published.
-
No release notes published.
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Stable update with no detailed release notes
-
Maintenance release with no detailed release notes
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with no detailed release notes
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
-
Patch release with a single cherry-picked fix
Showing the newest 100 of 169.