Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026

Antigravity CLI changelog: what's new each month

Every stable Antigravity CLI release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 6 months covered; the current month updates daily.

November 2026 · so far

13 releases: 1.5.0 → 1.5.12

November brought major new capabilities including swarms for parallel agent work, improved remote agent sessions, and broader model support. The month focused on stability, security fixes, and enhanced editor integration through the Agent Client Protocol.

Action needed
  • Auto mode is now the default for new sessions; sessions without a saved mode start in Auto (or Default if unavailable).
  • --dangerously-skip-permissions now blocks commands on the denylist instead of approving them automatically.
  • macOS Option-key shortcuts like alt+m now work by default without needing terminal configuration changes.
  • Enterprise model gateway URLs must use HTTPS unless running on localhost.

Highlights

  • Swarms let you coordinate multiple worker agents in parallel on complex tasks, with a live board to track progress.
  • Agent Client Protocol (ACP) now works as an editor server, letting editors like VS Code and JetBrains use Antigravity as their AI agent.
  • Remote agent sessions gained parity with local mode, including accept-edits, skills from cloud plugins, and improved permissions.
  • New models added: Gemini 4 Hydrogen and Neon lines with varying effort levels, plus an Orchestrator model for Gemini Enterprise.
  • Auto mode became the default for new sessions and learned to decide faster on sandbox-leaving commands by using a lighter model for review.

New

  • Swarms coordinate multiple worker agents in parallel with /swarm <task>, a live board view, and status tracking in the active items tray.
  • Agent Client Protocol server mode (agy acp) lets editors connect as ACP clients with full model access and image support.
  • Remote agent sessions can now use skills from cloud plugins and offer persistent file access choices in prompts.
  • run_workflow tool lets the agent run Python scripts that fan work to subagents in parallel or stages.
  • Split view in Remote Control browser shows multiple conversations side by side.
  • Conversation navigator in Remote Control with preview hovers and jump-to functionality.
  • Click-to-expand for tool groups, thoughts, and command output in full screen mode.
  • PLUGIN_ROOT and PLUGIN_DATA environment variables for plugin hook commands to access plugin files and data.

Improved

  • Background task handling now takes about 1 second instead of 5 for the agent to continue after a command exits.
  • Subagent rows show which tool they're blocked waiting on instead of generic "Processing" messages.
  • Plugin OAuth credentials are now protected so tokens appear only in Authorization headers.
  • Auto mode asks less often about safe commands like GitHub API reads and existing repository fetches.
  • Rate-limit handling now retries for delays up to 4 minutes instead of stopping the turn.
  • Undo in Vim mode now works like Vim, undoing one command at a time.
  • Markdown headings in responses now display cleanly with formatting instead of showing ## markers.
  • Conversation loading in Remote Control shows status and error messages while fetching older messages.

Fixed

  • File edits on Windows failing with CRLF line endings now preserve the file's line ending style.
  • Agent-run commands on Windows with non-ASCII output no longer return garbled text.
  • Several ACP session bugs including hanging prompts when sessions closed or reloaded.
  • Multi-second UI freezes when commands printed very long lines without spaces.
  • Agent replies stalling while animating long tool call output.
  • Security issue where custom headers for HTTP MCP servers could be forwarded to different hosts on redirects.

October 2026

23 releases: 1.2.15 → 1.4.12

October brought major improvements to the CLI's responsiveness, permissions model, and diff viewing experience, along with significant new features for plugins, Auto mode, and conversation management. The month saw consistent refinements across performance, reliability, and user workflow efficiency.

Action needed
  • Changed default Verbosity setting from high to medium; conversations now show concise summaries of related tool calls and thoughts.
  • Changed j and k keys in /diff file view to move line cursor like arrow keys instead of jumping between files, matching file list and Vim behavior.
  • Changed --dangerously-skip-permissions to skip only tool permission prompts, not the implementation plan review in /plan mode.
  • Changed execute_url permissions so approving Google domains no longer grants access to sensitive Google apps like mail, docs, or sheets.

Highlights

  • Added Auto mode to approve routine file edits and safe commands without prompts while still requesting approval for irreversible actions.
  • Introduced the /plugin interactive manager to browse, install, and configure plugins from the official marketplace with improved reliability on all platforms.
  • Added /compact command to summarize conversations and free up context window on demand, with automatic compaction improvements throughout the month.
  • Redesigned /diff viewer with syntax highlighting, hunk folding, word-level change highlighting, and improved navigation across multiple file views.
  • Dramatically improved CLI responsiveness by reducing memory allocation on updates, optimizing rendering performance, and pausing refreshes during idle periods.

New

  • Prebuilt native Android binaries for Termux without requiring proot-distro environment.
  • Announcement cards showing model launches, deprecations, and service notices.
  • Pinned prompt header in no-flicker mode that stays at the top while scrolling through conversations.
  • Terminal status reporting over OSC 7501 showing whether the CLI is working, waiting, done, or failed.
  • Image pasting through terminal paste shortcuts and dragging image files into the prompt.
  • Alt+m shortcut to send messages using the opposite of your Queued Messages setting.
  • Billing model choice (Gemini subscription or Pay As You Go) after Google Cloud sign-in.
  • Support for MCP servers to mark tools as always loaded via _meta field.

Improved

  • Dramatically reduced memory allocation on every update and redraw for better responsiveness while typing, scrolling, and streaming.
  • Terminal sandbox on Windows no longer requires administrator rights and includes Python, Git, and npm out of the box.
  • Always-allow suggestions for repo, go, and other commands now limited to specific subcommands instead of all variants.
  • Permission prompts now apply across main agent and all subagents instead of asking separately.
  • Agent now respects denied permission requests instead of attempting to work around them.
  • Improved Auto mode review of commands leaving sandbox with faster performance and better understanding of repository context.
  • Improved Remote Control with better support for tablets, foldables, and Business sign-in displays.
  • Improved plugin compatibility including remote MCP servers with OAuth and plugins using relative paths.

Fixed

  • Conversation view freezing after quick terminal resizes in long conversations.
  • Terminal garbled text left behind in macOS Terminal.app after exiting over SSH.
  • Trackpad and mouse-wheel scrolling jumping back and forth, especially over SSH or in tmux.
  • Files and folders with spaces, #, %, or other special characters being mishandled in file mentions and search results.
  • Markdown tables leaving misaligned borders and repeated text in scrollback during streaming replies.
  • Questions from subagents and workflow agents being lost, leaving subagents waiting forever.

September 2026

21 releases: 1.1.23 → 1.2.14

September brought substantial improvements to reliability, responsiveness, and user experience across the CLI. The month featured expanded remote control capabilities, better error handling, and refinements to core workflows.

Action needed
  • Default agent and subagent toolset retired find_by_name, grep_search, and list_dir tools from the baseline; custom agents can still access them.
  • Changed how directory entries in skills.json, rules.json, agents.json, and plugins.json are scanned to load only items directly inside the directory.
  • Plugins placed directly in ~/.gemini/config/plugins with MCP servers requiring configuration variables now start disabled until enabled.
  • --print-timeout default behavior changed: when it expires mid-turn, the CLI returns partial output and exits successfully with a warning instead of failing with a timeout error.

Highlights

  • Added remote-control start and stop subcommands to run the CLI as a background service registered with your operating system.
  • Introduced @<subagent> syntax to send messages directly to subagent conversations with autocomplete support.
  • Expanded Vim mode with numeric count multipliers and improved word-wise navigation in the prompt editor.
  • Added Kitty graphics rendering for LaTeX equations and Mermaid diagrams in the artifact viewer.
  • Improved headless runs to exit promptly, report failures clearly, and default to unlimited timeout instead of 5 minutes.

New

  • Remote Control feature to connect companion UIs to active sessions for following and controlling the CLI.
  • Half-page scrolling with Ctrl+D and Ctrl+U across all views.
  • Workspace-grouped view toggle in /resume conversation picker with Ctrl+F.
  • /model <name> <prompt> command to consult a different model mid-conversation without switching sessions.
  • /copy btw subcommand to copy side-question responses to clipboard.
  • /skills reload subcommand to reload discovered skills without restarting the session.
  • Automatic subagent guidance in Markdown custom agents and improved agent Markdown frontmatter with agents list.
  • Queued Messages option in /config to control whether follow-up messages interrupt or queue until the current turn ends.

Improved

  • Model API error resilience with longer retry windows and exponential backoff for transient failures.
  • Sign-in and startup speed by reading stored credentials instead of making network requests.
  • Headless runs to exit promptly, deliver partial output on timeout, and report errors with stable markers on stderr.
  • Tool approval prompts to show exactly what is being approved with appropriate verbs.
  • MCP panel navigation with arrow keys and improved action cycling.
  • /settings panel to explain Verbosity options as you highlight them.
  • Context compaction to spread user-request budget and size summaries from the model's full context window.
  • Scrolling performance by coalescing mouse-wheel events before redrawing and improving rendering efficiency overall.

Fixed

  • Subagent streaming overhead and overall latency by optimizing metadata transmission and startup processes.
  • Duplicate agent entries in the /agents picker when agents have the same name across configuration sources.
  • MCP configuration parsing to accept single-line and multi-line comments and trailing commas in mcp_config.json.
  • Tool invocations omitting tool-call IDs for Gemini models when reconstructing request history.
  • Headless runs leaving daemon background processes that hung script output until end-of-file.
  • Artifact viewer text selection capturing line-number gutter and fixing markdown re-wrapping after terminal resize.

August 2026

13 releases: 1.1.10 → 1.1.22

August brought major authentication and usability improvements to Antigravity CLI, with enterprise sign-in options, Vim editing support, and numerous tool and interface enhancements. The month saw consistent refinement across interactive and print modes, with fixes for credential handling and terminal compatibility.

Highlights

  • Added Vim editing mode with modal support and Insert First toggle for seamless prompt submission without leaving edit mode.
  • Enterprise users can now sign in with Business accounts, Workforce Identity Federation, or Application Default Credentials.
  • Added voice dictation through /voice command to transcribe speech directly into prompts.
  • Improved file path completion with typo tolerance and better ranking across large workspaces.
  • Added MCP server management subcommands to configure servers without hand-editing configuration files.

New

  • Voice dictation via /voice command with microphone forwarding support through mic-serve subcommand.
  • Heading outline for artifact viewer opened with t key for quick navigation of long markdown documents.
  • Vim editing mode with Normal, Insert, Visual and Visual Line modes across prompt and comment editors.
  • Non-interactive slash command answers in print mode for usage, quota, credits, model, effort, skills, permissions, hooks, help, changelog and config.
  • MCP subcommands (add, remove, list, enable, disable) for managing MCP servers without hand-editing.
  • Support for GEMINI_API_KEY environment variable to run against Gemini API without signing in.
  • Markdown alert callout rendering ([!NOTE], [!TIP], [!IMPORTANT], [!WARNING], [!CAUTION]) and carousel blocks.
  • Disable-slash-command flag for skill frontmatter to hide skills from the / menu while keeping them discoverable.

Improved

  • Terminal hyperlink support now works over SSH and in tmux, with word-wrapped URLs remaining clickable.
  • Hook ordering improved so hooks in hooks.json run before built-in termination checks.
  • Tool call headers now show a short summary of what the tool is doing instead of bare names.
  • Schedule tool accepts DurationSeconds and MaxIterations as bare JSON numbers instead of requiring strings.
  • Model retries honor server-supplied retry delays instead of client backoff.
  • Artifact detail view wraps long lines in code files instead of clipping them.
  • File path completion runs through bundled ripgrep for better performance in large workspaces.
  • Terminal sandbox now grants read-only access to .git directory instead of writable access.

Fixed

  • Fixed --model and --effort flags being ignored in interactive sessions and headless runs.
  • Fixed CLI exiting silently when language server failed at startup or during execution.
  • Fixed Enter key being swallowed while path-suggestion popup was still loading.
  • Fixed corrupted edits to files containing non-ASCII text due to inexact match splicing.
  • Fixed print mode treating benign tool execution errors and permission denials as fatal failures.
  • Fixed credentials and billing project being lost when CLI restarted and restored from keyring.

July 2026

12 releases: 1.0.15 → 1.1.9

July brought major execution-mode improvements, structured output for headless runs, and comprehensive stability fixes across interactive and print modes. The month focused on usability refinements, permission handling, and support for custom agent definitions in Markdown.

Action needed
  • Removed /fast and /planning slash commands; use /plan mode and shift+tab cycling instead.
  • Default execution mode changed to request-review; set Agent Mode in /settings to change behavior.
  • Changed default new project name from "Untitled" to "CLI Project" for clearer workspace identification.

Highlights

  • Added request-review execution mode with line-level diff previews for file operations, and mode cycling via shift+tab.
  • Introduced structured JSON and streaming JSON output formats for headless runs, enabling machine-readable results in CI and scripts.
  • Added /codesearch command with regex and literal matching to search code interactively across your workspace.
  • Launched Markdown-based custom agent definitions with YAML frontmatter as an alternative to JSON.
  • Added /effort command to adjust model reasoning effort on a timeline gauge for trading latency against depth.

New

  • Interactive status indicator below the input box showing active subagents and background tasks in real-time.
  • Agent mode selection via --agent flag and agent/agents subcommand to choose custom agents at launch.
  • In-file keyword search (/) and jump navigation (n/N) in the artifact detail viewer.
  • /plan mode replacing the legacy /planning command, with consolidated mode switching.
  • Full-screen diff view (f shortcut) for create-file confirmations matching file-edit experience.
  • Copy-on-select in no-flickering mode to auto-copy ANSI-stripped text when mouse selection is released.
  • Stacking multiple leading slash commands in a single prompt (e.g., /plan /grill-me <prompt>).
  • Support for nested subagents (grandchild and deeper) with recursive subtrajectory updates.

Improved

  • /tasks detail panel now automatically scrolls to bottom as background task logs stream in.
  • Model generation resilience improved with automatic client-side retries on transient errors.
  • File-edit diff preview with accurate line-level diffs, 3-line context, and hunk separators.
  • Interactive startup latency reduced by loading skills and MCP servers asynchronously.
  • Responsiveness on large conversations (5000+ steps) by optimizing line-count methods in no-flickering mode.
  • Permission prompts for compound shell commands now show the full command when any part needs approval.
  • /codesearch renders results progressively with live count while loading.
  • Customization discovery now sorts rules and paths deterministically to prevent prompt-cache misses.

Fixed

  • Windows print mode no longer silently discards output in non-TTY environments like pipes or subprocesses.
  • Windows editor fallback now uses "edit" or "notepad" instead of attempting "vim" when auto-configured.
  • Dynamically defined subagents now work by transitioning from JSON to Markdown format.
  • Crash when executing background tasks or terminal commands producing empty outputs.
  • Shutdown resource leaks including goroutine and database connection leaks on CLI exit.
  • Print mode now hard-fails with non-zero exit when --model cannot be resolved, instead of silently downgrading.

June 2026

2 releases: 1.0.13 → 1.0.14

June brought stability improvements and usability enhancements to the Antigravity CLI. The month focused on fixing synchronization issues, improving permission handling, and resolving bugs that affected core workflows.

Highlights

  • Goals can now run indefinitely without a length limit.
  • Command permission rules are now strict by default, with optional regex matching for advanced users.
  • Fixed critical synchronization bugs during UI transitions that could cause lockups.
  • Subagents can now auto-approve artifacts to prevent hanging when the parent is blocked.
  • Image pasting from clipboard now works in local tmux sessions.

New

  • Image pasting from clipboard in local tmux sessions.
  • Auto-approve mode for subagents to handle artifacts without parent approval.
  • Explicit opt-in for regex matching in command permission rules via prepended syntax.

Improved

  • Command permission security made strict by default instead of using regex matching.
  • Redirection checks relaxed to allow safe commands with output redirection (e.g., tool > file).
  • Exit output simplified by removing the "Resume in the same project" hint line.
  • Goal command now runs indefinitely until completed or cancelled instead of timing out.

Fixed

  • Skill commands temporarily displayed without slash prefix during optimistic updates.
  • UI transition lockups caused by missing synchronization mechanism.
  • Undo and redo history stacks desynchronized during rapid mutations in the prompt editor.
  • Browser-related prompt sections missing from the agent's prompt registry.
  • Plugin import logic stripping non-skill directories like shared/.
  • MCP configuration path mismatch between CLI and permission manager.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Antigravity CLI: pricing, features and alternatives · All changelogs