Antigravity 2

Antigravity 2.0

Use this skill when the user asks Codex to connect to, set up, inspect, compare, or automate the local Antigravity 2.0 desktop app.

Core jobs:

• Set up or verify the plugin on another Windows machine.
• Open Antigravity and report install/runtime status.
• Read model quota and AI credit state from Antigravity's local language server.
• Inspect what is happening live in Antigravity.
• See projects and chats through the live UI.
• Continue an existing chat after verifying the selected project, conversation, and model.
• Start a new chat in an existing project.
• Start a new project and then start a chat there.
• Report quota, model, UI, or submission errors without repeatedly retrying.
• Switch the active chat to an available model automatically when the selected model is exhausted or unsuitable for cost-saving handoff.
• Create durable bridge jobs and read compact artifacts instead of inspecting full Antigravity chats.

Primary operating model:

• Codex delegates nontrivial work to Antigravity first.
• Codex waits instead of doing the same exploration itself.
• Codex reads only compact status artifacts, targeted diffs, or concise visible status.
• For coding/workspace tasks that do not need visible desktop UI state, Codex should prefer submit-agy-job first. It uses official Antigravity CLI print mode and writes durable bridge jobs under .antigravity-bridge/jobs/<jobId>/.
• Use desktop UI jobs only when visible project/chat, Manager, Editor, or model-picker state is required.
• Codex should later read only result.md, changed-files.txt, diff.patch, test-output-summary.md, and status.json.
• Codex reviews Antigravity's result, identifies gaps, and gives Antigravity the next compact improvement task.
• Codex performs final user-facing synthesis, safety checks, and narrow code patches only when that is cheaper or higher quality than another Antigravity pass.
• Codex should not duplicate Antigravity's file reading, broad search, browser operation, or long reasoning unless Antigravity is blocked, unavailable, or the task is tiny.

MCP Tool Surfaces

This plugin exposes two MCP servers:

• antigravity-local: direct tools for quick, setup, doctor, status, open, repair-live, inspect, live, devtools-health, submission-guide, prepare-offload, create-job, submit-job, agy-status, agy-models, submit-agy-job, list-jobs, read-job, cancel-job, retry-job, switch-model, submit-offload, limits-summary, limits, models, offload-advice, handoff-template, and privacy.
• antigravity-devtools: Chromium DevTools controls for inspecting and driving the Antigravity UI.

Prefer antigravity-local.submit-agy-job for nontrivial coding/workspace work unless the task requires visible desktop project/chat state. It creates a durable .antigravity-bridge/jobs/<jobId>/ folder, runs official Antigravity CLI print mode, avoids desktop RAM overhead, and lets Codex stop watching the UI. Use antigravity-local.read-job later to read only the artifact files. Use antigravity-local.submit-job only when a visible Antigravity project/chat must be used. Use antigravity-local.submit-offload only for lightweight selected-chat handoffs that do not need a durable job folder. If Sonnet/Opus/GPT-OSS is exhausted in desktop UI, do not wait for the user to say so: call antigravity-local.switch-model with modelPreference=flash-medium or pass modelPreference=flash-medium to submit-job / submit-offload. If the MCP tool list is stale and does not show the job/model tools, use the PowerShell helper equivalents before falling back to DevTools choreography. Use antigravity-local.prepare-offload when Codex should first show the plan or when the selected chat is uncertain. For any nontrivial workspace, repo, browser, UI, research, planning, debugging, review, implementation, or job-application task, default to Antigravity for exploration and long reasoning, while Codex stays the planner, safety gate, patch reviewer, and final summarizer. Use antigravity-local.quick for general setup checks. If ReadyForLiveUiInspection is false or PageCount is zero, call antigravity-local.repair-live once before using DevTools. If repair-live restarts Antigravity, do not keep using an already-started stale DevTools MCP connection; let it reconnect to the new port before UI calls. If antigravity-devtools fails with Transport closed, do not repeatedly call list_pages in that session. Call antigravity-local.devtools-health; if it reports pages are ready, restart Codex to recreate the DevTools MCP transport or use handoff-template for a manual paste this turn. Use limits-summary for normal quota checks and full limits only when complete per-model JSON is needed. Use antigravity-devtools for project/chat selection only when the selected chat is not already correct. If this skill file cannot be read in a Codex session, run the PowerShell helper fast path.

Existing-chat rule: if the user asks for an existing project/chat, do not create or use a new chat. expectedChat must match the active Antigravity document title before model switching or submission. If the helper reports expectedChatActiveTitle or activeExistingChat, stop and select the correct existing chat before submitting.

Submission rule: do not report success or wait for artifacts unless the helper returns Submitted: true. If it returns Submitted: false, ComposerStillHasPrompt: true, or submit_failed, tell the user the prompt was not accepted and fix selection/submission first.

Startup rule: plugin MCP startup must be passive. Do not open, close, restart, or repair Antigravity just because Codex opened. agy-status and agy-models are passive and do not open desktop UI. Only call open, repair-live, switch-model, submit-job, or submit-offload after the user asks to use desktop Antigravity or the task explicitly requires visible UI state.

Requirements

• Windows desktop environment.
• Antigravity installed at the standard per-user location: %LOCALAPPDATA%\Programs\Antigravity\Antigravity.exe.
• Antigravity user data at: %APPDATA%\Antigravity.
• Plugin installed at: %USERPROFILE%\plugins\antigravity-2.
• Node.js available on PATH when using the bundled chrome-devtools-mcp bridge.

The helper scripts compute %LOCALAPPDATA%, %APPDATA%, and %USERPROFILE% at runtime. Do not hardcode another user's home directory, ports, project names, chats, email addresses, or runtime tokens.

First-Run Setup

After a user clones the GitHub repo into %USERPROFILE%\plugins\antigravity-2, run:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" setup

Use the setup report to decide the next step:

• If Installed is false, ask the user to install Antigravity or provide its install path.
• If Node.Found is false, install or expose Node.js on PATH before using the DevTools bridge.
• If AntigravityRunning is false, run the open command.
• If ReadyForModelLimits is false after opening Antigravity, wait a few seconds and rerun setup.
• If ReadyForLiveUiInspection is false, run repair-live once. It restarts Antigravity, clears a stale DevToolsActivePort, and waits for inspectable pages. Use Browser/Computer Use fallback only if repair still cannot connect.

Helper Commands

Check status:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" status

Fast combined readiness and quota summary:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" quick

Open Antigravity:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" open

Repair live DevTools inspection if Antigravity is running but exposes zero pages:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" repair-live

Generate a compact handoff prompt without touching the UI:

Call antigravity-local.handoff-template with goal, workspace, statusFile, and nextStep.

Check whether a task should be offloaded before using UI tokens:

Call antigravity-local.offload-advice with goal, hasWorkspaceWork, and estimatedCodexInputTokens.

Prepare the whole token-saving handoff in one call:

Call antigravity-local.prepare-offload with goal, workspace, statusFile, nextStep, hasWorkspaceWork, and estimatedCodexInputTokens.

Create and submit a durable bridge job:

Call antigravity-local.submit-job with goal, workspace, mode, nextStep, expectedProject, expectedChat, modelPreference=auto, and submit=true.

Create and submit a low-RAM Antigravity CLI job, without opening desktop UI:

Call antigravity-local.agy-status.
Call antigravity-local.submit-agy-job with goal, workspace, mode, nextStep, model=gemini-3.5-flash-low, and start=true.

Read job results without reading the Antigravity chat:

Call antigravity-local.list-jobs with workspace.
Call antigravity-local.read-job with workspace and jobId=latest.

Use modes:

• fast: inspect only directly relevant files, make the smallest safe change, run targeted verification.
• deep: inspect related modules and prior patterns, run the strongest practical tests, include risks.
• review: inspect and report only; do not edit.
• patch: make a narrow safe edit, run relevant verification, produce a diff.

Switch the selected Antigravity chat to a cost-saving available model:

Call antigravity-local.switch-model with modelPreference=flash-medium, expectedProject, and expectedChat.

Submit the handoff in one call when the correct Antigravity chat is already selected:

Call antigravity-local.submit-offload with goal, workspace, statusFile, nextStep, expectedProject, expectedChat, modelPreference=auto, and submit=true.

Use expectedProject and expectedChat when known. If either expected string is not visible, submit-offload stops before filling or submitting. Use modelPreference=flash-medium when the user wants Flash or when a quota warning/exhausted Sonnet/Opus model is visible. submit-offload refuses to submit if it cannot verify the chosen model. Use submit=false for verify-only; it must not fill the composer. Use fillOnly=true only when the user wants to manually review the handoff before sending.

If MCP submit-offload is not visible in the current Codex session, use the raw helper:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" switch-model -ModelPreference flash-medium -ExpectedProject "<project text>" -ExpectedChat "<chat text>"
powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" submit-offload -Goal "<goal>" -Workspace "<path>" -StatusFile "notes/antigravity-status.md" -NextStep "<next step>" -ExpectedProject "<project text>" -ExpectedChat "<chat text>" -ModelPreference auto -Submit true

If MCP tools are unavailable and Codex can only run shell commands, use the PowerShell helper equivalent:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" agy-status
powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" submit-agy-job -Goal "<goal>" -Workspace "<path>" -Mode fast -NextStep "<next step>" -AgyModel gemini-3.5-flash-low

Inspect integration details:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" inspect

Inspect the live DevTools connection:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" live

Check DevTools transport health after Transport closed:

Call antigravity-local.devtools-health.

Check how to submit a chat prompt without invalid key names:

Call antigravity-local.submission-guide.

Report model limits:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" limits-summary
powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" models

limits is an alias for models. Use limits-summary unless full per-model JSON is required.

Run the public-repo privacy scanner:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" privacy

Model Limits

Use models or limits for model quota checks. The helper discovers the running Antigravity language_server.exe, reads its CSRF token from the local process command line, finds its local HTTPS gRPC-web port, and calls:

• exa.language_server_pb.LanguageServerService/GetAvailableModels
• exa.language_server_pb.LanguageServerService/GetLoadCodeAssist

This is the same local language-server surface the Antigravity Models tab uses. It returns per-model quota fraction/reset metadata and AI credit tier data when Antigravity exposes it. It is not a raw all-model token ledger unless Antigravity exposes one through these responses.

Do not parse old chats, task transcripts, or logs for model limits except as supporting evidence after the language-server path fails.

Live UI Inspection

For project and chat work, the live UI is the source of truth.

Preferred path:

1. Run setup or live to confirm DevTools readiness.
2. Use the plugin MCP server antigravity-devtools when available. It starts Antigravity's bundled chrome-devtools-mcp and connects through DevToolsActivePort.
3. Inspect the visible UI text and interactive elements through DevTools.
4. Confirm the project, conversation, composer state, and selected model before sending a message.

If antigravity-devtools/list_pages or another DevTools tool fails with Transport closed, the child MCP process has died. This is a transport failure, not proof that Antigravity is unavailable. Call antigravity-local.devtools-health to confirm the live pages. If pages are ready, restart Codex before trying DevTools again; if the user wants to continue immediately, use antigravity-local.handoff-template and ask for manual paste into Antigravity.

Manual DevTools endpoint check:

$port = (Get-Content "$env:APPDATA\Antigravity\DevToolsActivePort")[0]
Invoke-RestMethod "http://127.0.0.1:$port/json/list"

If DevTools cannot interact with a native dialog or OS-level control, use the Computer Use plugin as a fallback.

Project And Chat Workflows

See Chats In Projects

Use DevTools to inspect the live Antigravity project list, selected project, conversation list, and conversation titles. If the UI has a project or conversation search control, use it rather than reading private storage first.

Storage under %APPDATA%\Antigravity\User can support investigation, but live UI remains authoritative.

Continue An Existing Chat

Before submitting:

• Verify the intended project is selected.
• Verify the intended conversation title or visible context.
• Verify the intended model in the composer.
• Verify the composer is idle and not already streaming.
• Check models if the user asks about limits or if a quota warning is visible.

Then paste the user's instruction into the composer and submit it. Report whether Antigravity accepted the message, started working, requested confirmation, or showed a quota/error state.

Submission reliability:

• Fill or type the prompt first without a submitKey.
• Prefer clicking the visible Send/arrow button after the prompt is in the composer.
• If a keyboard submit is needed, use a separate key call with a simple accepted key such as Enter.
• Do not use Control+Enter, Ctrl+Enter, or chord strings unless the active tool schema explicitly lists that exact value. Some DevTools tools reject those strings with Unknown key.
• After submission, verify a new user message or working state. If one submit method fails once, stop retrying the same method and use handoff-template or report the blocker.

Start A New Chat In An Existing Project

Before submitting:

• Select or search for the existing project.
• Use the UI control for a new chat/conversation.
• Verify the new composer belongs to that project.
• Verify the selected model.

Then submit the user's initial prompt. Report the new conversation title or visible identifier if Antigravity shows one.

Start A New Project

Before creating:

• Ask for a project name only if the user's intended name cannot be inferred.
• Use the visible Antigravity project creation flow.
• Select local folders/workspaces only when the user explicitly requests those paths or they are already visible and clearly intended.
• Avoid broad filesystem access without explicit user instruction.

After creation, verify the project is selected, then start a chat if requested.

Tell What Is Happening Live

Inspect the active page text and state through DevTools. Summarize:

• selected project,
• selected conversation,
• selected model,
• whether the agent is idle, working, waiting for approval, or blocked,
• visible quota/errors,
• last visible meaningful action.

Do not expose unrelated private chat content unless the user asked for that specific context.

Token-Saving Offload Workflow

Use Antigravity as an offload worker when the user wants to save Codex tokens or asks Codex to "ride on" Antigravity.

Offload Decision Gate

Run antigravity-local.prepare-offload before opening or driving the Antigravity UI unless the user explicitly asks to inspect the live UI. This is the cheapest token-saving step because it replaces several separate checks.

Do not offload:

• arithmetic, short factual answers, tiny transformations, or simple status questions,
• small shell checks where Codex can run one command and summarize compactly,
• prompts that can be answered without reading project files, logs, browser state, or long chat history,
• tests inside an existing project chat when the task does not need that project's workspace.

Do offload:

• most nontrivial workspace tasks that would make Codex read files, large logs, browser state, or transcripts,
• ongoing Antigravity project/chat work where the Antigravity context is already useful,
• debugging, implementation, job-search/application workflows, UI operation, reviews, planning, research, and analysis that can write a compact artifact,
• tasks where Antigravity can keep working while Codex waits and then reads only a small result.

Cost split:

• Antigravity explores, searches, reads files, inspects UI, drafts plans, and writes status artifacts.
• Codex gives Antigravity tasks, waits, reviews the compact output, tells Antigravity how to improve, verifies final diffs/tests, and explains the result to the user.
• If Antigravity output is incomplete, Codex should usually send a short follow-up task back to Antigravity rather than expanding its own context with broad file reads.

Important lesson from project chats: Antigravity may automatically inspect attached folders before answering, even for a tiny prompt. That is useful for real project work but wasteful for tests like 2+2. If Antigravity starts broad folder exploration for a small task, cancel it and report that offload is not token-efficient.

Core rule:

• Codex is the router, verifier, and final summarizer.
• Antigravity is the long-running worker.
• Files are the compact memory between them.

Do not copy large files, long logs, full source, or full Antigravity chat transcripts into Codex. Savings happen only when Codex sends compact instructions, lets Antigravity inspect the workspace locally, and reads back a small artifact or status checkpoint.

Fast preferred flow:

1. If the task is Antigravity-capable but does not require visible desktop UI/project/chat state, call antigravity-local.agy-status, optionally agy-models, then submit-agy-job; read artifacts with read-job.
2. If the correct desktop project/chat is already selected and visible UI context matters, call antigravity-local.submit-job with submit=true, modelPreference=auto, and expected visible project/chat text.
3. If a quota warning is visible or the user mentions Sonnet/Opus/GPT-OSS exhaustion in desktop UI, call antigravity-local.switch-model with modelPreference=flash-medium first.
4. If MCP job/model tools are not visible, run antigravity.ps1 submit-agy-job / antigravity.ps1 submit-job / antigravity.ps1 switch-model with the same fields.
5. If the selected chat is uncertain and UI context is actually needed, call antigravity-local.prepare-offload, then use DevTools only to select the project/chat; use switch-model for the model.
6. If the decision is codex-direct, do not open or drive Antigravity.
7. After a successful submit-agy-job or submit-job, stop watching every step. Read only the job artifacts with read-job.
8. If the artifact is weak, issue another compact Antigravity follow-up or retry-job with the exact missing point. Do not pull broad context back into Codex unless needed for final verification.

Detailed fallback flow:

1. Run antigravity-local.quick.
2. If ReadyForLiveUiInspection is false, run antigravity-local.repair-live once.
3. Run antigravity-local.offload-advice with the goal. Continue only if the decision is offload-to-antigravity.
4. Run antigravity-local.limits-summary; avoid full limits unless model-level JSON is actually needed.
5. Run antigravity-local.switch-model with modelPreference=auto or flash-medium; this is mandatory when Sonnet/Opus/GPT-OSS is exhausted.
6. Use antigravity-devtools only if needed to verify the target project, chat, idle composer, and whether workspace context is appropriate.
7. Use antigravity-local.submission-guide, then send Antigravity a compact handoff prompt with only the goal, workspace/path, constraints, next step, and required output format.
8. Ask Antigravity to inspect files locally, write progress/results to a small status artifact, and avoid pasting full files or logs.
9. Stop monitoring every step. Wait until Antigravity visibly stops or writes the status artifact.
10. Read only the small artifact or changed-file list, then summarize to the user in a few bullets.

If DevTools UI submission fails because a stale port is still attached, do not spend more tokens probing CDP from Codex. Use antigravity-local.handoff-template to prepare the compact prompt, report that UI submission was blocked by stale DevTools, and ask the user to restart Codex or paste the handoff manually. The next Codex session should load the new DevTools port.

Recommended handoff prompt:

Goal: <goal>
Workspace: <path>
Constraints: inspect files locally; do not paste full files, full logs, or full source; use search before reading whole files.
Token rule: work token-efficiently; write progress to <small-status-file>; output max 10 bullets plus changed file list.
Next step: <specific next action>
If blocked: ask one concise question; otherwise continue autonomously.

Recommended artifacts:

• notes/antigravity-status.md
• plans/antigravity-next.md
• reports/antigravity-result.json
• a Git diff, branch, or commit with a short summary

When checking back, Codex should read only the artifact, targeted diffs, or a compact visible UI status. Do not ask Antigravity to restate the whole conversation. Use delta prompts such as:

Continue from reports/antigravity-result.json. Only fix the failing item. Return max 5 bullets and changed file paths.

Report back to the user with:

• project/chat used,
• model selected,
• whether Antigravity accepted the task,
• current status,
• next action needed.

Public Plugin Hygiene

Before committing or publishing:

powershell -ExecutionPolicy Bypass -File "$HOME\plugins\antigravity-2\scripts\antigravity.ps1" privacy
git diff --check

Also run a local targeted scan for any private terms the current user mentions. Do not commit names, emails, project names, runtime ports, CSRF tokens, OAuth tokens, cookies, logs, screenshots, or Antigravity user data.

Boundaries

• This plugin is a local bridge, not a cloud service.
• It does not patch Antigravity internals.
• It does not commit runtime language-server tokens.
• It does not bypass Antigravity quota, billing, authentication, or safety controls.
• It should not read sensitive files, credentials, browser cookies, or private chat contents unless the user asks for that specific context.
• If Antigravity reports quota limits, do not keep resubmitting. Report the reset time and prepare the next action plan.