Ask Antigravity MCP Server

Community

Bridge Claude with Google's Antigravity CLI for a second model's review

How to install ↓ Compare GitHub source repository ↗

Category: Dev Tools #133 of 438
Stars: ★ 18 Very popular
Transport: stdio (local process)
Runtime: Node.js 20+
Credentials: API key / credential required
License: MIT
Last commit: today
Tools: 1

59FMRS · C

This is a single-provider MCP server that wires Google's Antigravity CLI (agy) into 40+ MCP clients such as Claude Code, Cursor, Codex CLI, and Claude Desktop, letting the host model get code review and second opinions from a different model. Strengths: simple setup and subscription-backed calls suited to one-shot reviews. Limitations: experimental, one-shot only with no session resume, and it depends on a locally installed and authenticated agy CLI; the default sandbox may need to be disabled to read directories. MIT-licensed and unofficial, with no affiliation or endorsement from Google or Anthropic.

Strongest · Documentation 14/20 Weakest · Reliability 9/20

Reliability

9/20

Security and permissions

11/20

Maintenance

12/20

Documentation

14/20

Setup experience

13/20

Why each score

Reliability 9/20: The manifest is coherent: stdio transport, npm package, version 0.0.1, and a single narrow purpose (bridge the agy CLI for a second opinion). Timeout (ASK_ANTIGRAVITY_TIMEOUT_MS, default 300000) and a sandbox toggle (ASK_ANTIGRAVITY_SANDBOX) show the author considered execution boundaries, and log-level variables have defaults. However, there is no verifiable execution evidence for this server: the repo hosts 7 MCP servers, the CI/test badges cover the whole repository rather than packages/antigravity-mcp, and no committed tests for this package are visible. The README itself labels Antigravity 'Experimental, one-shot', and the existence of a sandbox-off switch implies --sandbox may block --add-dir context reads, a known unstable path. With missing per-package test/CI evidence, the score is capped per the static calibration and set at 9.
Security and permissions 11/20: No red lines are crossed: no malware, no credential theft, no covert exfiltration, no unconfirmed delete/pay/deploy defaults, and no real tokens in install examples. There are positive signals: Antigravity runs with --sandbox on by default and can be explicitly disabled via ASK_ANTIGRAVITY_SANDBOX=0, a controllable least-privilege tradeoff; all env vars are non-secret (isSecret: false); the README discloses that prompts go to an external CLI/model vendor, requires a Google AI Pro/Ultra subscription, and states non-affiliation with vendors. Deductions: bridging to a third-party LLM inherently transfers code/prompt content off-machine, yet this package's own disclosure of data boundaries, session retention, and whether debug logs may contain source content is thin; GMCPT_LOG_LEVEL mentions debug verbosity without stating whether sensitive content is persisted; disabling the sandbox has no risk warning or confirmation step. Main risks are visible but permissions/confirmation/scoping are incomplete, so 11.
Maintenance 12/20: There are signs of active maintenance: a clear MIT license, repository not archived, CI/release/npm/version badges, a GitHub release channel, 11 open issues indicating users and a response surface, plus a migration guide (old package names to the @ask-llm org) and a docs site, all suggesting tracked version evolution. What cannot be verified statically: commit cadence, issue response time, and dependency-update policy for packages/antigravity-mcp specifically; version 0.0.1 is very early and the README calls this provider experimental; there is no SECURITY.md or explicit security-response channel, and no governance/ownership detail attributable to this subpackage. With governance and versioning gaps, the score is 12.
Documentation 14/20: Documentation is layered and reviewable: installation for Claude Code, Cursor, Codex CLI, Claude Desktop, Pi, and generic STDIO; per-provider prerequisites and auth (agy >= 1.1.5, logged in with Google AI Pro/Ultra); a tool inventory with purposes; session/sessionId semantics; doctor and REPL CLI; a migration table; a host support matrix; and disclosed limitations (experimental, one-shot, Gemini CLI enterprise gating, no Grok/Ollama fallback), plus links to a docs site and llms.txt. Deductions: the material is written for the unified package and plugin, so server-level tool parameters (e.g. ask-antigravity input schema and error codes) must be inferred; the failure mode where agy's sandbox blocks --add-dir reads is buried in an env-var description rather than a troubleshooting section, leaving hidden assumptions; cost/quota guidance is generic with no concrete subscription-cost detail. Usable but with hidden assumptions, so 14.
Setup experience 13/20: The install path is clear with examples for mainstream clients: one-line registration via npx -y @ask-llm/antigravity-mcp or claude mcp add, plus Cursor/Codex/Claude Desktop config snippets. Prerequisites state Node.js 20+, Linux/macOS, and a separately installed/authenticated agy CLI, with agy --version and doctor as self-checks, making a working connection a small, predictable number of steps. Deductions: it is not zero-config since users must install and log into the Antigravity CLI (version >= 1.1.5 with a subscription), platforms are limited to Linux/macOS (Windows unaddressed), and failure diagnosis depends on a neighboring package's doctor rather than a server-specific one; with only the manifest and no README, env-var semantics require guessing. Capped per static calibration (15) and reduced for platform and external-dependency constraints, giving 13.

Static review · not runReviewed 2026-09-25

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesRuns commands or codeUses the network

Best for

Users who already have a Google AI Pro/Ultra subscription and the agy CLI installed
Developers who want an independent review from a different vendor's model
One-shot, quota-backed second-opinion calls without metered API billing

Not for

Environments without the Antigravity CLI (agy) installed and authenticated
Workflows needing multi-turn sessions or session resume (this server is one-shot)
Production-critical flows requiring stable, non-experimental guarantees
Users who cannot send code to Google models (use the Ollama server for fully local runs)

Required permissions

Spawn the agy CLI (Antigravity CLI) process on the host machine
Use the locally stored Google AI Pro/Ultra credentials and subscription quota
Read code and file context passed in as prompt content (sandbox on by default; reading directories via --add-dir may require setting ASK_ANTIGRAVITY_SANDBOX=0)
Write local logs and npm cache

Risks and side effects

Reviewed code is sent to Google's Antigravity service, which may expose sensitive information
Sandbox is on by default; setting ASK_ANTIGRAVITY_SANDBOX=0 to read context reduces that isolation
The server is experimental and one-shot, and behavior or quota may change upstream
Long runs may time out (default 5 minutes), especially with large context reads
Subscription quota may be consumed or rate-limited
It is an unofficial third-party tool with no affiliation to Google or Anthropic

Setup

Before you start

Runtime：Node.js 20+

Other optional settings (3)

ASK_ANTIGRAVITY_TIMEOUT_MS optional Timeout in milliseconds for Antigravity (agy) execution; default 300000 (5 minutes), set it in your MCP config if needed.

ASK_ANTIGRAVITY_SANDBOX optional Set to '0' to drop agy's --sandbox flag when it blocks --add-dir context reads; sandbox is on by default (value 1).

GMCPT_LOG_LEVEL optional Log verbosity: debug, info, warn, error; default warn.

Ensure Node.js 20+ is installed (Linux or macOS).
Install the Google Antigravity CLI (agy), version >= 1.1.5, and log in once with a Google AI Pro/Ultra account; verify with agy --version.
Register the server in your MCP client. Claude Code example:

claude mcp add --scope user antigravity -- npx -y @ask-llm/antigravity-mcp

Generic STDIO client config: {"command": "npx", "args": ["-y", "@ask-llm/antigravity-mcp"]}; for Cursor write .cursor/mcp.json, for Claude Desktop write claude_desktop_config.json.
Optional environment variables: ASK_ANTIGRAVITY_TIMEOUT_MS (default 300000 ms), ASK_ANTIGRAVITY_SANDBOX (set to '0' to drop agy's --sandbox flag so --add-dir context reads work), GMCPT_LOG_LEVEL (debug/info/warn/error, default warn).
Optionally run npx @ask-llm/mcp doctor to check Node, PATH, and provider CLI presence.

.mcp.json

{
  "mcpServers": {
    "antigravity": {
      "command": "npx",
      "args": [
        "-y",
        "@ask-llm/antigravity-mcp"
      ]
    }
  }
}

Shown for Claude Code. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json

{
  "servers": {
    "antigravity": {
      "command": "npx",
      "args": [
        "-y",
        "@ask-llm/antigravity-mcp"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal

claude mcp add antigravity -- npx -y @ask-llm/antigravity-mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

Confirm that the ask-antigravity tool appears in your client's tool list, then send a prompt like “ask antigravity to review src/auth.ts”; a review coming back from Antigravity (agy) proves the connection works.

Troubleshooting

Run npx @ask-llm/mcp doctor to check Node version, PATH resolution, and whether agy is present and its version
Confirm agy --version is at least 1.1.5 and that you are logged in with a Google AI Pro/Ultra account
If the command is not found, check that npx resolves @ask-llm/antigravity-mcp, or use the globally installed executable
If sandboxing blocks --add-dir directory reads, set ASK_ANTIGRAVITY_SANDBOX=0 and assess the security impact
On timeout errors, raise ASK_ANTIGRAVITY_TIMEOUT_MS (default 300000)
Set GMCPT_LOG_LEVEL=debug for more detailed logs when calls fail
On rate limits or quota issues, wait for quota to reset or switch to another provider such as Codex, Claude, or Ollama

Things to try

Once connected, you can ask your AI assistant things like:

Ask antigravity to review src/auth.ts for security issues
Ask antigravity to debate the plan in docs/design.md
Use antigravity to get an independent second opinion on this approach

Tools 1

ask-antigravity read-only

Runs a one-shot prompt through the Google Antigravity CLI (agy) for code review, plan debate, or a second opinion; subject to subscription quota, experimental.

Use cases

Have Antigravity review a code diff and surface issues the primary model missed

Send an architecture proposal to Antigravity for critique and trade-off analysis

Get an independent second opinion on code before committing

Use Google subscription capacity for large-context reads and reviews

Supported clients

Claude Code

Cursor

Codex

Claude Desktop

Listed from the project's documentation, not tested by this site.

Overview

Ask Antigravity MCP Server is the standalone provider server inside the Lykhoyda/ask-llm repository. Over MCP it connects a host model such as Claude to Google's Antigravity CLI (the agy command), exposing an ask-antigravity tool so the host can send a prompt for code review, plan debate, or a second opinion and get an independent answer from a different model. It runs on a subscription (Google AI Pro/Ultra), is experimental, and is one-shot only — no session resume. It uses stdio transport and starts from the npm package @ask-llm/antigravity-mcp. Note: the repository also contains six other MCP servers (Codex, Claude, Grok, Ollama, Gemini, and a unified orchestrator); this profile covers only the Antigravity server and does not attribute sibling tools to it. It is an unofficial third-party tool with no affiliation, endorsement, or sponsorship from Anthropic, Google, OpenAI, or xAI.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 62.4k · Tools 2 Compare with this →

NPM Sentinel MCP Server 79 · B

AI-powered NPM package analysis MCP server

★ 18 · Tools 19 Compare with this →

MartinLoop MCP Server 79 · B

An independent control layer that gives AI coding agents budgets, brakes, and receipts

★ 191 Compare with this →

GitHub MCP Server 76 · B

GitHub's official MCP server for managing repos, issues, PRs, and workflows via natural language

★ 33.2k · Tools 15 Compare with this →

Compare with other MCP servers

More servers from this repository

Data synced 2026-09-25 Read the FMRS scoring method