Skip to main content
Glama

agyenvoy

A Claude Code plugin (MCP server + skill) that lets one agent dispatch the agy (Antigravity / Gemini) CLI. Delegate prompts to agy's models, continue/resume conversations, list models, check auth/credit, log in, and generate real images via agy's native image model.

Claude Code only. The MCP server is a single self-contained server.mjs (deps bundled) — no npm install needed to run it.

Tools

Tool

Purpose

agy_ask

Run a prompt through agy --print. Params: model, add_dir[], project, new_project, continue_conversation, conversation_id, timeout_ms. Blocks with a hint if logged out.

agy_list_models

List models available to agy. No args.

agy_auth_status

Report login state + per-model quota (credit). Read-only, never returns the token.

agy_login

Ensure agy is logged in; if not, open a terminal for Google sign-in and poll until done (macOS).

Related MCP server: Sayalabs MCP

Skills

Each skill teaches the host agent to reach for agy where it beats Claude's native abilities:

Skill

Covers

agy-image-gen

Generate/edit raster images (native-model phrasing, matplotlib-trap avoidance, verify step).

agy-media-understanding

Transcribe audio and describe/summarize video — Claude can't hear or watch media; Gemini can.

agy-second-opinion

Cross-family review of a plan/diff/answer via Gemini or GPT-OSS models.

Prerequisites

  • Node.js >= 18.

  • agy CLI installed — agy models should work from your shell. If it isn't on $PATH, set AGY_BIN to the full path. (Auth itself is handled by agy_login.)

Install (as a plugin)

/plugin marketplace add yinjs/agyenvoy      # local; or <git-owner>/agyenvoy once pushed
/plugin install agyenvoy@agyenvoy

Then restart Claude Code (or /mcp) to load it.

Install (raw server, no plugin)

Copy .mcp.json.example to your project's .mcp.json, fix the absolute paths, done.

Gotcha: request timeout

agy responses exceed the default 60s MCP request timeout. Raise it in the environment launching Claude Code, or calls report a spurious client-side timeout:

export MCP_TOOL_TIMEOUT=600000   # 10 minutes

agy_ask's own timeout_ms (default 5m) governs the server-side wait.

Security

agy_ask runs agy with --dangerously-skip-permissions so tool actions inside agy are auto-approved (required for non-interactive use — no TTY to approve). Use with trusted prompts only.

Develop

npm install
git config core.hooksPath .githooks   # once: blocks commits with a stale server.mjs
npm run build     # tsc typecheck
npm run bundle    # -> server.mjs (self-contained, shipped)

Source: src/index.ts. The tracked server.mjs is the built artifact the plugin runs.

Available Tools

4 tools
agy_askAsk agy (Antigravity/Gemini)A

Send a prompt to the agy CLI (Antigravity, Gemini-backed) and return its response.

Runs agy non-interactively in print mode. Use this for a second opinion from Gemini/Claude/GPT models, to delegate a self-contained task, or to continue a multi-turn agy conversation.

Args:

  • prompt (string, required): The prompt to send.

  • model (string, optional): A model from agy_list_models (e.g. "Gemini 3.1 Pro (High)"). Omit for default.

  • add_dir (string[], optional): Absolute dirs to add to agy's workspace so it can read those files.

  • project (string, optional): Existing agy project ID.

  • new_project (boolean, default false): Create a new project for this session.

  • continue_conversation (boolean, default false): Continue the most recent conversation.

  • conversation_id (string, optional): Resume a specific conversation by ID (overrides continue_conversation).

  • timeout_ms (number, default 300000): Max wait in ms (1000..1800000).

Returns JSON: { "response": string, "model": string, "exit_code": number }. On failure exit_code is non-zero and response contains agy's stderr (e.g. binary not found, timeout).

Note: runs agy with --dangerously-skip-permissions so tool actions inside agy are auto-approved (required for non-interactive use). Only invoke with prompts you trust to run unattended.

Inputs: prompt (required), model, add_dir, project, +4 more.

agy_auth_statusCheck agy auth status & creditA
Read-onlyIdempotent

Report whether agy is logged in and how much per-model quota (credit) remains.

Reads agy's local account store (~/.antigravity_tools). Never returns the auth token.

Takes no arguments. Returns JSON: { "logged_in": boolean, // true if a current account has a valid, non-disabled token "email": string | null, // current account email "name": string | null, // current account display name "accounts": string[], // all known account emails "models_quota": [ { "display_name": string, "percentage": number, "reset_time": string } ] // percentage = remaining credit for that model (100 = full) }

Use before agy_ask to confirm agy is usable; if logged_in is false, call agy_login.

Inputs: no named parameters.

agy_list_modelsList agy modelsA
Read-onlyIdempotent

List the models available to the agy CLI (runs agy models).

Use before agy_ask to discover valid values for its 'model' argument.

Takes no arguments. Returns JSON: { "models": string[] } — e.g. ["Gemini 3.5 Flash (High)", "Gemini 3.1 Pro (High)", "Claude Opus 4.6 (Thinking)"].

Inputs: no named parameters.

agy_loginLog in to agy (Google auth)A

Ensure agy is logged in. If already authenticated, returns immediately.

Otherwise opens a terminal window running agy so you can complete Google sign-in in your browser (agy drives the OAuth flow itself), then polls agy's account store until a valid account appears and reports the signed-in email.

Args:

  • timeout_ms (number, default 180000): how long to wait for login to complete (10000..600000).

Returns JSON: { "logged_in": boolean, "email": string | null, "already": boolean, "message": string }.

  • already=true means agy was already logged in (no terminal opened).

  • logged_in=false with a message means the wait timed out (finish login and call agy_auth_status).

macOS only (opens Terminal.app). On other platforms, run agy yourself to log in, then use agy_auth_status.

Inputs: timeout_ms.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.6/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clear, distinct purpose: asking the CLI, checking auth status, listing models, and logging in. There is no functional overlap or ambiguity among them.

    Naming Consistency5/5

    All tool names follow a consistent 'agy_verb' pattern (ask, auth_status, list_models, login) using snake_case. The pattern is predictable and easy for an agent to learn.

    Tool Count5/5

    With only 4 tools, the server is tightly scoped to the essential operations of the agy CLI: authentication, model discovery, and making requests. Each tool is necessary and there is no bloat.

    Completeness4/5

    The tool surface covers the core interaction loop: authenticate, check status, list models, and send prompts. Minor gaps exist—e.g., no tool to list ongoing conversations or manage projects directly—but these are partially addressed via parameters in agy_ask, so agents can still function.

    Related MCP Connectors

    Related MCP Servers