GAS MCP

A Model Context Protocol (MCP) server for Google AI Studio (Gemini API),
with a built-in Ink terminal UI for API key, model, and token-usage management.

---

GAS MCP exposes every model in the Google AI Studio catalog to any MCP-compatible client through three tools: chat_completion, list_models, and get_model_capabilities. It accepts OpenAI-style messages and function schemas and translates them to Gemini's native format, supports multimodal (vision) input, and logs token usage locally. The bundled TUI lets you manage multiple API key profiles, switch the default model, and review usage — changes take effect on the running server instantly, no restart needed.

Windows only. Setup, launcher, and shortcut scripts (setup.bat, run.bat, make_shortcut.vbs) are Windows-specific.

Installation

Prerequisite: Python 3.10+ on PATH.

Option A — let a code agent install it

Paste this into any code agent (Claude Code, Cursor, Windsurf, ...) and it will download and install everything for you:

Clone https://github.com/liewcc/GAS_MCP to a local folder of my choice,
then run setup.bat inside it and confirm it finishes without errors.
It needs Python 3.10+; everything else (venv, portable Node.js, TUI deps)
is handled by the script.

Option B — manual

Clone the repo and run:

setup.bat

The script is fully automated:

1. Creates a default config.json if missing
2. Verifies Python 3.10+
3. Creates .venv and installs server dependencies (mcp, httpx, python-dotenv)
4. Downloads portable Node.js v20 into .node_venv (no system-wide Node install)
5. Installs TUI dependencies (ink, react, esbuild)
6. Creates a desktop shortcut

Connecting to an MCP client

The easiest way: paste this into any code agent (Claude Code, Cursor, Windsurf, ...):

Help me register the GAS MCP server in my MCP client config.
It is in the GAS_MCP folder cloned from https://github.com/liewcc/GAS_MCP —
locate it on this machine. The command is .venv\Scripts\python.exe server.py
run from that folder, transport is stdio, and the server name is "gas".

Or add it manually, e.g. for Claude Desktop / Claude Code:

{
  "mcpServers": {
    "gas": {
      "command": "D:\\path\\to\\GAS_MCP\\.venv\\Scripts\\python.exe",
      "args": ["D:\\path\\to\\GAS_MCP\\server.py"]
    }
  }
}

API key

Get a free API key at Google AI Studio. Then either:

• Add it as a profile in the TUI's API KEY tab (recommended — supports multiple profiles and instant switching), or
• Set GAS_API_KEY in a .env file next to server.py (fallback when no profile is active)

One key works for every model in the catalog.

TUI

run.bat

(or use the desktop shortcut — a second launch is refused while a TUI window is already open)

Tab	What it does
API KEY	Create / edit / delete / switch API key profiles (keys shown masked)
MODELS	Browse the live model catalog and pick the default_model
USAGE	Token usage per MCP call, newest first, auto-refreshing while open
EXIT	Quit (with confirmation)
AI STUDIO	Opens aistudio.google.com in the browser
REPO	Opens this repository in the browser

Keys: ←/→ switch tabs · Enter/↓ enter a tab · ↑/↓ move within lists · → jump to the right panel · Tab/Esc go back · Ctrl+C quit.

MCP tools

chat_completion

Generate a completion with any chat model in the catalog.

Arg	Default	Notes
messages	required	OpenAI-style {role, content}; system becomes Gemini systemInstruction
model	TUI's default_model	Falls back to gemini-2.5-flash if unset
temperature	0.7
max_tokens	8192
top_p	1.0
seed	none	For deterministic output
tools	none	OpenAI-style function schemas, translated to Gemini functionDeclarations

If the model requests a function call, the raw candidate content JSON (with functionCall parts) is returned instead of plain text, so the caller can execute the tool and continue.

Vision input — pass Gemini-style content parts instead of a string:

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"text": "What is in this image?"},
        {"inline_data": {"mime_type": "image/png", "data": "<base64>"}}
      ]
    }
  ]
}

list_models

Returns the sorted list of model ids available to the active API key (live query to /v1beta/models).

get_model_capabilities

Returns {type, vision, tools, context, notes} for a model id. Call it before chat_completion to check whether a model supports vision or tool calling — models of type embed, image, video, or other are rejected for chat.

Architecture notes

• Instant hot-swap — the server re-reads config.json on every call, so switching the active API key profile or default model in the TUI takes effect immediately on the running server.
• Token tracking — Gemini's usageMetadata (prompt / candidates / thoughts / cached / total tokens) is appended to usage_log.jsonl after each call, trimmed to the last 500 entries. Logging is best-effort behind a thread lock and never breaks a completion.
• Capability inference — models not listed in model_capabilities.json are classified by naming heuristics (embed, imagen, veo, aqa, gemini), so unknown models still get sensible vision/tools flags instead of malformed-payload errors.