Skip to main content
Glama
Antigravity Gemini MCP

Antigravity Gemini MCP

by zz327455573

Antigravity Gemini MCP

Drive your Antigravity IDE from the Gemini app you already have on your phone. No app to install, no browser tab to keep open.

License: MIT MCP OAuth Platform


The idea

Every mobile remote for a coding agent builds its own UI. That is a lot of work, and you end up with a second place to type.

This project does the opposite: it exposes your local Antigravity as a standard remote MCP server, and lets Gemini be the UI. You open Gemini — the app already on your phone — and say:

"Search what changed in Next.js 16 middleware, then have Antigravity apply it to my project and show me the diff."

Gemini's own search, reasoning and the Antigravity session all run in one conversation. That is something a browser-based IDE remote cannot do, because it only shows you the IDE.

Nothing to install on the phone. Nothing to host for the client. Just one HTTPS URL.


Related MCP server: AKIMCP

What it looks like in practice

Gemini drives your Antigravity

One conversation where Gemini searches the web, dispatches the task, and shows you the diff — an IDE-only remote can't do that.

How the pieces connect

Architecture

  Phone: Gemini app  ──►  Gemini cloud  ──►  your HTTPS endpoint
                                              │ (OAuth 2.1 + PKCE)
                                              ▼
                                     gateway.py + mcp_bridge.py
                                              │  :18900
                                              ▼
                                  agentapi.bat → Antigravity IDE
                                              │
                                        your files on disk
  • 9 MCP tools: list_projects, list_conversations, check_server_health, create_and_send_task, send_message, get_task_result, stop_task, list_files, read_file_snippet

  • Sync mode for tasks under 50 s (wait_seconds=45), async mode for long jobs (wait_seconds=0 → returns conversation_id, then poll get_task_result with offset / max_chars)

  • Multi-window routing by project_path, plus newline folding so cmd /c agentapi.bat does not truncate multi-line prompts


Compared with Google's Remote Control

Google Remote Control

This (MCP bridge)

Where you type

Antigravity's own web UI

Gemini — search, reasoning, tools, same thread

Install on phone

add to home screen

nothing

Can combine web research + code change in one turn

—

✅

Can read files off the host

—

✅ list_files / read_file_snippet

Access surface

Google's own client

Google's officially supported custom MCP connector

Setup

zero

self-hosted gateway + HTTPS endpoint

The two are complementary. Many people will run both: MCP bridge for "think + dispatch", official Remote Control for "review the diff in the real IDE UI".


Requirements

Side

Needs

Windows host

Windows 10/11, Python 3.9+, Antigravity IDE running

Phone

The Gemini app (or gemini.google.com)

Network

Public HTTPS domain with a trusted CA certificate — Gemini's backend will not reach a self-signed cert or a local address

Optional

A domain on Cloudflare (recommended for a stable tunnel URL)


Authentication

No shared secret is typed into Gemini. The bridge speaks real OAuth 2.1:

  • /.well-known/oauth-protected-resource

  • /.well-known/oauth-authorization-server

  • PKCE + RFC 7591 dynamic client registration

Token lifetimes: access 30 days, refresh 180 days. The authorization page simply asks for the auth_token you put in config.json — that string is the only credential, so make it long.


Quick start

1. Host

cd server
pip install -r requirements.txt
copy config.example.json config.json

Edit server/config.json:

{
  "host": "127.0.0.1",
  "port": 18900,
  "auth_token": "use-a-long-random-string",
  "workspace_roots": ["D:\\YourWorkspace"],
  "default_model": "MODEL_PLACEHOLDER_M526"
}

Start Antigravity IDE, then:

python gateway.py

Optional silent keepalive: watch_gateway.py, watch_cloudflared.py, launch_hidden.vbs, or a scheduled task.

2. Public HTTPS endpoint

cloudflared tunnel --url http://127.0.0.1:18900          # quick test (random URL)
cloudflared tunnel run agy-remote                        # named tunnel + your domain

A named tunnel with your own domain is strongly preferred — the MCP connection stores the URL, and a random trycloudflare.com host changes on every restart.

3. Connect Gemini

  1. Open gemini.google.com → Settings → Connected Apps → Add a custom app

  2. Paste https://your-domain.com/mcp

  3. Approve the OAuth page (enter your auth_token)

  4. Once added there, the connector also appears in the Gemini mobile app

Then just talk to it. Ask for a task, or ask it to check list_conversations first.


FAQ

Does this need a cloud PC? No. Any Windows machine that stays powered on — a desktop, a mini PC, a laptop. The only hard requirement is a public HTTPS endpoint.

Why HTTPS with a real certificate? Gemini's servers call your endpoint from the public internet. Self-signed certificates and plain HTTP are rejected by design. Cloudflare Tunnel gives you a valid cert for free.

Is my Antigravity account involved? The agent runs locally inside Antigravity with your normal quota. This bridge does not move your Antigravity credentials anywhere; Gemini authenticates to your bridge via OAuth. Access to Antigravity itself is through Google's supported custom-MCP surface.

Can a remote task brick the connection? Yes — if you ask a Gemini-dispatched session to edit or restart gateway.py / mcp_bridge.py, it kills its own link. Keep those two files out of remote scope.

Only Gemini? The server is a standard Streamable HTTP MCP with OAuth discovery, so any MCP client that supports remote servers can connect — Claude, Cursor, and others. Gemini is simply the one that needs no extra app on your phone.



中文说明

这个项目解决的不是"再造一个遥控器",而是"不要再造 UI"。

市面上 40 多个 Antigravity 手机遥控项目全都在自己写界面;这个项目反过来——把本地 Antigravity 封装成标准远程 MCP 服务器,让你手机里现成的 Gemini 当界面。于是"查资料 + 派任务 + 拉结果"能在同一个对话里完成,这是浏览器式遥控做不到的。

三个关键点

  1. 手机端零安装:不装 App、不加主屏幕,用你已有的 Gemini

  2. 走官方插槽:接入的是 Google 2026-06 开放的自定义远程 MCP 连接器,不是逆向

  3. 不需要云电脑:任何一台常开的 Windows 都行,家用台式机即可

三步接入

  1. 电脑端:cd server → pip install -r requirements.txt → 复制 config.example.json 为 config.json(改 auth_token、workspace_roots)→ 开着 Antigravity,跑 python gateway.py

  2. 公网 HTTPS:cloudflared tunnel --url http://127.0.0.1:18900 快速试用;长期建议建命名隧道 agy-remote 挂自己域名(MCP 会记住地址,随机域名重启就失效)

  3. 接进 Gemini:浏览器打开 gemini.google.com → 设置 → Connected Apps → 添加自定义 app → 填 https://你的域名/mcp → OAuth 页输入 auth_token。加完之后手机 Gemini 里也会出现

9 个工具

list_projects / list_conversations / check_server_health / create_and_send_task / send_message / get_task_result / stop_task / list_files / read_file_snippet


🔒 Security

  • 切勿提交 server/config.json、server/oauth_store.json、真实域名或令牌。
    Never commit server/config.json, server/oauth_store.json, private domains, or tokens.

  • 远程自改保护:不要让 Gemini 派发的远程会话修改或重启 gateway.py / mcp_bridge.py,否则会切断它自己的连接。
    Self-modification guard: a remote session that edits or restarts the bridge kills its own link.

License

MIT — see LICENSE.

Tool Schema Changelog

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

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      An OAuth-protected MCP server exposing a coding-tool harness over Streamable HTTP, enabling remote MCP clients like ChatGPT to securely access a local development environment.
      22
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Exposes an OAuth-gated MCP endpoint with 38 governed tools that let web-based AI chat clients read and edit local files, run whitelisted shell commands, search, use Git and SQLite, fetch localhost/intranet APIs, automate Chromium and DevTools, and manage background tasks, ports, clipboard, notifications, and Postman. Local IDEs can connect directly over 127.0.0.1, while remote clients reach the same machine through a secure public ingress.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Connects web-based AI chat clients such as Claude, ChatGPT, and Grok to a local machine through a single OAuth-gated MCP endpoint so they can browse, read, write, and edit files, run whitelisted shell commands, search content, use Git, query SQLite, drive browsers and DevTools, and manage background tasks, ports, and clipboard. It also exposes localhost and LAN services via a protected HTTP client, letting remote or mobile clients debug and operate local projects without granting unrestricted shell access.
      -
    • A
      license
      Not graded
      quality
      A
      maintenance
      Publishes an OAuth-gated MCP endpoint with 38 governed tools that let web-based AI clients read and edit local files, run whitelisted shell commands, search code, use read-only Git and SQLite, drive browser and DevTools automation, manage background tasks, ports, clipboard, and localhost services. It keeps that broad control safe by default through a strict command whitelist and configurable allowed folders instead of unrestricted shell access.
      MIT