# How to add an MCP server to Antigravity

Level 3: Connect AI to your tools · 6 min · Updated 19 November 2026 · Path: Connect MCP servers to your assistant

By the end you'll add an MCP server to the Antigravity app, IDE or CLI and know which transports and settings it accepts.

Antigravity, Google's agent environment, comes as an app, an IDE and a CLI, and all three take [MCP](https://ai-powered.com/academy/glossary/mcp) servers. Most of the time you'll pick one from a list; for anything else, there's one JSON file.

## The quick way: from the list

As of 19 November 2026, Google's docs describe:

- **App**: **Customizations** in the sidebar → **Installed** → under **MCP Servers**, **Add MCP Server** (**+**), find the server and **Add**. Use **More actions** (**⋮**) to restart, disable or delete it.
- **IDE**: **…** at the top of the agent side panel → **MCP Servers** → **Install** on a supported server.
- **CLI**: type `/mcp` in the prompt panel to open the MCP manager.

## Any server: mcp_config.json

In the IDE, **…** → **MCP Servers** → **Manage MCP Servers** → **View raw config** opens the file. It lives at:

- `~/.gemini/config/mcp_config.json` for every workspace,
- `.agents/mcp_config.json` inside a workspace for that workspace.

In the app, **Copy Config File Path** in the MCP section header gives you the path.

```json
{
  "mcpServers": {
    "sqlite-explorer": {
      "command": "node",
      "args": ["/usr/local/bin/sqlite-mcp-server.js"],
      "env": { "SQLITE_DB_PATH": "/var/data/app.db" }
    },
    "my-remote-server": {
      "serverUrl": "https://api.example.com/mcp/",
      "headers": { "Authorization": "Bearer YOUR_API_TOKEN" }
    }
  }
}
```

A server needs either `command` (local) or `serverUrl` (remote; `url` works as an alias).

**Streamable HTTP only**

The docs say remote servers must use Streamable HTTP: the older HTTP+SSE transport and the `httpUrl` field from Gemini CLI configs aren't supported. Copying a Gemini CLI entry as-is is the usual reason a remote server won't connect.

## Approvals

Tools from a server you haven't configured permissions for run in Ask mode, so you approve each [tool call](https://ai-powered.com/academy/glossary/tool-call). You can allow tools by pattern: `mcp(server/tool)`, `mcp(server/*)` or, for everything, `mcp(*)`. Keep the wide patterns for servers you trust completely.

Default timeouts are 300 seconds to connect and 180 seconds per tool call; `timeoutSeconds` overrides both.

## Servers developers use

MCP servers for developer tools from the directory:

- [GitHub](https://ai-powered.com/mcp/github-2): Official MCP server connecting AI assistants to GitHub repos, issues and workflows
- [fastmcp](https://ai-powered.com/mcp/fastmcp): Python framework for building MCP servers, clients and interactive apps
- [Codebase Memory](https://ai-powered.com/mcp/codebase-memory): Local MCP server that builds a code knowledge graph for AI coding agents
- [XERJ](https://ai-powered.com/mcp/xerj): Local search engine for AI agents with cited answers over your code and docs
- [REA](https://ai-powered.com/mcp/rea): CLI and MCP server that lets coding agents reverse engineer apps and binaries

## Key takeaways

- Pick from the built-in list, or edit `mcp_config.json` (global or per workspace).
- `command` for local servers, `serverUrl` for remote ones, Streamable HTTP only.
- Allow tools by pattern only when you trust the server.

## What to try next

- Vet a server before you allow it: [Is this MCP server safe?](https://ai-powered.com/academy/connect-mcp-servers/is-this-mcp-server-safe)
- Connect our read-only server to practise: [hands-on lab](https://ai-powered.com/academy/connect-mcp-servers/connect-ai-powered-mcp-server).

## Sources

Official documentation, checked on 19 November 2026:

- [MCP in Antigravity (Google Antigravity docs)](https://antigravity.google/docs/mcp)

Next lesson: [Let your AI read your Postgres database, safely (read-only)](https://ai-powered.com/academy/connect-mcp-servers/read-only-postgres)
