Connect Fine Structure to Gemini

Gemini is the platform where you need to know which Gemini you mean. The consumer app has no published way to add an arbitrary remote MCP server, while Antigravity, Gemini Enterprise and the Gemini API managed agents all accept one. Antigravity is our recommended path, because it handles the OAuth flow for you.

Before you start

A Fine Structure account on the free tier or above.
For Antigravity: the Antigravity IDE or CLI installed locally.
For Gemini Enterprise: a team administrator, since this is an admin action. Google marks it a Pre-GA offering with possibly limited support.
For the Gemini API: access to managed agents on the Agent Platform, also Pre-GA, plus a scoped MCP token from the Fine Structure Studio MCP Server tab.

Step by step

Antigravity, recommended: click the three dots at the top of the agent side panel and select MCP Servers.
Antigravity: click Manage MCP Servers, then View raw config to open mcp_config.json, at ~/.gemini/config/mcp_config.json globally or .agents/mcp_config.json in a workspace.
Antigravity: add the entry from the config snippet below, using serverUrl. Google's docs state that url and httpUrl are not supported.
Antigravity: save, then complete the sign in. It performs automatic OAuth flows with dynamic client registration, so no client ID or secret is needed.
Gemini Enterprise: an administrator opens Settings & help, selects Team, then Manage team, then Connected apps, then Add MCP Server.
Gemini Enterprise: enter https://finestructure.ai/api/mcp as the MCP Server URL, plus a Name and Description.
Gemini Enterprise: the Authentication settings section asks for an Authorization URL, Token URL, Client ID, Client Secret and Scopes. Our endpoints are https://finestructure.ai/api/mcp/oauth/authorize and https://finestructure.ai/api/mcp/oauth/token, but clients register dynamically, so there is no static Client ID and Client Secret to enter. Contact Fine Structure support for this route.
Gemini Enterprise: enable the connection once created, because a new custom MCP server connection is disabled by default and invisible to the team until enabled.
Gemini API managed agents: pass a tool at interaction time with type mcp_server, a name, url set to https://finestructure.ai/api/mcp, and headers carrying Authorization set to Bearer plus your Studio MCP token.

Configuration

{
  "mcpServers": {
    "finestructure": {
      "serverUrl": "https://finestructure.ai/api/mcp"
    }
  }
}

What to watch out for

Consumer Gemini at gemini.google.com: Google publishes no general way to attach an arbitrary remote MCP server there. Treat it as not supported and use Antigravity instead.
Gemini Enterprise supports streamable HTTP only and explicitly rejects legacy SSE. Fine Structure serves streamable HTTP, so that requirement is met.
Google marks both the Gemini Enterprise connection and the managed agents route as Pre-GA, so click paths and field names can change. Re-check Google's page first.
In the Gemini API, tools supplied at interaction time completely override the agent's preconfigured tools for that turn, so include everything the turn needs in one call. Antigravity also accepts a headers object with an Authorization Bearer value.

Questions people ask

Can I add Fine Structure to the Gemini app on my phone?

No. There is no published path for adding an arbitrary remote MCP server to the consumer Gemini app on any device. Antigravity, Gemini Enterprise and the Gemini API are the three routes today.

Why does Gemini Enterprise ask for a Client ID and Client Secret when other platforms do not?

Because its form models a pre registered OAuth app rather than dynamic client registration. Fine Structure registers clients on demand as public clients with PKCE, so there is no fixed secret to paste. Antigravity does the registration itself and has no such problem.

Is serverUrl really the right key for Antigravity?

Yes. Google's Antigravity MCP documentation declares remote connections with serverUrl and states that url and httpUrl are not supported, which is why this snippet differs from the Cursor one.

Vendor documentation we checked

All platforms: connect Fine Structure to any AI. Also the documentation and pricing.