gigi-mcp

MCP server for GIGI — fiber-bundle persistent memory and geometric reasoning for Claude and other MCP-aware LLM clients.

The "Claude as voice, GIGI as brain" pattern, working out of the box.

Quick start

Step 1. Install:

pip install gigi-mcp
# or, for one-line trial without persistent install:
uvx gigi-mcp

Step 2. Add to your Claude Desktop config. Open Claude Desktop → Settings → Developer → Edit Config, and add:

{
  "mcpServers": {
    "gigi": {
      "command": "uvx",
      "args": ["gigi-mcp"]
    }
  }
}

Step 3. Restart Claude Desktop. That's it — Claude can now query GIGI directly.

Out of the box, gigi-mcp connects to the public read-only gigi-stream.fly.dev instance, so you can try it before setting up your own GIGI server. To point at your own GIGI instance, set environment variables:

{
  "mcpServers": {
    "gigi": {
      "command": "uvx",
      "args": ["gigi-mcp"],
      "env": {
        "GIGI_URL": "https://your-gigi-instance.example.com",
        "GIGI_API_KEY": "your-key-here"
      }
    }
  }
}

Try it

After restarting Claude Desktop, ask Claude things like:

• "What datasets are in GIGI?"
• "Show me the schema for the cities bundle."
• "How many cities have population over 10 million?"
• "Use GIGI Query Language to cover the cities bundle where lat > 40."
• "Export the cities bundle as DHOOM so I can see the format."

Claude will discover bundles via gigi_list_bundles, examine schemas via gigi_get_schema, run filtered queries via gigi_query_bundle, construct expressive queries with gigi_gql, and export to DHOOM when round-tripping data through other geometric tools.

What it exposes

About "GQL"

The "GQL" in gigi_gql is GIGI Query Language — GIGI's own SQL-flavored DSL. It is not GraphQL. GQL statements look like:

CREATE BUNDLE events BASE (id CATEGORICAL) FIBER (ts TIMESTAMP, val NUMERIC)
INSERT INTO events (id, ts, val) VALUES ('e1', 1700000000, 42.0)
COVER events WHERE val >= 10
SCAN events LIMIT 100

Claude reads bundle schemas via gigi_get_schema before constructing complex GQL, so the queries it builds are grounded in actual data structure.

What it doesn't do (yet)

This is v0 — intentionally minimal but covering the daily-driver surface. The roadmap:

• v0.1: Resources (gigi://bundles, gigi://schema/{name}) so Claude reads schemas as context rather than via tool calls
• v0.2: Vector search tool (gigi_vector_search) once we settle the embedding-vector serialization shape
• v0.3: Aggregation tool (gigi_aggregate) for GROUP BY / summary statistics
• v1.0: Commercial-tier operations surfaced (gigi_curvature, gigi_spectral, gigi_holonomy, gigi_transport) returning LICENSE_REQUIRED from the GIGI execution layer for non-commercial callers
• v1.x: Streamable HTTP transport, hosted endpoint option

Configuration

Why GIGI?

GIGI is a geometric database built on fiber-bundle structure. For LLMs, it provides:

• Persistent structured memory with schema that survives serialization
• Geometric reasoning primitives — curvature, spectral gap, holonomy, transport (commercial-tier operations; not exposed in v0 of this MCP server)
• GIGI Query Language for expressive SQL-flavored queries grounded in fiber-bundle semantics
• DHOOM-native export for round-tripping through geomstats and other geometric tooling

See davisgeometric.com for the broader stack.

License

MIT — see LICENSE.

GIGI itself is free for non-commercial use; commercial deployments are patent-protected (US Provisional Patent Application 64/045,889). This MCP server is free to use unconditionally — gating happens at the GIGI execution layer, not in this server.

Status

v0.0.1 — covers the six listed tools against the configured GIGI endpoint. Issues and feature requests welcome at the main repo.