MCP

Connect Claude, Gemini CLI, or Cursor to your Lemma specs over the Model Context Protocol. Ask an assistant to evaluate rules, explain results, inspect specs, and, when you allow it, draft or update specs in your workspace.

Prerequisites

Install the Lemma CLI so lemma is on your PATH. See Installation.

Point the server at a workspace directory (or a single .lemma file) with --prefix. That is the set of specs the assistant can see.

Connect

Clients can use stdio (spawn lemma mcp) or Streamable HTTP (lemma mcp --http). Replace /path/to/workspace with your project directory. Add --write to the args only when the assistant should create, update, remove, or install repositories from LemmaBase (see Read-only vs write).

Stdio

Shared shape:

{
  "command": "lemma",
  "args": ["mcp", "--prefix", "/path/to/workspace"]
}

Older MCP clients use initialize / notifications/initialized (protocol 2025-11-25; later tools/list / tools/call need no per-request _meta). Clients on 2026-07-28 send _meta.io.modelcontextprotocol/protocolVersion on every request and may call server/discover first.

Streamable HTTP

Start a long-lived server (default http://127.0.0.1:8013/mcp):

lemma mcp --http --prefix /path/to/workspace

Point an HTTP MCP client at that URL. Example Cursor / Claude Desktop url entry:

{
  "mcpServers": {
    "lemma": {
      "url": "http://127.0.0.1:8013/mcp"
    }
  }
}

Optional flags: --host, --port (default 8013), --cors. No built-in auth or TLS; prefer localhost or a reverse proxy.

Claude

Claude Desktop: Settings → Developer → Edit Config. Add a server under mcpServers in claude_desktop_config.json, then fully restart Claude Desktop:

{
  "mcpServers": {
    "lemma": {
      "command": "lemma",
      "args": ["mcp", "--prefix", "/path/to/workspace"]
    }
  }
}

Claude Code: from a terminal:

claude mcp add --transport stdio lemma -- lemma mcp --prefix /path/to/workspace

Gemini CLI

Add the same mcpServers entry to ~/.gemini/settings.json, or to .gemini/settings.json in your project:

{
  "mcpServers": {
    "lemma": {
      "command": "lemma",
      "args": ["mcp", "--prefix", "/path/to/workspace"]
    }
  }
}

Or:

gemini mcp add --transport stdio lemma lemma mcp --prefix /path/to/workspace

Restart Gemini CLI (or /mcp reload) and confirm the server is connected.

Cursor

Cursor Settings → MCP, or a project .cursor/mcp.json / user MCP config. Add:

{
  "mcpServers": {
    "lemma": {
      "command": "lemma",
      "args": ["mcp", "--prefix", "/path/to/workspace"]
    }
  }
}

Enable the server in Cursor’s MCP UI if it is listed but off.

What you can do

Once connected, ask the assistant in natural language, for example:

Evaluate a spec with given inputs and explain how a result was reached
List specs in the workspace or show what data and rules a page defines
Check whether draft Lemma source is valid before you keep it
Install a repository from LemmaBase into the workspace (needs write access)

The server already ships authoring and evaluate guidance for the model. For learning the language yourself, see Learn and LLMs.txt.

Read-only vs write

By default the server is read-only: evaluate and inspect, no changes to the engine or disk.

Pass --write when you want the assistant to load or replace specs, remove them, clear the workspace load, or install (download a LemmaBase repository into lemma_deps/ and load). Only enable that for workspaces and agents you trust.

How to work

Install the CLI and set --prefix to your specs directory.
Connect Claude, Gemini CLI, or Cursor via stdio or Streamable HTTP as above (restart the client after config changes).
Ask questions against your specs; ask for explanations when you care how a rule fired.
For drafting new specs, enable --write and ask the assistant to validate, then update the workspace when you are ready.

MCP