MCP server

@artano-ai/mcp-server exposes the cards corpus and the verification engine over the Model Context Protocol. Any MCP-aware runtime — Claude Code, Cursor, Codex, Goose — can drop it in.

Install

npx @artano-ai/mcp-server

That runs the server over stdio. You normally point an agent runtime at it via config rather than running it directly.

Register

{
  "mcpServers": {
    "lemma": {
      "command": "npx",
      "args": ["@artano-ai/mcp-server"]
    }
  }
}

Tools the server exposes

Tool	What it does
`cards_list`	List the curated cards. Optional `domain` substring filter.
`cards_get`	Fetch a full card record by `id`. Refuses to fabricate — an unknown id returns a structured error listing valid ids.
`ops_get`	Fetch an `OpsCard` rendered as Markdown for direct LLM consumption.
`hypothesis_crosscheck`	Run the cross-check engine on a `HypothesisCard` (by `id` or inline) — dimensional analysis, reference-corpus resolution, declared limit/conservation claims, `derivedFrom` resolution.
`usce_check`	Run USCE on a finished output — range-check its numeric values against a principle card’s `validationEnvelopes` (by `id`). Within → pass, outside → HIGH.
`rag_lookup`	Retrieve passages from a pgvector corpus over the SIESTA manual, ASE, pymatgen, numerical-methods notes, and SLURM / MareNostrum docs.

The server deliberately omits read_file / write_file / run_shell — every modern runtime already has those. Lemma adds the scientific layer on top.

Environment

By default the server reads the bundled corpus. Point it at a private fork with:

LEMMA_CARDS_DIR=/path/to/your/cards npx @artano-ai/mcp-server