MCP server¶

← Home

lit-monitor ships a Model Context Protocol server so any MCP-capable agent — Claude Desktop, Claude Code, Cursor, Continue — can query your knowledge graph and vector index directly. It's the same query layer the HTTP API exposes, surfaced as MCP tools.

Starting the server¶

lit-monitor mcp serve

stdio transport. Register it in your MCP client config:

"lit-monitor-graph": {
  "command": "lit-monitor",
  "args": ["mcp", "serve"]
}

The server reads the same local state DB, ChromaDB store, and KuzuDB graph as the rest of lit-monitor — nothing leaves your machine, and it's read-only by design.

The twelve tools¶

Tool	What it does
`find_papers_by_entity`	Find papers mentioning an entity (alias-resolved).
`find_papers_by_relationship`	Find papers participating in a typed relationship (`CITES`, `EXTENDS`, `CONTRADICTS`, …).
`get_paper_details`	Full snapshot of one paper — metadata, entities, and graph edges.
`get_corpus_stats`	Aggregate paper / entity / edge counts for the whole corpus.
`list_entities_by_type`	List entities of a given type, ranked by mention count.
`get_schema`	Markdown-formatted description of the graph schema (for prompt grounding).
`find_papers_by_query`	Free-text query resolved through the graph entity alias chain.
`find_papers_by_query_hybrid`	Free-text query via RRF-fused graph + vector retrieval (the strongest default).
`semantic_search`	Pure vector retrieval over the ChromaDB index — paper-level or chunk-level hits.
`run_cypher`	Execute a read-only, safety-guarded Cypher query against the graph.
`get_recent_discovery_runs`	The most recent discovery runs, newest first.
`get_discovery_run_papers`	The ranked paper results for one discovery run, score-descending.

The Cypher escape hatch¶

run_cypher lets an agent run arbitrary graph queries, but every statement passes through a safety guard that rejects anything mutating — only read paths (MATCH / RETURN and friends) are allowed, write clauses (CREATE, DELETE, SET, MERGE, …) are blocked, and a LIMIT is appended when absent. Pair it with get_schema so the model knows the node labels, properties, and the closed-vocabulary predicates before it composes a query.

Zotero hand-off¶

get_paper_details returns each paper's metadata with a zotero_key and a derived zotero_deeplink (or null when the paper has no linked Zotero item), so a paper found through lit-monitor can be handed straight to a Zotero MCP client without a title round-trip.

To see the live tool registry with full input schemas, run lit-monitor mcp serve and inspect the tools/list response from your client.