| name | lumen-mcp |
| description | Search, query, and explore the Lumen knowledge base via MCP tools. TRIGGER when user invokes /lumen or asks about their knowledge base, sources, concepts, or graph. |
Lumen Knowledge Base
You have access to a local Lumen knowledge base via MCP tools. All data is stored in SQLite on the user's machine.
Available MCP Tools
| Tool | When to use |
|---|
search | Find content by keyword — returns ranked chunks with snippets |
query | Ask a question — synthesizes an answer from chunks via LLM |
status | Quick stats — source count, chunk count, concept count, edges |
god_nodes | Find the most connected concepts (highest edge count) |
concept | Deep dive into one concept — summary, edges, sources |
path | Find how two concepts connect (shortest path) |
neighbors | Explore around a concept within N hops |
pagerank | Rank concepts by structural importance |
communities | List topic clusters detected by label propagation |
community | Members of a specific cluster by ID |
add | Ingest a new source (URL, file, arXiv, YouTube) |
Decision Guide
- "What do you know about X?" →
search first, then query if user wants synthesis
- "What are the main topics?" →
god_nodes or pagerank
- "How does X relate to Y?" →
path
- "Show me everything near X" →
neighbors
- "What clusters exist?" →
communities, then community for detail
- "Add this article" →
add
Gotchas
search returns raw chunks. query calls the LLM to synthesize — it costs API tokens.
community IDs are ephemeral — recomputed on each call. Don't cache them across sessions.
add deduplicates by content hash. Adding the same URL twice is safe.
path returns null if no path exists — concepts may be in disconnected components.
god_nodes and pagerank return empty arrays if lumen compile hasn't been run yet.