Statistical Genomics Foundry
Design docs design record

knowledgebase-mcp

Agent-facing knowledge-retrieval MCP servers — their framing, strengths, and where they sit on our values.

A cluster of independently-maintained, similarly-named agent-facing knowledge-retrieval MCP servers — no single canonical repo. Reviewed: jeanibarz/knowledge-base-mcp-server (exact name; FAISS), Geeksfino/kb-mcp-server (most stars; txtai, portable .tar.gz), olafgeibig/knowledge-mcp (LightRAG). None is genomics-specific. Verified 2026-06-26 against the three repos.

Its framing

Each is a knowledge-retrieval server: point it at a folder of documents (or a .tar.gz archive), it chunks and embeds them into a vector store (FAISS / txtai / LightRAG), and exposes MCP tool calls — retrieve_knowledge, ask_knowledge, semantic or graph search — so an LLM client (Claude Desktop, Cursor, Cline) queries the corpus mid-reasoning. The source files stay as readable text on disk; some ship an operator CLI for build / search / curate. The product surface is a retrieval API for an agent, not a destination a human reads.

Strengths relative to our approach

  • Transparent, git-friendly storage — source docs stay plain-readable, "nothing is obfuscated"; the index is derived.
  • Citation-grounded answers — ask_knowledge returns answers with citations; chunk metadata carries source paths and line ranges.
  • Corpus portability — entire knowledge bases save as compressed archives, shareable and loadable in one command (runnable via uvx, no install).
  • Hybrid vector + graph retrieval (txtai, LightRAG) — beyond naive RAG.
  • Operator ergonomics — diagnostics (kb doctor), retrieval testing (kb eval), guarded mutation paths.

Where it sits on our values

Value knowledgebase-mcp
Produces skills No — a server you run; the output is a queryable index. A .tar.gz is a portable corpus, not a self-contained skill.
Progressive disclosure No — knowledge exists to feed an agent via retrieval; where a human surface exists at all it is an operator/curator CLI, not a reader's layered learning surface.
Traceability Yes, at the storage and citation layer — source files stay plain text and git-friendly, and answers carry citations with source paths and line ranges. Strong here — but operator-facing, not a reader's foregrounded surface.
CLI instructions derived from CLI docs N/A — the CLI is an operator console for the index, not derived CLI-tool invocation.
Portability Partial — bound to MCP as the agent interface (many clients, one protocol); the corpus is portable (.tar.gz), but there is no runtime-agnostic skill artifact.
Human scrutiny Partial — knowledge is stored inspectably as files (operator-inspectable), but not foregrounded for a human to read and learn; no skills are produced.
Knowledge-base backed Yes — it is a knowledge base, but an agent-facing retrieval one, not a human-foregrounded reading surface.

Where it diverges from the Foundry

knowledgebase-mcp is the canonical instance of the pattern the Foundry keeps observing: agent-facing retrieval with the human-reading surface absent. It is a substrate an agent queries (a RAG / retrieval backend), not a destination a human learns from; it builds no portable skill artifacts and does no validity refereeing. The relationship is complement, not competition — one of these (most plausibly the .tar.gz-portable Geeksfino server) could be a backend component for exposing Foundry content to agents via retrieval, orthogonal to and downstream of the Foundry's differentiator: a navigable human knowledge base cast into portable, provenance-stamped skills.

Sources