Without a docs server

• LLMs hallucinate API signatures that don't exist
• Entire files dumped into context — 3,000–15,000 tokens per doc
• Architecture decisions buried across dozens of files
• Every repeated lookup pays full context cost

With Gnosis MCP

• search_docs returns ranked, highlighted excerpts — typically 300–800 tokens
• Real answers grounded in your actual docs, not guesses from training data
• One local index across hundreds of files — instant multi-doc search
• 5–10× token savings per lookup when your corpus covers the question

What makes gnosis-mcp different

• Your data stays on your machine. SQLite by default, PostgreSQL at scale — nothing leaves the host.
• Index anything that's docs-shaped. Markdown, git commit history, crawled websites — one index, one search API.
• Measured, not marketed. Ships BEIR SciFact numbers (0.671 nDCG@10 — within 1 % of the Lucene BM25 baseline), a reproducible eval harness (gnosis-mcp eval), and a chunk-size sweep showing where the quality plateau actually sits.

Full side-by-side vs Context7 / docs-mcp-server / mcp-local-rag: gnosismcp.com#compare.

Features

• Zero config — SQLite by default, pip install and go
• Hybrid search — keyword (BM25) + semantic (local ONNX embeddings, no API key). Tune RRF fusion with GNOSIS_MCP_RRF_K.
• Cross-encoder reranking — optional [reranking] extra with a 22M-param ONNX model. Off by default. Test on your own corpus before enabling — the bundled MS-MARCO reranker hurts dev-doc retrieval in our measurements.
• Git history — ingest commit messages as searchable context (ingest-git)
• Web crawl — ingest documentation from any website via sitemap or link crawl
• Multi-format — .md .txt .ipynb .toml .csv .json + optional .rst .pdf
• Auto-linking — relates_to frontmatter creates a navigable document graph
• Watch mode — auto-re-ingest on file changes
• Prune stale docs — gnosis-mcp ingest --prune removes chunks whose source file was deleted. --wipe for a full reset before re-ingest.
• Built-in eval harness — gnosis-mcp eval prints Hit@K / MRR / Precision@K in one command, against a bundled fixed fixture set
• PostgreSQL ready — pgvector + tsvector when you need scale

Performance

Fast. 8.7 ms mean MCP round-trip. Hybrid search p50 < 30 ms on a 700-doc corpus. Keyword QPS scales from 9,463 @ 100 docs to 471 @ 10,000 docs (full numbers).

Finds the right answer. On 558 real dev docs with 25 hand-written golden queries: Hit@5 = 0.92, nDCG@10 = 0.87, MRR = 0.79. On BEIR SciFact (5,183 docs, public retrieval benchmark): nDCG@10 = 0.671 — within 1 % of the Lucene BM25 baseline.

Tokens saved. Each search_docs call returns 200–500 tokens of on-point snippets instead of the 3,000–15,000 tokens a full-file Read would have cost. Track your own with gnosis-mcp savings (v0.12.0+) — the ledger writes to search_access_log on every call and aggregates per tool per --days N:

$ gnosis-mcp savings --days 7
  Tool calls:               142
  Tokens returned:        7,104
  Tokens baseline:      231,580
  Tokens saved:         224,476
  Ratio:                   32.6×

Typical compression runs 10–60× depending on corpus coverage and query specificity — verify on yours. access_log is on by default; GNOSIS_MCP_ACCESS_LOG=false opts out.

The same ledger answers the other question — what is this corpus being asked, and where does it come up empty — with gnosis-mcp usage (v0.17.4+): calls, the queries that matched nothing, which documents are served most, which have never been served, and which clients are doing the asking.

Reproducible. gnosis-mcp eval runs a bundled retrieval-quality harness locally in one second — but it ingests nine hardcoded sample documents into a temporary database and answers ten bundled queries, so it returns the same numbers for every corpus and is a smoke test, not a measurement of your docs. To score your own corpus, run python tests/bench/bench_real_corpus.py --corpus <docs-root> --golden <golden.jsonl> (the numbers above come from tests/bench/golden-knowledge.jsonl). tests/bench/*.py reproduce every number. Methodology: docs/benchmarks.md.

Rerankers stay off by default. The bundled MS-MARCO cross-encoder drops nDCG@10 by 27 points on dev-docs and adds 400× latency; BGE-reranker-v2-m3 drops it 31 points at 2400×. Test on your corpus before enabling — full write-up: bench-experiments-2026-04-18.

Quick Start

pip install gnosis-mcp           # or: uv tool install gnosis-mcp
gnosis-mcp ingest ./docs/        # loads docs into SQLite (auto-created)
gnosis-mcp serve                 # starts MCP server

That's it. Your AI agent can now search your docs.

Connect your client — see llms-install.md for copy-paste JSON snippets for Claude Code, Claude Desktop, Cursor, Zed, opencode, Windsurf, VS Code, JetBrains, Cline, and any other MCP client.

Re-organized your docs? gnosis-mcp ingest ./docs --prune re-ingests and removes any DB chunk whose source file no longer exists. --wipe resets the entire index first. Or run gnosis-mcp prune ./docs --dry-run to preview what would be deleted. Pruning only touches documents this root is responsible for: crawled URLs and generated documents (git history) are left alone unless --include-crawled / --include-generated says otherwise.

Want semantic search? Add local embeddings — no API key needed:

pip install gnosis-mcp[embeddings]
gnosis-mcp ingest ./docs/ --embed   # ingest + embed in one step
gnosis-mcp serve                    # hybrid search auto-activated

Test it before connecting to an editor:

gnosis-mcp check                                # FTS5 + schema + row counts; non-zero exit = stop
gnosis-mcp search "getting started"             # keyword search
gnosis-mcp search "how does auth work" --embed  # hybrid semantic+keyword
gnosis-mcp stats                                # see what was indexed

Then wire it into the clients you actually use — one command, no paths to edit:

gnosis-mcp setup                # preview what each client's config would become
gnosis-mcp setup --write        # Claude Code, DeepSeek Harness, Codex, Cursor, VS Code, …
gnosis-mcp doctor               # is it wired, and has anything called it yet?

setup resolves the command path for the machine it runs on instead of the one the README assumed, and each block it installs is marker-delimited so re-running rewrites it in place. Where a client does not read the server's own MCP instructions, it also installs the short rule that gives the agent a reason to prefer gnosis over reading files — a mounted server nobody calls is the silent failure mode. doctor is the check for that: it reads the access log and tells you whether a client has really called the server. Details: docs/cli.md.

gnosis-mcp check is the gate: it exits 0 only when the backend started and the schema the server needs is present, and it names whatever is missing before exiting 1. Keyword search needs SQLite FTS5, which is compiled into your Python's SQLite rather than guaranteed by it — check reports it as FTS5: ready. Anything else: docs/troubleshooting.md.

# Serve your ./docs on http://localhost:8000 — MCP at /mcp, REST at /api/*
docker run -p 8000:8000 \
  -v "$PWD/docs:/docs:ro" -v gnosis-data:/data \
  ghcr.io/nicholasglazer/gnosis-mcp:latest