Tacit MCP server

Tacit

Your team's memory, wired into every AI tool.

Tacit syncs your team's GitHub, Slack and Notion into a time-aware knowledge graph that your team owns, and serves it to Claude, Cursor and Claude Code through a local MCP server. Ask "why did we do X, who decided, and where's the evidence?" and get an answer with links to the exact PRs, threads and pages.

• Local-first. Each teammate runs tacit on their own machine: sync, indexing and the MCP server all run locally. The only shared piece is your team's Neo4j (Aura Free or your own server). We host nothing and never see your data.
• Free by default. Search uses local embeddings and full-text, with no API keys. LLM extraction of people, decisions and time-aware facts is optional, uses your own key, and has a hard budget per run.
• Deterministic reads. search never calls an LLM. Your AI client writes the answer; Tacit returns the evidence.
• Every access logged. Every read and write goes through one function that logs it first, and refuses to run if it can't. tacit log answers "who looked at what".

 Each teammate's machine                                  Your team's Neo4j
┌──────────────────────────────────────────────┐         ┌─────────────────────────────┐
│ Claude / Cursor ──stdio──▶ tacit mcp         │         │ docs + chunks (search)      │
│                            (search, get_doc) │         │ graph facts (optional)      │
│ tacit sync ──▶ GitHub / Slack / Notion       │──Bolt──▶│ scope, cursors, write lease │
│   (your own tokens, only what's in scope)    │         │ access log                  │
└──────────────────────────────────────────────┘         └─────────────────────────────┘

Quickstart

uv tool install git+https://github.com/PhaseYon/tacit
tacit init                                # your team's Neo4j URI + password, your email
tacit auth github && tacit scope add github:acme/api
tacit sync
tacit mcp install --client claude         # or cursor / claude-code

Full walkthrough, including Slack, Notion and graph facts: docs/quickstart.md.

Trust model

Tacit is built for small, high-trust teams:

• One password per team graph. Whoever has it can read everything in the graph. Handing it out carefully is the team's job.
• Scope is the permission boundary. Only the repos, channels and pages you add are synced. Private Slack channels need an explicit --private; DMs are never synced.
• Access log. Every query is logged with who, what and when, and kept 90 days by default. Known gap: someone with the password can query Neo4j directly and bypass the log. Teams that need more than this need a hosted gateway (design doc §11), which is not built.

Commands

Command	What it does
tacit init	Connect to the team graph and prepare it (safe to re-run)
tacit auth github|slack|notion|anthropic|openai	Store a token (keychain, or a 0600 file)
tacit scope add|ls|rm	What the graph syncs
tacit sync [--extract] [--max-usd N] [--watch MIN]	Pull new and changed items
tacit search "…" [--as-of DATE] [--json]	Hybrid search with citations, no LLM
tacit get <url>	Full stored text of a cited item
tacit log [--who] [--since] [--grep]	Who read or wrote what
tacit forget <url>	Remove one item and its graph facts
tacit mcp / tacit mcp install	Local MCP server / client config
tacit doctor	Check config, secrets, database and tokens

How it works

• Sync polls each source with the user's own token, skips anything whose content hash hasn't changed, and writes under a single write lease so two laptops never ingest the same thing twice. Cursors live in the graph, so whoever syncs next picks up where the last person stopped.
• Search fuses vector search (local bge-small embeddings) and full-text search over chunks with reciprocal rank fusion. With extraction on, it adds Graphiti facts that know when they became true and when they were superseded.
• Details and trade-offs: docs/design/local-first.md.

Benchmark

52 "why" questions over the fictional Larkspur Labs company (demo/), local bge-small embeddings, k = 10. Retrieval costs $0; building the graph (124 items → 746 facts, 213 entities) cost $1.37 with Claude Haiku 5.5. Repeated runs give identical retrieval numbers.

arm	recall@10	evidence found	MRR
vector-only RAG	0.942	0.837	0.711
hybrid (RRF)	0.962	0.849	0.704
graph facts only	0.846	0.700	0.515
hybrid + graph facts	0.962	0.889	0.605

Caveats, in full in benchmark/README.md:

• Hybrid is no better than keyword-only here. The same author wrote the questions and the corpus, so they share wording; a held-out, paraphrased question set is still needed.
• Not comparable one-to-one with v0 (hybrid recall@10 0.865 on Postgres): v0 indexed every comment as its own document, while the new stack bundles a PR or issue with its discussion, so a top-10 hit is easier.
• The graph finds more evidence but ranks it lower. Adding facts raises the share of supporting evidence found (84% → 89% vs vector-only; 90% → 100% on "who decided" questions) but pushes the first right document down (MRR 0.71 → 0.61). The graph alone is worse than plain RAG. Fusion is untuned on purpose.

Development

uv sync
docker compose --profile test up -d --wait neo4j-test   # throwaway Neo4j on :7688
uv run pytest -q
uv run ruff check . && uv run ruff format --check .

docker compose up -d neo4j starts a persistent local Neo4j on :7687 for trying Tacit solo. The previous hosted architecture (Next.js, FastAPI, Celery, Postgres RLS, OAuth MCP) is preserved at tag v0-hosted.

License

Apache 2.0