Open source · Apache 2.0 · local-first
Your team's memory, wired into every AI tool.
Tacit syncs your team's GitHub, Slack and Notion into your own Neo4j and serves it to Claude and Cursor through a local MCP server. Ask “why did we do X, who decided, and where's the evidence?” and get answers with citations.
uv tool install git+https://github.com/PhaseYon/tacitNeeds Python 3.13 + uv. Then run tacit init. No binary to download.
Hands-on example
See what Tacit finds
Ask the history of Larkspur Labs, a made-up startup with three repos and a few decisions they later reversed.
Pick a question
Tacit will pull the evidence from Larkspur Labs' history.
Fictional company, so the links don't resolve. No LLM involved: this is what Tacit hands to Claude, which then writes the answer.
How it works
Runs on your laptops. Stores in your Neo4j.
One package per teammate, one database per team. Details in the design doc.
Only what's in your team's sync scope
…one per teammate
Your team's Neo4j
Aura Free · Docker on a VPS
The numbers
Fast, free to search, and it finds the evidence
- 96%
- of questions found the right source in the top 10 (50 of 52)
- 33 ms
- typical search, on a laptop
- $0
- to search: local embeddings, no API calls
- $0.007
- to turn a decision record into graph facts with Claude
Measured on 52 “why did we…” questions about a fictional company. How we measured →
Trust model
Simple on purpose
Tacit is for small, high-trust teams. Here's exactly what protects your data, and what doesn't.
One shared password per team graph
Whoever has your team's Neo4j URL and password can read the whole graph. Tacit is for small, high-trust teams; handing the password out carefully is your team's job. There are no per-user permissions inside a graph.
Sync scope is the boundary
Nothing outside tacit scope is ever fetched. GitHub repos and Notion pages must be listed explicitly. Slack is public channels only by default; private channels need --private and a warning. DMs are never synced. You can only sync what your own token can read.
Every query is logged, first
Every read and write goes through one function that writes an access entry (who, tool, params, when) before the query runs. If the log write fails, the query doesn't run. The LLM never writes Cypher.
Your keys stay on your laptop
Source tokens live in your OS keychain (or a 0600 file where there's none). Graph extraction is optional and uses your own LLM key. Tacit sends us no telemetry.
Read the log from any laptop on the team:
tacit log --who alice@ --since 7d --grep redisKnown gaps, stated plainly
- Anyone with the password can query Neo4j directly (Neo4j Browser, cypher-shell), skip the log, or delete log entries.
- The identity in the log is self-asserted (the email you gave
tacit init). - The access log is not tamper-proof. Closing that needs per-person database logins or a gateway, which is out of scope for now.
- If someone shouldn't see a channel or repo, keep it out of scope. There is no finer-grained control.
Quickstart
From zero to “why did we…?” in six commands
Create your team's Neo4j
A free Neo4j Aura instance is enough to start. Or run our Docker Compose file on a small VPS. Keep the URL and password.
Install and connect
Installs the tacit command, then asks for the URL, password and your email (used in the access log).
uv tool install git+https://github.com/PhaseYon/tacit && tacit initSign in to GitHub
Uses a fine-grained token, or reuses
gh auth token.tacit auth githubChoose what to sync
Scope is shared by the whole team and is the permission boundary.
tacit scope add github:owner/repoSync
Unchanged items are skipped by content hash, so re-runs are cheap.
tacit syncWire it into Claude
Also --client cursor or --client claude-code. Then ask: “why did we do X?”
tacit mcp install --client claude
Teammates: get the URL and password from whoever set it up, install Tacit, then run tacit init and tacit mcp install.