Tailpipe

Every day you think out loud with AI — decisions, problems, things you figured out, things you're still chewing on. That's your most valuable thinking, and it comes out the tailpipe of your conversations as exhaust: scattered across ChatGPT, Claude, and Gemini, locked in vendor silos, unowned and unsearchable.

Tailpipe captures that exhaust and turns it into memory you own — a private, local, queryable second brain that any AI agent can plug into.

• Capture every conversation across providers.

• Own it — full-fidelity transcripts on hardware you control. Nothing leaves the house.

• Extract the actual knowledge — decisions, problems, solutions, lessons — from the raw chat.

• Connect it into a typed, dated knowledge graph of your entities and how they relate.

• Serve it over MCP, so any agent, any vendor, can query your whole history.

• See it — an interactive 3D graph of your own mind.

Not a product you rent. A system you run.

How it works

  ┌── Capture ──────────────┐     ┌── Memory Core (yours) ──────────────┐
  │  browser extension      │     │                                     │
  │  Claude · ChatGPT ·      │──▶  │   raw transcripts (kept forever)    │
  │  Gemini normalizers      │     │   FTS + vector search               │
  │  local code sessions     │     │   MCP server ◀── any AI agent       │
  └─────────────────────────┘     │        │                            │
                                  │        ▼                            │
                                  │   Extract → mentions & assertions   │
                                  │        │                            │
                                  │        ▼                            │
                                  │   Knowledge graph (typed, dated)    │
                                  │   + 3D visualization                │
                                  └─────────────────────────────────────┘

Everything runs on your own box — a NAS, a mini-PC, a spare machine. The only outbound calls are the ones you choose (e.g. a frontier model for extraction). Your memory never leaves.

Related MCP server: Rapport MCP Server

Runs anywhere, with the models you already have

You don't need a NAS. The memory core is just a Docker container — run it on a NAS, a mini-PC, a spare laptop, or a cloud VM. Wherever Docker runs, Tailpipe runs.

Two models are involved, and you control both:

• Embeddings (for search) — a small local model, BAAI/bge-small-en-v1.5 (~130 MB), auto-pulls on first run via fastembed. It's the only model Tailpipe downloads for you; runs on CPU, nothing to configure.

• Extraction (chat → knowledge) — an LLM you pick. Tailpipe doesn't ship one. Point it at:

  • Ollama or any OpenAI-compatible local server (EXTRACT_ENGINE=openai_compat),

  • a frontier API like Anthropic (EXTRACT_ENGINE=anthropic), or

  • an edge box.

  Most people already have Ollama running or an API key — either works out of the box.

Coming soon: a purpose-trained extraction model, dropping in the next few weeks — a small model tuned for exactly this pipeline, so extraction runs better and cheaper than a general LLM. Swap it in with one env change.

Repository layout

Quick start

Full setup lives in docs/SETUP.md. The short version:

1. Copy .env.example → .env; set INGEST_TOKEN and your extraction engine. Never commit .env.

2. docker compose up -d --build — brings up the memory core + MCP.

3. Load the browser extension (capture/extension/) and/or run the normalizers on your exports to feed conversations in.

4. Run extraction + graph build (extract/, graph/) to turn raw chat into a knowledge graph.

5. Point any MCP-capable agent at your server and ask it about your own history.

Privacy & the MCP surface

Tailpipe is local-first by design. Your transcripts, your extracted knowledge, and your graph live on your hardware, served only under your own bearer token. The extraction API is the one leash you keep on purpose — everything else runs on your box.

MCP is local-network only, for now — on purpose. Tailpipe currently serves its MCP tools over your LAN, for local agents (Claude Desktop, Claude Code, etc.). Exposing them to web-based agents — which dial in from a vendor's servers, not your machine — is on the roadmap but deliberately not shipped yet. Doing it safely requires a public endpoint, an OAuth 2.1 gateway, a read-only remote surface, and an audit log; get it wrong and you leak your entire memory to the internet. Local-first until that's airtight.

License

AGPL-3.0. Use it, run it, modify it — but if you offer it as a service, your changes stay open too. Commercial licensing available separately.