A search index for your team's notes and documents that you run yourself.

What it does

1. A document comes in

   From a connector, an upload, or a call to the API.

2. It is cut into passages

   About 512 tokens each. Neighbours share 64, so a sentence is not lost at a cut.

3. Each passage is indexed twice

   By its words (top) and as a vector of 1,536 numbers (bottom).

4. Two rankings become one

   Reciprocal rank fusion. By default the vector ranking counts double.

Twelve questions and what came back

$ node bench/run.mjs --verbose

Asking instead of searching

Asking sends the best passages to a language model along with your question. These are two real replies to “how do we ship a release?”, from the same twelve notes.

Here is the most relevant knowledge I found:

[1] We cut releases every Thursday. Tag main, run the release workflow, then announce in the #engineering channel. Hotfixes can ship any day with a lead's approval.

[2] To roll back a bad deploy, re-run the release workflow with the previous git tag. Page the on-call engineer if customer traffic is affected.

[3] After any incident we run a blameless retro within three business days. The output is a written timeline, root cause, and action items with owners and due dates.

Sources:
[1] Release process
[2] Deploy rollback
[3] Incident retro

Tag main, run the release workflow, then announce in the #engineering channel.

Sources:
[1] Release process
[2] Deploy rollback
[3] Incident retro
[4] Expense policy
[5] Onboarding checklist
[6] Billing system
[7] Data retention
[8] Support SLA

Run it

1. Run npx companybrain. It clones the repository into ~/.companybrain/app and asks one thing: which model provider to use. Choose “Local only” to start without a key.
2. It builds two Docker images and starts three containers: Postgres with pgvector, the API on port 3333, and the dashboard on port 3000. Those two ports and 5432 need to be free.
3. Open http://localhost:3000. There is no sign-in. Paste some text, drop in Markdown files, or set up a source under Connections.
4. Stop it with npx companybrain down. Your data stays in a Docker volume until you run npx companybrain uninstall.

Installer 0.6.2, the current release, has a bug: once everything is running it prints error: webPort is not defined and exits. The stack is up regardless, so open the dashboard. The fix is merged and goes out with 0.6.3.

Where it reads from

There are 44 connectors in the repository. Every one is tested against recorded responses, not live accounts, so expect some rough edges with real data.

How you reach it

Several of these exist in the repository but have not been published to a package registry yet. This list says which.

Dashboard

Runs with the install

Search, ask, browse what is stored, and manage sources and providers, at port 3000.

REST API

Runs with the install

At port 3333, with an interactive reference at /docs and an OpenAPI 3.1 document at /v1/openapi.json.

MCP server

In the repository only

Four tools: search_memory, ask_memory, add_memory and list_spaces. It is not on npm yet, so the npx -y @companybrain/mcp config that the installer prints does not work. Clone the repository, run pnpm install, and point your client at the checkout:

{
  "mcpServers": {
    "companybrain": {
      "command": "pnpm",
      "args": [
        "--dir", "/path/to/companybrain",
        "--filter", "@companybrain/mcp",
        "exec", "tsx", "src/index.ts"
      ],
      "env": { "COMPANYBRAIN_API_URL": "http://localhost:3333" }
    }
  }
}

n8n node

On npm

Published as n8n-nodes-companybrain: search, ask and add.

Command line and SDKs

In the repository only

A cb command, a TypeScript SDK and a Python SDK. None is on npm or PyPI yet, so use them from a checkout.

Framework adapters

In the repository only

A LangChain retriever, Vercel AI SDK tools, and OpenAI function-calling tools.

Chrome extension

In the repository only

Saves a page or a selection. Load apps/extension as an unpacked extension.

Slack

Runs with the install

A /brain slash command is built into the API. You create the Slack app yourself; the Slack guide has the steps.

A larger test

What it does not do

• It does not understand meaning without a model. Out of the box, search matches words. That works when your question uses the words in the document, and fails when it does not.
• It does not know when it is wrong. A wrong passage comes back looking exactly like a right one, with a score of 1.00, as the last question in the table shows.
• It does not run itself. There is no hosted version. You install it, back up the database, and upgrade it.
• It does not sign in to your accounts for you. Sources take a token that you create and paste in. There is no OAuth sign-in flow.
• It is not finished. The installer is at version 0.6 and one person builds this. Things break. If something does, open an issue.