BrainStack

Shared second brain for humans and AI assistants. Self-hostable. MCP-native.

BrainStack keeps your notes as Markdown in Postgres. It exposes them to any MCP-compatible AI assistant: Claude Code, Claude Desktop and claude.ai, Cursor, Codex, Gemini CLI, Continue, Cline and others. People use the web app, which covers what Obsidian did: a file tree, a Markdown editor with wikilinks, full-text search, a graph, and import of the .md files you already have. Assistants read and write the same notes through the MCP server.

Every account has its own vault, and any folder can be shared with other people, read-only or with write access.

Status: pre-alpha. Built by Impulse Labs.

Why

If you work with several AI assistants, you keep retyping the same context everywhere: design docs in the repo, in Claude Projects, in Cursor rules, in Gemini, in your head. BrainStack is the single source of truth. Your assistants pull what they need through MCP and write back when you tell them to. One brain, every AI.

Related MCP server: big-brain

What's inside

apps/
  server/     Hono + tRPC + MCP server (Node 20+), with its own OAuth 2.1 provider for MCP clients
  web/        Next.js 15 web app: react-aria-components + CodeMirror 6
packages/
  core/       Markdown parsing, wikilink resolution, Postgres store and full-text search
  skill/      Canonical AI instructions + adapters for every major client

Storage is any Postgres. Sign-in is email and password, with optional Google sign-in and optional TOTP. MCP clients connect through OAuth or with an API key.

Quick start (self-host)

You need Docker. The compose file runs Postgres, the server, the web app and Caddy, which serves them all on one origin with automatic TLS.

git clone https://github.com/Impulse-Labs-SAS/brainstack.git
cd brainstack
cp .env.example .env
# edit .env: at least DOMAIN, PUBLIC_ORIGIN and POSTGRES_PASSWORD
docker compose up -d

Then open your PUBLIC_ORIGIN and create your account right away. The first account on an instance is its owner and needs no email verification. After it, sign-up is closed.

The schema is created on first start and updated by every new version on its own. Already have a Postgres? Set DATABASE_URL in .env and BrainStack uses it instead of the bundled one.

Email: optional for one person, needed to add others

BrainStack sends email for three things: verifying new accounts, password resets, and folder invitations. For a personal instance you can skip it. To bring anyone else in, configure one of these in .env, plus AUTH_EMAIL_FROM:

• SMTP, with any mail server (Gmail with an app password, Outlook, your company's relay, Resend's SMTP): SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASSWORD.

• Resend's API: RESEND_API_KEY.

Links are only ever sent by email, never written to the log. Without email:

• only the first account can sign up with a password;

• email invitations are refused, but link invitations still work;

• a forgotten password is reset on the server:

  docker compose exec server node dist/cli.js reset-password you@example.com

  It prints a new password and signs the account out everywhere.

The server logs a warning at startup when email is not configured.

Who can sign up

After the first account, sign-up is closed. People get in when one of these applies:

• you invite them to a shared folder by email;

• their address is in AUTHORIZED_EMAILS (comma-separated);

• OPEN_SIGNUP=true, which lets in anyone who can reach the server.

New accounts verify their address by email, so all three need email configured. Any account holder can invite by email, so everyone you let in can bring someone else in, into a folder they share. Keep that in mind before you add people you don't know.

Local development

pnpm install
cp apps/server/.env.example apps/server/.env   # set DATABASE_URL to any Postgres
pnpm lint
pnpm typecheck
pnpm test        # runs on PGlite, in process; no database needed
pnpm build
# server on :3000 (HTTP + MCP; stdio MCP when MCP_STDIO=true)
pnpm --filter @brainstack/server dev
# web app on :3001
pnpm --filter @brainstack/web dev

To try sign-up with a second account, or password resets, point SMTP at a local mail catcher such as Mailpit. apps/server/.env.example shows how.

Connecting an AI assistant

The web app's sidebar has Connect AI and Skill dialogs with this instance's MCP URL and copyable setup for each client. The MCP endpoint is https://<your-domain>/api/mcp.

Claude Desktop, claude.ai and other clients that support OAuth

Add https://<your-domain>/api/mcp as a connector (in Claude: Settings → Connectors). The client sends you to BrainStack to sign in and approve access.

Claude Code, Cursor and clients that take a token

Generate an API key in the web app (Settings → API keys), then:

claude mcp add brainstack --transport http https://<your-domain>/api/mcp \
  --header "Authorization: Bearer <api-key>"

Loading the skill

BrainStack ships canonical AI instructions for every major client. Build them with:

pnpm --filter @brainstack/skill build
# Outputs land in packages/skill/dist/<client>/...

Assistants can also fetch the same instructions at runtime through the get_brainstack_guide MCP tool, which helps clients without a dedicated skill format.

Architecture

Start with CLAUDE.md. It holds the stack, the workspaces, and the non-negotiable rules. It is written for AI assistants, but it binds everyone. Design notes for the larger subsystems (sharing and permissions, the graph, wikilinks and facets) are in docs/.

License

BrainStack is free software, released under the GNU Affero General Public License v3.0 (AGPL-3.0). You can use it, modify it and self-host it, for yourself or for your company, commercially or not. If you run a modified version as a service for other people, you must offer them its source code under the same license.

Contributing

Issues are open to everyone. Code contributions start with an issue, not a pull request; see CONTRIBUTING.md.

Built by Impulse Labs

BrainStack is built by Impulse Labs, a small team building AI-first tools. We use it ourselves every day.