AI Brain

A self-hosted, web-first Markdown knowledge base that doubles as an AI brain. Documents and the links between them are stored in Postgres and exposed over both a REST API and an MCP server, so Claude, Codex, and other tooling can read, write, search, and traverse your notes. A drop-in replacement for Obsidian, built AI-native from the ground up.

Stack

Related MCP server: Notemd MCP Server

Repository layout

apps/
  web/        Next.js app — UI, REST API, Auth.js
  mcp/        MCP server (added in Phase 6)
packages/
  core/       Domain/service layer (documents, links, search, embeddings, auth)
  db/         Drizzle schema, migrations, Postgres + pgvector client
infra/
  postgres/   First-boot init (enables the vector extension)

Deploy with Docker Compose

The whole stack runs from one image:

cp .env.example .env        # set AUTH_SECRET (openssl rand -base64 32)
docker compose up -d --build
# web → http://localhost:3002 · MCP → http://localhost:8787/mcp

docker compose brings up db (Postgres+pgvector), a one-shot migrate, then web, mcp, and worker (async embedding + trash purge). The first registered user becomes the admin. For a dev mail catcher (captures password-reset/verification emails, UI on :1080): docker compose --profile mail up -d and set SMTP_HOST=maildev, SMTP_PORT=1025.

Local development

pnpm install
cp .env.example .env         # edit AUTH_SECRET etc.
docker compose up -d db      # just Postgres
pnpm db:migrate              # apply the schema
pnpm dev                     # web :3002, mcp, and worker (turbo)

Connect Claude / Codex (MCP)

Generate a Personal Access Token at /settings/tokens, then point your client at the MCP server. Tools exposed: search_documents, list_documents, get_document, create_document, update_document, delete_document, get_backlinks, list_links, plus a brain://documents/{id} resource.

Local (stdio) — e.g. Claude Code:

claude mcp add ai-brain \
  --env AI_BRAIN_TOKEN=<your-PAT> \
  -- node --import tsx /absolute/path/to/ai-brain/apps/mcp/src/stdio.ts

Remote (Streamable HTTP) — run pnpm --filter @ai-brain/mcp start:http (defaults to :8787), then:

claude mcp add --transport http ai-brain http://localhost:8787/mcp \
  -H "Authorization: Bearer <your-PAT>"

REST API

All endpoints accept Authorization: Bearer <PAT> (or a session cookie):

Build phases — all complete ✅

1. Scaffold — monorepo, Next.js + Tailwind, Postgres + pgvector, base schema.

2. Auth — login + Personal Access Tokens.

3. Documents — CRUD, Markdown parsing, editor.

4. Links — [[wiki-links]] + backlinks.

5. Search — full-text + semantic (hybrid).

6. MCP — stdio + HTTP server over the service layer.

Each phase has a verify-phaseN*.mts script under packages/core/scripts (or apps/mcp/scripts) demonstrating it end-to-end.

Production-readiness (added)

• Async embedding via a Postgres-backed job queue + worker; real index-status badges.

• Version history (coalesced snapshots) + soft-delete/Trash with retention purge.

• Markdown import/export (zip) for spaces and documents.

• Account hardening: password reset, email verification, admin + registration modes.

• Full Docker Compose stack (db + migrate + web + mcp + worker).

Not yet done (follow-ups)

• Graph view, folders/workspaces, SSO, real-time collaboration, attachments/images.