cortex

How It Works

1. Capture - send a Telegram message, write in the web editor, or add a thought via MCP from any AI tool.

2. Classify - an LLM sorts it into one of five categories (People, Projects, Tasks, Ideas, Reference) and extracts structured fields.

3. Store - PostgreSQL with pgvector embeddings. Search by meaning, not just keywords.

4. Access - web dashboard, daily/weekly email digests, or query your brain from any MCP-compatible tool.

Related MCP server: Memoria

When crtxboard fits (and when it doesn't)

Fits: low-friction capture of individual thoughts from anywhere — voice from a phone, text on the go, or a quick note at the desk. Personal logistics (tasks, follow-ups, appointments, people context, idea backlog). Agent-assisted recall via MCP. Self-hosted with local LLM support.

Doesn't fit: long-form writing, progressive-summarization knowledge work, cross-linked knowledge graphs, team wikis, curated reference libraries. For those, Obsidian, Notion, or Logseq are purpose-built — crtxboard is deliberately a different shape.

Features

• Capture - Telegram bot with text and voice (faster-whisper), web dashboard with quick capture and full editor, MCP server

• Intelligence - LLM classification into 5 categories with confidence scoring, context-aware (uses recent + similar entries), inline Telegram buttons for low-confidence entries, /fix to reclassify, automatic task completion detection

• Search - semantic search via pgvector + qwen3-embedding with text fallback, peer-dropdown filters (category, tag, status, activity), multilingual (EN/DE)

• Google Calendar - automatic event creation from classified entries, multi-calendar support with LLM-based routing

• Digests - daily briefing and weekly review, delivered by email and shown on the dashboard

• Display - PNG endpoint for e-ink devices (e.g. TRMNL) showing today's calendar, pending tasks, and weather

• Self-hosted - local embeddings, local voice transcription, LLM-agnostic (Anthropic, OpenAI, or any compatible endpoint), 7 MCP tools

How to use

Capture / correction / search / MCP workflows with examples: USAGE.md.

Quick Start

git clone https://github.com/moguls753/cortex.git && cd cortex

Create a .env file with a database password:

echo "POSTGRES_PASSWORD=changeme" > .env

Start everything:

docker compose up -d

This boots PostgreSQL, Ollama (embeddings), Whisper (voice transcription), and the app. The embedding model is pulled automatically on first start.

Open http://localhost:3000 and the setup wizard will walk you through:

1. Account - create your login credentials

2. LLM - pick a provider (Anthropic, OpenAI, Groq, Gemini, LM Studio, Ollama) and enter your API key

3. Telegram - optionally connect a Telegram bot for capture

4. Done - start using crtxboard

Everything is reconfigurable later from the Settings page.

MCP

Add to your Claude Code config (~/.claude.json):

{
  "mcpServers": {
    "cortex": {
      "command": "docker",
      "args": ["exec", "-i", "cortex-app-1", "node", "dist/mcp.js"]
    }
  }
}

Tools: search_brain · add_thought · list_recent · get_entry · update_entry · delete_entry · brain_stats

Resource Requirements

Minimum: ~4-5 GB RAM with a cloud LLM provider (Anthropic, OpenAI).

If you run classification and digests through a local LLM via Ollama instead of a cloud provider, add the RAM for that model on top. Recommended minimum: Qwen 2.5 7B (~5 GB) or Llama 3.1 8B (~5 GB) — smaller models tend to struggle with reliable structured output. 10 GB+ RAM recommended for fully local setups.

Configuration

All configuration is done through the setup wizard and the Settings page. No .env editing required beyond POSTGRES_PASSWORD.

If you prefer env vars, they serve as defaults and are overridden by settings saved in the database:

Development

Stack: Node.js, TypeScript, Hono, PostgreSQL + pgvector + Drizzle ORM, Ollama, grammY, Vitest, Tailwind CSS, Docker Compose.

npm install && npm run dev    # local dev
npm test                      # all tests
npm run test:unit             # fast, no Docker
npm run test:integration      # needs Docker (testcontainers)

Architecture, schema, and prompt contracts are documented in ARCHITECTURE.md. Spec artifacts in docs/specs/.

License

AGPL-3.0