Self-hosted email platform with LLM-powered organisational intelligence.
┌──────────────┐ ┌─────────────┐ ┌──────────────┐
│ Stalwart │────►│ Mail Bridge │────►│ NATS │
│ Mail Server │ │ (webhook) │ │ JetStream │
└──────────────┘ └─────────────┘ └──────┬───────┘
│
┌─────────────────────┼─────────────────────┐
│ │ │
▼ ▼ ▼
┌─────────────┐ ┌──────────────┐ ┌──────────────┐
│ Ingestion │────►│ Classifier │────►│ Graph Writer │
│ Worker │ │ (LiteLLM) │ │ (AGE/Cypher) │
└─────────────┘ └──────────────┘ └──────┬───────┘
│
▼
┌──────────────┐
│ PostgreSQL │
│ + Apache AGE │
└──────┬───────┘
│
▼
┌──────────────┐
│ FastAPI │
│ Intel API │
└──────────────┘
Stalwart receives mail and fires a webhook to the Mail Bridge, which publishes normalised messages onto NATS JetStream. Three workers consume the stream in sequence: Ingestion parses and normalises, the Classifier extracts entities and intent via an LLM, and the Graph Writer persists classification output (never message bodies) into a PostgreSQL + Apache AGE graph. The FastAPI Intel API exposes graph queries and derived metrics.
git clone https://github.com/manwithacat/overmind.git
cd overmind
cp .env.example .env
# Edit .env to add your ANTHROPIC_API_KEY (or switch to ollama/mistral)
docker compose up --buildOnce all containers are healthy, the Intel API is available at http://localhost:8000.
People discovered from email headers and content:
curl http://localhost:8000/api/v1/graph/persons | python -m json.tool{
"persons": [
{"id": "p-0a1b", "name": "Alice Chen", "email": "alice@example.com", "message_count": 42},
{"id": "p-9c3d", "name": "Bob Martinez", "email": "bob@example.com", "message_count": 17}
]
}Attention cost — how much time each topic is consuming:
curl http://localhost:8000/api/v1/metrics/attention-cost | python -m json.tool{
"window_days": 30,
"topics": [
{"topic": "Q3 Budget Review", "thread_count": 8, "participant_count": 5, "estimated_hours": 3.2},
{"topic": "Office Move Logistics", "thread_count": 14, "participant_count": 9, "estimated_hours": 6.7}
]
}The default .env uses the Anthropic cloud API for convenience during proof-of-concept development. For production use with real organisational email, switch to a local model (e.g., ollama/mistral) to ensure no message content leaves your infrastructure. Set OVERMIND_LLM_PROVIDER in .env accordingly.
Message bodies are never stored in the intelligence layer — only classification output (entities, topics, intent).
Proof of concept — the full pipeline works end-to-end:
- Stalwart mail server receiving and storing mail
- Webhook-driven ingestion into NATS JetStream
- LLM classification (entity extraction, topic detection, intent)
- Graph persistence via Apache AGE with Cypher queries
- FastAPI intelligence API serving graph and metric endpoints
- CI pipeline (lint, typecheck, test, Docker build)
Planned:
- React analytics dashboard
- k-anonymity enforcement on derived metrics
- OIDC authentication for the API and dashboard
- Production hardening (rate limiting, alerting, backup automation)
The full specification lives in docs/:
| Document | What it covers |
|---|---|
| Overview | System summary and spec index |
| Architecture | Layers, message flow, NATS topology |
| Roadmap | Four-phase delivery plan |
| Privacy | UK GDPR, data minimisation, DPIA |
See CONTRIBUTING.md for guidelines on submitting patches and reporting issues.
Business Source License 1.1 — free for non-production use and production use by organisations with fewer than 50 employees. Converts to Apache 2.0 on the change date (2029-03-14).