Harbor 🚢

AI web chat agent framework — Chatwoot + LangGraph

Harbor gives any business an AI-powered chat widget that actually knows their business. Built on Chatwoot (MIT) as the platform layer, with LangGraph as the AI brain.

Self-hosted alternative to Intercom + AI. $0 platform cost. Your LLM, your data.

---

Architecture

Client website
    └── Chatwoot widget embed
            │
            ▼
    Chatwoot (self-hosted)     ← platform: widget, inbox, human handoff
            │ webhook
            ▼
    Harbor service             ← brain: LangGraph agent, persona, tools
            │
            ├── Local LLM (Qwen3-32B)   ← $0/call
            └── GPT-4o (Pro tier)       ← optional fallback

Rule: We deploy Chatwoot, we never modify its source. Harbor is a separate service.

---

Quick Start

# 1. Clone
git clone https://github.com/teamauresta/agent-harbor
cd agent-harbor

# 2. Install deps
pip install -r requirements.txt

# 3. Configure
cp .env.example .env
# Edit .env with your Chatwoot URL + token + LLM endpoint

# 4. Add a client persona
cp personas/examples/dental-demo.yaml personas/examples/my-client.yaml
# Edit the YAML with your client's details

# 5. Run
python main.py

---

Adding a Client

1. Create personas/examples/{client_id}.yaml (see examples)
2. In Chatwoot: create an account + inbox for the client
3. Set the inbox webhook URL to: https://your-harbor-url/webhook/{client_id}
4. Done — Harbor handles all conversations for that inbox

---

Tiers

Tier	Features
Starter	AI-only, 500 conversations/mo
Growth	+ Human escalation, proactive triggers, 1 integration
Pro	+ Multi-channel, 3 integrations, custom persona
Agency	White-label, unlimited, all integrations

---

Deploying Chatwoot (K3s)

See ~/k3s-manifests/chatwoot/README.md

Deploying Harbor (Railway)

railway login
railway init
railway up

Set environment variables in the Railway dashboard from .env.example.

---

Stack

• FastAPI — webhook server
• LangGraph — agent orchestration
• Chatwoot — chat platform (MIT, self-hosted)
• Qwen3-32B — default LLM (local, OpenAI-compatible)
• GPT-4o — optional Pro/Agency tier LLM

---

Built by SOTAStack