A self-hosted AI agent platform you control
Griffin runs a small team of AI agents on your own machine. Agents are rows in a table, not code: instructions, a model, a set of tools, and a list of who may call them — with no SaaS control plane in the middle.
What you get
Agents as data
Create, edit and import agents from the web UI, the HTTP API or a JSON profile. Nothing about an agent needs a deploy.
Per-caller permissions
Each caller gets exactly the tools granted to it, enforced before tools reach the model; a caller with no grant gets nothing.
Any model
Claude Agent SDK, Cursor SDK, or any OpenAI-compatible endpoint — hosted or local — chosen per agent and per chat.
Delegation
An orchestrator delegates to specialist agents without blocking and reports back when subtasks finish.
Every interface
Web UI in English or Persian, terminal client, Bearer HTTP API, MCP endpoint, Telegram and Bale, scheduled jobs — or fully headless.
Rehearsal mode
Simulation chats run the whole agent tree but record side effects instead of executing them, with scenario evals.
Runs on one small box
Node 24 and SQLite, one docker compose up. A fresh install is a chat with one agent; optional tool packs switch on only when you configure what they need.
Self-hosted, open source, MIT
- Your conversations, prompts and credentials stay on your server.
- Login is local and keeps working when your SSO is down.
- MIT licensed, with no telemetry and no hosted dependency.