Documentation
OpenAgent is both an embeddable modular agent runtime and a complete standalone product. The standalone Server, Desktop App and CLI share sessions and whichever optional modules the product profile enables. Other products can import Core and supply their own identity, storage, tools and infrastructure.
It's model-agnostic: Claude, GLM, Ollama, LM Studio, vLLM, or any OpenAI-compatible provider. Swap the model without losing memory or tool state. Run multiple independent agents side-by-side, each in its own folder.
Start here
- Downloads — current canonical release artifacts
- Getting Started — install, first run and embedding
- Configuration reference
Core concepts
- Invitation System & Networking — Iroh P2P transport, coordinator, device certs, invite tickets
- Models — pick a provider
- MCP and capabilities — uniform tools, catalog modes and client sources
- Using connectors — unified marketplace, consent and local controls
- Shared components — Core, Tools, runner, MCP surfaces and product boundaries
- Composable applications — architecture requirements for multi-source data, interactive views, durable state and synchronization
- Gateway — WebSocket + REST surface over Iroh QUIC
- Channels — Telegram, Discord, WhatsApp, Webhook
- Events — inbound webhook triggers for workflows, tasks, and chats
- Memory & vault — the markdown-based Obsidian vault
- Scheduler & Dream Mode — recurring tasks and background work
Clients & operations
- Desktop App
- Architecture — component map and system flow
- Deployment
- Connector setup and recovery
- Connector release status — delivered code, verification and outstanding gates