Jan Agent
Jan Agent is a standalone Rust workspace for running autonomous AI agents powered by local models. It decouples the agent system from the Jan Desktop (opens in a new tab) monorepo for faster builds, shorter dependency chains, and independent development.
What it does
- Serves local models via an OpenAI-compatible API using LlamaCPP
- Runs autonomous agents with a ReAct loop (read, write, edit, bash, web search tools)
- Wires to external agents like Claude Code, OpenClaw, or opencode with zero config
- Sandboxes tool execution using a WASM-based runtime with dual metering
- Downloads models from HuggingFace with interactive quantization selection
- Shares the Jan data folder — models installed via Jan Desktop are automatically available
Key features
| Feature | Description |
|---|---|
| Local-first | All inference runs on your hardware — no cloud, no API keys, no fees |
| OpenAI-compatible | localhost:6767/v1 works with any OpenAI-compatible client |
| LlamaCPP engine | Serves models via the LlamaCPP inference backend |
| HuggingFace integration | Download GGUF models directly from HuggingFace repos |
| WASM sandbox | Tool execution runs in a sandboxed WebAssembly environment |
| Agent TUI | Terminal UI with real-time streaming, resource monitoring, and tool output |
Quick links
- Installation — Build from source
- Quickstart — Get running in under a minute
- Core Concept — How the pluggable trait system works
- Architecture — Crate layout, dependency graph, embedding API
- CLI Reference — Full command reference
- Development — Contributing and building