Local AI harness for coding agents, tuned for AMD Strix Halo

Run a coding agent on whatever you have. It works with any OpenAI-compatible engine (llama.cpp, vLLM, Ollama, LM Studio, Strata) or a cloud model. On a Ryzen AI Max+ 395 it runs a 125B Qwen3.8 Flash-Next on the iGPU with 262k context, while NPU models screen and route prompts and a sidecar writes commits and summaries. Built on pi.

Source on GitHub · Releases · Windows setup (beta) · Local vs cloud comparison

What you get

• One-command, idempotent setup that probes each piece instead of trusting it.
• Crash recovery: relaunch and say "continue".
• NPU side jobs: prompt triage in about 120 ms and near-duplicate file scans.
• Token-saving extensions and symbol-precise code tools.
• Logged benchmarks for engines and models.

Install steps, the full stack and the evidence are in the README.