Most coding agents send your code to someone else's server and hide the run behind a chat bubble. Small local models, on the other hand, tend to break tool calls and loop until the run is wasted.
betav2.0.1
PatchPilot
A coding agent for your terminal that runs on local models and asks before it touches anything.
PatchPilot works inside a repository with a model you run yourself: Ollama, LM Studio, MLX, llama.cpp or vLLM. Every file edit and shell command shows up in the transcript first, and the risky ones wait for your approval. Since 2.0 there are no cloud providers and no API keys.
A TypeScript terminal UI with nine focused tools, native function calling plus a validated fallback for smaller models, isolated child agents with narrow permissions, and a context meter that shows when the window is about to fill up.
Latest release
v2.0.1
- 2.0 removes Gemini, OpenRouter, NVIDIA and Codex and rebuilds the agent around local inference.
- Model discovery across every supported runtime, including a measured context window.
- 2.0.1 hardens session IDs and DOCX text extraction.
Developer setup
install
npm install -g @jx-grxf/patchpilotclone
git clone https://github.com/jx-grxf/PatchPilot.git && cd PatchPilotHighlights
- Local only: Ollama on your machine, your LAN or over Tailscale, plus OpenAI-compatible servers like LM Studio, MLX, llama.cpp and vLLM.
- Every file read and shell command lands in the transcript before it runs, with plan, build and bypass modes and a workspace boundary.
- Small models stay usable: tool calls are validated and repaired step by step when they break.
- Resumable sessions, pinned and compacted context, and child agents that work in their own context.