About this project

ipsupport-code is a lightweight, self-learning local coding agent designed to operate with LM Studio or any OpenAI-compatible provider. Distributed as a single static Go binary, it drives local models through a reason-act-observe loop using core tools like file, run, git, web, and calc. A key feature is its ability to distill lessons from past runs to recover from errors and improve over time. The interface is a Claude-Code-style TUI built with Bubble Tea, offering live-streamed tool calls, syntax-highlighted diffs, and a toggle between plan (read-only) and auto (execution) modes. It supports multi-provider switching, sub-agent delegation to different models or external CLI agents, and background job processing. Security is managed through a permission policy, workspace jail, and optional OS-level sandboxing (Seatbelt/Landlock). Additional capabilities include MCP server integration, on-demand skill packs, goal-oriented multi-turn tasks with a judge loop, and session rewinding.