qwen-harness
A privacy-focused, offline coding agent that runs entirely against a local Qwen model served by Ollama, with no network calls, accounts or API keys.
qwen-harness is pure TypeScript run directly by Node, with no dependencies to install, no build step and no registry package to fetch; the model, files and every transcript stay on the user's machine. It operates a permission-gated read, decide, act, observe loop with real tools for reading, searching, writing and editing files and running shell or PowerShell commands, running independent tool calls in parallel and splitting work across two parallel sub-agents shown in a live panel with each worker's task and token use.
File tools are confined to the workspace and are symlink-proof, the most dangerous commands are blocked by default, and full isolation is recommended by running it inside Docker, WSL or a sandbox wrapper. Argument validation and repair, recovery from malformed tool calls, and automatic context compaction are built in to keep small local models reliable, backed by 697 automated tests and a mandatory secret scan on every run.
It also supports remembering facts across chats, resuming any past conversation, streaming answers, and switching models on the fly.