Overview
Mingbird is a local-first agent harness designed to help small 2–9B models finish real tasks on laptops with integrated graphics and 16–32 GB of RAM. It works with Ollama and emphasizes offline operation, transparent benchmarks, and reliable task completion rather than dependence on a large discrete GPU.
The project supports Windows 10 and 11, while Linux and macOS support is described as experimental. Its interface is available in English and Chinese.
Key Features
- Runs tests automatically and returns exact file locations and original errors to the model for self-debugging.
- Backs up every code change and provides a rollback command for safer editing.
- Uses task layering and escalating anti-loop controls to stop repeated tool demonstrations or unproductive loops.
- Keeps the factory tool prefill to 797 tokens and enforces that size in continuous integration.
- Handles larger output with an 8,192-character per-call cap and chunked skeleton-then-append feedback when needed.
- Rechecks the original task before reporting completion.
Local Operation and Safety
Mingbird includes a one-click offline mode and is described as zero-telemetry, allowing work to remain on the local machine. Configuration is not hardcoded: the Ollama address, executable, and GPU-related environment settings can be adjusted.
Its safety approach combines automatic backups, rollback support, crawl limits based on file size, and a five-ring protection system intended to intercept destructive actions. The crawl guard uses a byte budget of twice the file size, bounded between 64 KB and 512 KB, rather than simply refusing to inspect large files.
Benchmarks and Maintenance
The repository reports a 288-cell public benchmark comparing agent harness behavior and cites an improvement from 0.017 to 0.821 for the same 2B model across four harnesses. The listed release is version 1.9.1, with 472 tests and continuous-integration builds for Linux and macOS artifacts.