A Go harness cherry-picking the best ideas from every coding agent

What it does

This is a coding-agent harness written in Go that ships as a single binary with no runtime dependencies. It wraps an LLM tool-use loop—bash, file read/write, edit, and background subagents—inside an interactive terminal UI built on bubbletea. You point it at any OpenAI-compatible endpoint, and it discovers models live from the provider’s catalog instead of hard-coding them.

The interesting bit

The authors are openly magpie-ing: every feature in the roadmap cites the project it was borrowed from. Rather than replicating TypeScript promise gymnastics, the codebase uses Go channels and compiler-checked concurrency for file locks and background subagents. It also treats model discovery as a first-class feature, surfacing new entries automatically so you don’t have to hand-edit config when a provider ships a new weights file.

Key highlights

• One static binary, no runtime; config is readable and minimal.
• Live model-catalog discovery from providers (inference.net by default, OpenRouter, or any OpenAI-compatible endpoint).
• Built-in OAuth flow for ChatGPT/Codex subscriptions, with local token refresh and no keys in chat history.
• MCP servers auto-appear when you drop a .mcp.json in a repo.
• Parallel tool calls and streaming throughout; the agent loop is intentionally kept as a single function.

Verdict

Try this if you want a fast, local-first coding agent that ditches Node runtimes and auto-discovers models. Skip it if you need a polyglot or deeply customizable plugin ecosystem; the tool is Go-native and opinionated by design.