Show HN: Building an open-source agentic terminal (davehudson.io)

A developer has built and open‑sourced an "agentic terminal" inside Humbug (Apache 2.0, https://github.com/m6r-ai/humbug): a Python-based environment that lets both a human and an LLM interact with a terminal emulator, run shell commands, and inspect results. The codebase (~65k LOC) intentionally minimizes dependencies (PySide6, qasync, aiohttp, certifi) so LLMs and humans can easily reason about the implementation. The project was prototyped with Claude Sonnet 4 guiding design decisions and demonstrates how to integrate conversational AI with real shell control while keeping the system transparent and extensible. Technically, the design separates emulator core, GUI rendering, and the underlying Unix/Windows shell so the terminal has an in‑memory display and scrollback that can be exposed to an LLM or run headless. The conversation engine is asynchronous and provider‑agnostic (Anthropic, OpenAI, Mistral, Ollama, etc.), supports forking conversations between models, and cleanly abstracts LLM interfaces. A tool framework provides deterministic glue (keystroke injection, sandboxed filesystem access, clocks, GUI control) with a user‑mediated approval flow enforced at the tool level (not by the LLM). The result is a reproducible, inspectable agentic setup that reduces hallucination by making source and architecture available to models, while offering a safety model for running potentially dangerous operations.