Show HN: Bash is All You Need for a language model REPL
A developer built a language model REPL agent using only standard Unix command-line tools like Bash, jq, and curl, with zero heavy dependencies. The system relies on transparent file-based state, piped text streams, and Unix exit codes for control flow, challenging the necessity of complex AI frameworks.
Background
- The author built a minimal "agent" (an AI assistant that can take actions) using only standard Unix command-line tools: bash, curl, jq, and a local LLM server. No Python, no npm, no heavy ML frameworks like LangChain or LlamaIndex.
- Key context: Most current LLM "agent" frameworks (LangChain, AutoGPT, CrewAI) are Python-heavy, dependency-laden projects that abstract away the underlying API calls and orchestration logic. This project strips that away to show the core loop is simple: send a prompt, get a response, check exit codes, repeat.
- The provocative title references a viral paper "Attention Is All You Need" (which introduced the Transformer architecture behind modern LLMs) and the popular "X is All You Need" meme in ML. It's making the point that you don't need complex agent frameworks—just bash piping.
- The project demonstrates a REPL (Read-Eval-Print Loop) where the agent's "memory" is literally just a .jsonl file you can truncate with `head`, and tool execution is handled by checking Unix exit codes in a while loop.