gi-dellav/zerostack

Lightweight coding agent written in Rust, optimized for memory footprint and performance

What it solves

zerostack is a high-performance, minimal coding agent designed to automate software development tasks. It provides a terminal-based interface for developers to interact with LLMs to write, debug, and refactor code while maintaining a small resource footprint.

How it works

Written in Rust, the tool operates as a CLI agent that connects to various LLM providers (OpenRouter, OpenAI, Anthropic, Gemini, Ollama). It uses a system of interchangeable system prompts (e.g., code, plan, review, debug) to tailor agent behavior to specific tasks. It can execute bash commands and edit files using a set of standard coding tools, with a granular permission system to prevent unauthorized or destructive operations.

Who it’s for

Developers who want a lightweight, fast coding agent that integrates directly into their terminal and supports advanced workflows like git worktrees, parallel subagents, and long-horizon task loops.

Highlights

  • Low Resource Usage: Extremely small binary size (26MB) and RAM footprint (~16MB average).
  • Flexible Prompting: Built-in modes for brainstorming, planning, reviewing, and debugging that can be chained together.
  • Permission System: Five configurable modes (from restrictive to yolo) with doom-loop detection to prevent runaway agents.
  • Advanced Tooling: Support for MCP servers, Exa search, and integrated git worktree management.
  • Extensibility: Gated features for persistent Markdown-based memory, lifecycle hooks, and a strategic "Advisor" model.
  • Developer Experience: Includes a TUI with markdown rendering, an input queue for non-blocking interaction, and a /btw command for parallel side-questions.