Ultra-lightweight AI coding agent zerostack built in Rust unveiled

Core features of zerostack

Developed by Giuseppe Della Vedova, zerostack is an AI coding agent that runs in the terminal, aiming for extremely low resource usage through its Rust implementation. The binary size is 26MB, and average memory usage is around 16MB (24MB max). Compared to JavaScript-based agents (opencode, pi, etc.) that use 300MB on average and up to 700MB, it is more than 20 times more efficient.

Diverse functionality

It supports multiple LLM providers including OpenRouter as default, along with OpenAI, Anthropic, Gemini, and Ollama. It offers prompt modes (code, plan, review, debug, brainstorm, review-security, etc.) that change the system prompt to fit the nature of the task, and custom prompts can also be added. It features subagents that explore the codebase in parallel, MCP integration, session save/restore, and automatic compaction.

Permission management system

It provides 5-tier permission modes: restrictive (confirm all actions), readonly (read-only allowed), guarded (read allowed, write/bash requires confirmation), standard (default, allows operations in the current directory), and yolo (mostly allowed). Decisions approved during a session are kept in an allowlist, and a doom-loop detection feature that warns when the same action repeats 3 times helps prevent infinite loops.

Installation and license

It can be installed via an install script, cargo, or Homebrew. It is released under the GPL-3.0 license, allowing free use, modification, and distribution of the source, but derivative works must comply with the same license when distributed.