Zerostack: 8MB Coding Agent in Pure Rust
HN front page yesterday at 529 points, 293 comments. Zerostack is a coding agent written entirely in Rust by a developer who runs out of memory when he opens two instances of opencode on his old laptop. The numbers: roughly 8MB RAM idle, 12MB while actively working a task, 8.9MB binary on disk. The reference point everyone in the thread used was Claude Code chewing multiple gigabytes.
Around 7,000 lines of Rust, GPL-3.0. Aggressive LTO, smallvec and compactstring crates to keep allocations on the stack, size-optimized link flags. The plumbing you would expect is all there. OpenRouter, OpenAI, Anthropic, Gemini, Ollama as model providers. MCP and ACP protocol support. Terminal UI with markdown. Git worktrees. Four permission modes from restrictive (every tool action prompts) up to standard (safe commands auto-approved). Iterative coding loop where the agent picks an item from the plan, works on it, runs tests, updates the plan, loops.
The interesting design choice is in the prompting layer. Zerostack does not implement Anthropic Skills. It uses a prompt library stored as markdown files instead. The author says the goal is a complete suite of prompts that can fully substitute Skills or obra/superpowers. Simpler, more portable, less framework lock-in. Whether that bet ages well depends on whether Skills becomes a dominant standard or just one of many.
The honest take. For most developers the Rust-vs-JS performance gap is going to be invisible. Agent loops are I/O bound on LLM latency. But for the laptop-poor and the principle-driven, this is a clean alternative to an Electron-stuffed 300MB dev tool. The fact that 529 HN points showed up the day after release tells you the market for sub-megabyte tooling is real. v1.1.0 shipped May 17.
https://github.com/gi-dellav/zerostack
← Back to all articles
Around 7,000 lines of Rust, GPL-3.0. Aggressive LTO, smallvec and compactstring crates to keep allocations on the stack, size-optimized link flags. The plumbing you would expect is all there. OpenRouter, OpenAI, Anthropic, Gemini, Ollama as model providers. MCP and ACP protocol support. Terminal UI with markdown. Git worktrees. Four permission modes from restrictive (every tool action prompts) up to standard (safe commands auto-approved). Iterative coding loop where the agent picks an item from the plan, works on it, runs tests, updates the plan, loops.
The interesting design choice is in the prompting layer. Zerostack does not implement Anthropic Skills. It uses a prompt library stored as markdown files instead. The author says the goal is a complete suite of prompts that can fully substitute Skills or obra/superpowers. Simpler, more portable, less framework lock-in. Whether that bet ages well depends on whether Skills becomes a dominant standard or just one of many.
The honest take. For most developers the Rust-vs-JS performance gap is going to be invisible. Agent loops are I/O bound on LLM latency. But for the laptop-poor and the principle-driven, this is a clean alternative to an Electron-stuffed 300MB dev tool. The fact that 529 HN points showed up the day after release tells you the market for sub-megabyte tooling is real. v1.1.0 shipped May 17.
https://github.com/gi-dellav/zerostack
Comments