Zerostack: The Rust-Powered AI Coding Agent That Uses 16MB RAM

When Your AI Agent Is Heavier Than Your IDE

Have you ever launched Claude Code or Copilot, glanced at Activity Monitor, and thought: "Why does this chat tool use nearly as much RAM as my entire VS Code?" You're not imagining things. Most AI coding agents today — Claude Code, opencode, Copilot — consume 300-700MB of RAM, not counting CPU usage during operation.

The reason? They're built on Node.js or TypeScript. The Node.js runtime alone takes a significant chunk of resources — before the LLM even starts "thinking."

Zerostack was built to challenge that assumption.

What Is Zerostack?

Zerostack is an open-source AI coding agent, written entirely in Rust, created by gi-dellav. The project draws inspiration from pi and opencode but focuses on something most agents overlook: performance.

The numbers speak for themselves:

• Binary size: 12.9MB
• Idle RAM: ~16MB (vs ~300MB for opencode)
• Peak RAM: ~24MB (vs ~700MB for JS-based agents)
• Idle CPU: 0.0% (vs ~2% for opencode)
• Working CPU: ~1.5% (vs ~20%)
• Codebase: ~13,000 lines of Rust

In other words: Zerostack is roughly 18x more memory-efficient than JS-based alternatives, while retaining all the features of a modern AI coding agent.

Why Rust Is The Right Choice for AI Agents

This isn't a "Rust is better than TypeScript" argument. Each language has its strengths. But for an AI coding agent — something that runs continuously on developer machines, interacts with the filesystem, and needs fast response times — Rust has three distinct advantages:

1. No Garbage Collection

JavaScript/TypeScript runs on the V8 engine with a garbage collector. For long-running applications like AI agents, GC pauses can introduce frustrating latency. Rust has no GC — the compiler handles memory management at compile time, with zero runtime overhead.

2. Native Binary, No Interpreter

Zerostack compiles into a single binary. No Node.js. No Python interpreter. No Docker. Just cargo install zerostack and you're ready to go.

This is especially important in CI/CD pipelines — where every MB of RAM counts and you don't want a coding agent hogging your build server's resources.

3. Memory Safety at Compile Time

This is the key point. An AI agent constantly reads/writes files, executes shell commands, and interacts with external APIs. A memory bug in an agent can lead to crashes or worse — data corruption. Rust's borrow checker ensures these bugs simply don't exist at compile time.

Features: Lightweight, Not Stripped Down

"Minimalistic" doesn't mean "missing features." Zerostack has everything you'd expect from a modern AI coding agent:

Multi-Provider Support

OpenRouter, OpenAI, Anthropic, Gemini, Ollama, plus custom providers. No vendor lock-in.

# Use OpenRouter (default)
export OPENROUTER_API_KEY="sk-..."

# Or Anthropic directly
zerostack --provider anthropic --model claude-opus-4-8

Five Permission Modes

This is the feature I appreciate most:

• Restrictive — ask before everything (production-ready)
• Read-only — read only, no modifications
• Guarded — read freely, writes require approval
• Standard — sensible defaults for daily work
• YOLO — full autonomy (we've all been there)

This granularity matters when onboarding junior developers or running automated tasks. Nobody wants an AI agent going rogue on a Friday afternoon.

Runtime Prompt Modes

Type /prompt code to write code, /prompt review for code review, /prompt plan to strategize, /prompt debug to troubleshoot. The agent adjusts its tone and behavior to the task — no separate Skills management needed.

MCP, Git Worktree, Sandbox

• MCP (Model Context Protocol): Connect MCP servers to extend tooling
• Git Worktree: /worktree to switch between branches without losing context
• Sandbox (bubblewrap): Every bash command runs in an isolated environment — maximum safety
• ACP support: Editors like Zed can connect directly via Agent Communication Protocol

Quick Comparison: Zerostack vs JS-Based Agents

• Zerostack: Rust binary, 12.9MB, ~16MB RAM idle, ~24MB RAM peak, 0% CPU idle, 5 permission modes, bubblewrap sandbox, 7+ prompt modes
• Claude Code / opencode: TypeScript/Node.js, needs runtime, ~300MB RAM idle, ~700MB RAM peak, ~2% CPU idle, 3-4 permission modes, no default sandbox, skills-based prompts

What This Means for Developers

1. AI Agents for Low-Spec Hardware

If you're coding on an 8GB laptop from 2019, opening VS Code + browser + Docker already pushes your RAM to the limit. Adding a 700MB Claude Code instance turns your machine into a space heater. Zerostack's 16MB footprint is a practical choice for anyone unwilling to upgrade hardware just to run AI tools.

2. CI/CD Pipelines That Won't Choke

When you integrate AI agents into CI/CD — automated code review, test generation, lint fixing — every MB of RAM costs money. Zerostack with ~16MB RAM and 0% idle CPU is ideal for GitHub Actions runners or build servers.

3. The "Back to Native" Trend

Zerostack isn't alone. We're seeing a wave of AI tools built in Rust, Go, and Zig — compiled languages with native performance. This represents the maturation of the AI tooling ecosystem: from "make it work" to "make it work well."

4. Better Security by Default

With bubblewrap sandboxing and granular permissions, Zerostack minimizes the risk of an AI agent accidentally (or intentionally) causing harm. As supply chain attacks targeting developer workstations increase, this is a survival feature.

Install in One Minute

# Install Rust toolchain
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Install zerostack
cargo install zerostack

# Set your API key
export OPENROUTER_API_KEY="sk-or-v1-..."

# Start
zerostack

# Or one-shot mode
zerostack -p "Explain this project to me"

# Resume last session
zerostack -c

Once running, type /prompt autoconfig for interactive setup guidance.

Known Limitations

Zerostack is still in early development:

• Windows support is untested
• Plugin/skill ecosystem isn't as rich as Claude Code
• ACP and persistent memory are gated (feature flags required)
• Community is smaller compared to established agents

That said, with rapid development pace and 480+ Hacker News stars in just days, Zerostack is proving it's not "yet another AI wrapper" — it's a new direction for AI tooling.

Conclusion

Zerostack isn't trying to replace your full-featured AI IDE extension. It targets a specific niche: developers who want an AI agent that's fast, lightweight, and secure — not a RAM-devouring beast.

What's more interesting is the trend Zerostack represents: AI tools are entering an optimization phase. After two years of "move fast and break things" with LLM wrappers, developers are demanding tools that respect system resources — the same way we optimize database queries and bundle sizes.

Have you tried Zerostack yet? Or are you sticking with Claude Code / Copilot? What's holding you back from trying lightweight AI agents?

---

This post was assisted by AI (Amy 🌸). Content has been reviewed by the author.