Zerostack: a 12MB Rust coding agent worth a look from SL
Zerostack is a new Unix-inspired Rust coding agent that runs in about 12MB of RAM. I walk through why that number matters more on a 4GB laptop than on an M3 MacBook.
Zerostack, a Unix-inspired Rust coding agent, landed on crates.io at version 1.0.0 — and the spec sheet alone makes it worth a second look. The listing is on crates.io with source on GitHub. What pulled me in: the slug 2026-05-17-zerostack-a-unix-inspired-coding-agent-written-in- points to a tool built around a single number — ~12MB of RAM during a working session. For a coding assistant, that is almost rounding error.
I want to think out loud about why that number matters more in a Sri Lankan CS lab than it does in a San Francisco office.
⚡ What Zerostack actually is
The crate's description is honest about its scope: "Minimalistic coding agent written in Rust, optimized for memory footprint and performance." About 6,685 lines of Rust across 45 files, GPL-3.0, single author.
| Feature | Detail |
|---|---|
| Install | cargo install zerostack |
| Providers | OpenRouter, OpenAI, Anthropic, Gemini, Ollama |
| Prompt modes | code, plan, review, debug, ask, brainstorm, frontend-design, review-security |
| Sandbox | Optional bubblewrap shell isolation |
| Memory (idle) | ~8 MB |
| Memory (working) | ~12 MB |
| Not included | Skills, Subagents (not implemented in v1.0.0) |
The prompt modes are plain markdown files you can swap out and check into git — giving a small team a single source of truth for shared system prompts. That alone is a cleaner workflow than pasting into a chat window every morning.
💻 Why 12MB matters when your laptop has 4GB
Here is the part that is specific to where many readers of this site actually work.
A typical second-year CS student at UCSC or NSBM is not coding on an M3 MacBook. They are on a 4–8GB Windows or Linux laptop already running:
- Chrome with 20+ tabs open
- VS Code plus an active dev server
- Docker Desktop running a database container
- WhatsApp Web or Discord in the background
By the time the project file is open, you are often at 80% memory before any AI tooling starts. Electron-based coding tools hit you with a second tax: they bring an entire browser engine along for the ride.
| Tool | Approximate RAM usage |
|---|---|
| Claude Code (during a session) | ~500MB–2GB |
| Opencode (large project, per HN reports) | Up to 6GB |
| VS Code Copilot extension overhead | ~200–400MB |
| Zerostack (working) | ~12MB |
The author attributes the footprint to specific technical choices: smallvec and compactstring for stack-allocated data instead of heap allocations, opt-level=z for size-optimised builds, and Link Time Optimisation enabled. None of these are exotic. Together they are the difference between a CLI that feels native and one that brings a JavaScript city with it.
💰 BYOK, Ollama, and what this means for Sri Lankan wallets
The provider list is where this gets practically interesting for anyone managing an LKR income.
OpenRouter takes pay-as-you-go top-ups and works with prepaid USD cards — a known advantage over monthly subscriptions billed to foreign-issued credit cards, which many Sri Lankan banks still restrict. The math is small but real: if you only run a coding agent a few hours a week, a $5 OpenRouter top-up can outlast a $20 monthly subscription by weeks.
| Provider option | What it costs | Works in Sri Lanka? |
|---|---|---|
| Anthropic direct | Monthly or pay-per-token | Needs Visa/Mastercard, some friction |
| OpenRouter | Pay-as-you-go top-up | Works with prepaid USD cards |
| Ollama (local models) | $0 recurring | Yes — no internet needed after download |
| Gemini API free tier | Free up to rate limits | Yes |
Ollama matters most for the budget-conscious case. Point Zerostack at a local model — Qwen 2.5 Coder, DeepSeek Coder, or anything you can run on your hardware — and the agent works without touching a paid API at all.
For students experimenting on a tight data cap or a prepaid mobile plan, this is the difference between "I can try this" and "I cannot justify the spend this month." Zero recurring cost once the model is downloaded.
🛠️ How to evaluate a 1.0.0 crate before trusting it
A single-author, 1.0.0 release on crates.io deserves a read before you give it shell access to a work project. Here is a fast checklist:
| Check | What to look for |
|---|---|
| Dependency count | Fewer is safer — less transitive attack surface |
| Last commit date | Has anyone been active this month? |
| Issue tracker | Open issues with no response = slow maintainer |
| License | GPL-3.0 means copyleft — check your employer's policy |
| Sandbox option | Does it offer process isolation before running shell commands? |
Zerostack passes most of these for a toy project. The bubblewrap sandbox option is a good sign. The single-author status is a caution for anything production-facing.
✅ What this means for you
I would not bet a paying client project on a 1.0.0 release from a single developer without a backup workflow. But that is not the only use case.
For studying how a modern Rust CLI agent is built from scratch, this is one of the cleaner small codebases available right now. For trying out local-model coding workflows without a hardware tax or a subscription, it is worth an afternoon. For keeping AI tooling running on a 4GB laptop while Chrome eats the rest of your RAM, it is the most practical option I have seen this week.
Clone the repo, read it for an evening, and you will come out knowing more about Rust CLI design than most tutorials will teach you.
🔗 Useful Tools
- AI Model Comparison — compare Claude, GPT, Gemini and others side by side before choosing a provider
- AI Token Counter — estimate token costs before committing to a paid API plan
- Sri Lanka Fixed Deposit Calculator — if you are saving toward a dev setup upgrade, see what your LKR earns in an FD first