DEV Community

Cover image for I Built an AI Coding Agent That Ships in Under 50KB — Zero Dependencies, Runs on a Phone
Emmanuel Ene Rejoice Gideon
Emmanuel Ene Rejoice Gideon

Posted on

I Built an AI Coding Agent That Ships in Under 50KB — Zero Dependencies, Runs on a Phone

#cli #javascript #opensource #ai

I'm a developer from Enugu, Nigeria. Most devs I know here code on second-hand laptops — and a lot of them code straight from cheap Android phones using Termux.

So when I set out to build an AI coding agent for the African market, "first, install 400 npm packages" was never going to work. Neither was "you need a $2,000 laptop."

I set myself one hard rule: the entire package must stay under 50KB, with zero dependencies.

Meet Stew Code — a terminal AI agent that hits that rule and still does all of this:

  • 👀 Vision — /image screenshot.png and it actually sees the screenshot
  • 🌐 Headless browsing — cookies, forms, clicks, login flows. No Puppeteer
  • 📸 Screenshots & screen recording — web pages or your whole screen
  • 🔒 Security scanning — audits your codebase for real issues
  • 🐝 Swarm mode — a planner splits your task, parallel agents execute it
  • 🗣️ Voice — it reads responses out loud using OS text-to-speech
  • 🔌 MCP client — connect any Model Context Protocol server
  • 🛠️ Self-healing installs — missing Chromium? It auto-installs it, and warns you first: "~250MB download, battery at 12% — plug in before we start"

Quick start 

npm i -g stew-ai
Enter fullscreen mode Exit fullscreen mode

That's the whole install. 49KB. On a bad 3G connection it takes seconds, not minutes.

Grab a free API key at https://stew-agent.onrender.com (free tier: 1,500 calls/month), then:

stew auth login   # paste your key
stew code         # launch the REPL
Enter fullscreen mode Exit fullscreen mode

And you're in.

What the brain looks like

Honest disclosure: the intelligence runs on my Stew Agent API — the CLI is the lightweight terminal client. The API is OpenAI-compatible (/v1/chat/completions), so you can also point any OpenAI-shaped tool at it by changing the base URL.

The API side packs 59 skills, a 100-agent swarm, 12 personas, and 6 AI providers (Groq, Mistral, NVIDIA NIM, OpenRouter, HuggingFace, OpenAI) with automatic failover — if one provider goes down, it switches mid-flight. And when you npm i -g stew-ai, none of that bloat lands on your machine. One package. Zero dependencies.

The zero-dependency receipts

How do you fit vision, browsing, screenshots, recording, a REPL, and an MCP client into 49KB with zero deps?

  • Node 18+ native fetch — no axios, no node-fetch
  • Raw ANSI escape codes for all terminal coloring
  • Hand-written argument parsing
  • Chrome headless --screenshot for page shots; OS-native tools (scrot, screencapture, ffmpeg, PowerShell) for everything else
  • The npm package ships minified; the GitHub repo stays fully readable
  • A regression suite runs offline against mocks, so nothing breaks silently between releases

Try it on your phone (yes, really) 

# Install Termux from F-Droid, then:
pkg install nodejs
npm i -g stew-ai
stew code
Enter fullscreen mode Exit fullscreen mode

That's an AI coding agent with vision and web access running on an Android phone. That's the entire point — AI tooling shouldn't require hardware most of the world can't afford.

Links

If you're building for emerging markets, steal this constraint: under 50KB, zero deps. It forces better engineering than any roadmap ever will.

Star the repo if it's useful. And if something breaks on your machine, run stew sysinfo and paste me the report — that command exists because debugging from a phone is hard. 🚀

Top comments (0)