Meet 3code: The Featherweight Coding Agent That Won't Eat Your Budget

markdown formatted blog content

Let's be honest: the current generation of AI coding agents is getting out of hand. We're talking 85 MB binaries, runtime dependencies, web UIs that need hosting, and system prompts that chew through thousands of tokens before you even type your first query. It's as if the industry decided that bigger equals better, and somewhere along the way, efficiency went out the window.

Enter 3code, a coding agent that's taking a radically different stance. Built in Nim and clocking in at just 1.6 MB, it's the programming equivalent of a Honda Civic in an era of gas-guzzling SUVs. And here's the kicker—it actually works.

Why Efficiency Should Matter to Developers

Before we dig into what makes 3code special, let's talk about why any of this matters. Token costs add up fast. When you're running an AI coding assistant all day, every prompt and every system message contributes to your bill. And it's not just about money—larger contexts mean slower responses and more chances for the model to lose track of what you're actually trying to do.

The folks behind 3code call their philosophy the "Three E's":

• Token efficiency — Cutting out the fat that doesn't serve your actual coding needs
• Computer efficiency — It should be a chat program, not a resource hog
• Brain efficiency — Automation where it counts, so you don't waste mental energy on repetitive tasks

This isn't just marketing speak. The numbers are pretty compelling.

The Numbers Don't Lie

Here's how 3code stacks up against the competition:

| Tool | Binary Size | System Prompt (tokens) | |------|-------------|------------------------| | 3code | 1.6 MB | ~600 | | opencode | ~22 MB | ~2,100 | | goose | ~35 MB | ~1,000 | | claude code | ~85 MB | ~2,900 |

That's a stark difference. 3code's binary is roughly 53 times smaller than Claude Code, and its initial context is about five times leaner. Early testing suggests this translates to 30-50% token savings in real-world usage—meaning your dollar goes further and the model stays focused on your actual task.

How It Works

3code is refreshingly no-nonsense. It's a command-line tool with no daemon running in the background and no web interface to maintain. You install it, point it at a supported model provider, and start coding.

Currently supported providers include:

• NVIDIA with GPT-OSS (free tier available, decent performance)
• z.ai with GLM-5.1 (excellent performance, affordable pricing)

The installation is dead simple—single commands for macOS/Linux or PowerShell on Windows, and you're off to the races.

Technical Highlights That Actually Matter

Some features caught my eye:

• Session persistence — Your .3log files are human-readable, so you can pick up any past session exactly where you left off
• Context compaction — The tool intelligently removes stale information, keeping context lean
• Loop guard — Prevents runaway autonomous edits by halting at configurable thresholds
• Native web search — Built-in Startpage integration without needing curl
• No telemetry — Everything stays local. Nothing phoned home.

The self-clearing execution model is particularly clever for larger tasks. When 3code finishes a planning phase, it resets context before moving to execution, preventing the kind of context pollution that makes other agents lose the thread.

Dogfooding at Its Finest

Perhaps the most convincing endorsement? The developers built much of 3code using 3code. They're running it in production with GLM 5.1 on z.ai's coding plan and haven't looked back. That's the kind of confidence you want to see from a tool's creators.

The Bottom Line

3code isn't trying to be the most powerful coding agent on the market. It's trying to be the most efficient. And for developers and startups watching their API costs, or anyone who values a snappy, no-frills tool that just works, it might be exactly what you need.

It's a reminder that in the race to add more features and more complexity, sometimes less really is more.

Check it out at 3code.capocasa.dev and see if lightweight is your thing.