frouter: I Built a CLI That Finds Free AI APIs So You Can Vibe Code for $0

Two people I’d wanted to meet for a while showed up unexpectedly at dinner last night. Somewhere in the conversation one of them pointed out that I’d shipped far less publicly than they expected. That stung, mostly because it was true.

I went home, couldn’t stop thinking about it, and by 3am I was writing code. By morning I had frouter (free router), a CLI tool that finds live free AI APIs and plugs them straight into your vibe coding setup.

The cost wall is higher than the skill wall

I keep hearing the same thing from people around me: “I want to try vibe coding, but I can’t justify the API bills.” For students especially, even $20/month is real money. The irony is that free model capacity already exists in surprising volume. NVIDIA NIM serves around 100 models at zero cost. OpenRouter lets you filter for prompt-cost-zero models specifically. Pair either with an open-source coding tool like OpenCode and your total spend is literally zero.

The problem is that nobody has connected these pieces into a single workflow.

Finding a fast free model right now is the actual hard part

Some people know free models exist. The harder question is which one is responsive at this exact moment. I’ve tested the same model returning in 200ms in the morning and 5 seconds in the afternoon. That kind of variance kills any real coding session.

frouter solves this by pinging every available model on a 2-second loop. It runs 20 concurrent connections, displays the full status on a single TUI dashboard, and applies progressive backoff to models that fail five times in a row. Each model also shows its SWE-bench tier (S+ through C), so you’re not just picking the fastest response but the most capable one that happens to be free.

One-key config means you don’t touch JSON files

A pattern I’ve seen repeatedly: someone finds a free model, opens the config file for their coding tool, gets confused by the schema, and gives up. frouter writes the config for you the moment you press Enter. It auto-generates the provider settings and model ID for OpenCode or OpenClaw, backs up your existing config with a timestamp so nothing is lost, and stores API keys with 0600 permissions.

For oh-my-opencode users, it also auto-swaps NIM models to their OpenRouter twin when needed. And if you’re scripting, --best mode prints the fastest model ID to stdout.

One line to install, four keys to learn

npx frouter-cli

That’s it. The TUI launches immediately. W/X adjusts the ping interval between 1 and 30 seconds. P opens the API key manager and provider toggle. ? shows all shortcuts. If you prefer a global install, npm i -g frouter-cli works too.

Code and docs are at tonylee.im/frouter and on GitHub.

What I’m not sure about yet

I haven’t stress-tested frouter with more than three providers running simultaneously. The backoff logic might be too aggressive for models that have intermittent cold-start delays rather than genuine outages. And honestly, the SWE-bench tier mapping is hand-curated from public benchmarks, so it will drift as new model versions ship. These are things I plan to fix in the open as people start using it.

On a side note, I’m joining the Ralphton hackathon hosted by Team Attention this weekend, where Goobong Jeong and 12 others will be building together. I want to bring something useful for the participants there too.

The range of what you can do for free is wider than most people assume. The barrier was never capability. It was just bad wiring.