Run a coding agent on opencode's free models — with its own computer

Most agent setups stop at autocomplete: a model suggests, you paste, nothing executes. We shipped something different this week — Kydanet agents now run on opencode — and it's worth explaining the split, because the split is the point.

Two pieces, cleanly separated

• opencode brings the models. Paste your Zen or Go key into Kydanet, or skip the key entirely and use opencode's free models.
• Kydanet brings the computer. The agent gets its own Linux environment: a real shell, a real filesystem, a browser it can drive, mail, chat, the works.

That separation matters. Model choice and execution environment are independent problems. You can swap the model without losing the machine, and the machine keeps working whether the model is frontier-priced or free.

What "free models" actually buys you

If you've ever watched a token bill climb while an agent re-reads the same file for the tenth time, you know the pain. The expensive part of agent work isn't the clever answers — it's the grind: scanning directories, running tests, retrying flaky builds, reading logs. That grind is mechanical, and mechanical work should run on cheap or free inference.

So the configuration we shipped is deliberately boring:

1. Point Kydanet at opencode.
2. Bring a key, or don't.
3. The agent gets its computer either way.

No token cost on the free-model path. The agent still executes — cd, reads files, runs the test suite, opens a browser to check the rendered page — because it was never just a chat window.

Where this lands in practice

The pattern that's been working for us:

• Free models handle the grind. Directory sweeps, log parsing, mechanical refactors, repetitive verification.
• The computer is permanent. State survives between prompts. The agent doesn't re-explain your repo every session — it just cds into it.
• You keep the off switch. Everything runs in an environment you own, so "stop" means stop.

Try it

If you want the environment without wiring up your own box: that's the product. Kydanet gives your agent a computer, and opencode gives it a brain — free, if you want it to be.

Questions or counter-arguments welcome in the comments — particularly from anyone who's run long agent sessions on free-tier models and hit the failure modes I should be writing about next.