The latest articles on DEV Community by Opulent Byte (@opulentbyte). https://dev.to/opulentbyte https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4169837%2Fab7a9622-cea3-476f-baeb-14ca0d7ce993.png https://dev.to/opulentbyte en Opulent Byte Fri, 09 Oct 2026 17:04:56 +0000 https://dev.to/opulentbyte/run-a-coding-agent-on-opencodes-free-models-with-its-own-computer-31fm https://dev.to/opulentbyte/run-a-coding-agent-on-opencodes-free-models-with-its-own-computer-31fm <p>Most agent setups stop at autocomplete: a model suggests, you paste, nothing executes. We shipped something different this week — <strong>Kydanet agents now run on <a href="https://opencode.ai" rel="noopener noreferrer">opencode</a></strong> — and it's worth explaining the split, because the split is the point.</p> <h2> Two pieces, cleanly separated </h2> <ul> <li> <strong>opencode brings the models.</strong> Paste your Zen or Go key into Kydanet, or skip the key entirely and use opencode's free models.</li> <li> <strong>Kydanet brings the computer.</strong> The agent gets its own Linux environment: a real shell, a real filesystem, a browser it can drive, mail, chat, the works.</li> </ul> <p>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.</p> <h2> What "free models" actually buys you </h2> <p>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.</p> <p>So the configuration we shipped is deliberately boring:</p> <ol> <li>Point Kydanet at opencode.</li> <li>Bring a key, or don't.</li> <li>The agent gets its computer either way.</li> </ol> <p>No token cost on the free-model path. The agent still executes — <code>cd</code>, reads files, runs the test suite, opens a browser to check the rendered page — because it was never just a chat window.</p> <h2> Where this lands in practice </h2> <p>The pattern that's been working for us:</p> <ul> <li> <strong>Free models handle the grind.</strong> Directory sweeps, log parsing, mechanical refactors, repetitive verification.</li> <li> <strong>The computer is permanent.</strong> State survives between prompts. The agent doesn't re-explain your repo every session — it just <code>cd</code>s into it.</li> <li> <strong>You keep the off switch.</strong> Everything runs in an environment you own, so "stop" means stop.</li> </ul> <h2> Try it </h2> <p>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.</p> <p>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.</p> ai opensource productivity coding