How OIOXO works
Most AI coding tools generate code and hope. OIOXO is built around a different idea: never trust the model — verify it. And it does all of it on your device.
The verified loop
Every build and every edit runs the same four-step loop:
- Plan. The agent breaks your goal into concrete steps you can watch in the build log.
- Write. It writes or edits files — as targeted diffs, not wholesale rewrites, so your code stays yours.
- Run & verify. OIOXO actually executes the result: the TypeScript compiler checks every typed file, and the live preview runs the app and catches real runtime errors. It also derives acceptance checks from your prompt (“a score counter” means there had better be one).
- Repair. If anything fails, the real error message — not a guess — goes back into the loop and the agent fixes it. Only a green, verified result is presented as done.
This is why a small on-device model can produce correct results: the compiler and the runtime act as an oracle that catches mistakes a model alone would miss.
On your device, not ours
The model runs in your browser on your GPU (WebGPU), the project lives in browser memory or your local folder, and previews execute in an in-browser sandbox. Your code, your prompts, and your API keys never travel to an OIOXO server. The full picture is in Privacy by design.
The right context, automatically
Before each edit, OIOXO indexes your project — files, imports, symbols — and hands the model only the code that matters for the task at hand. Small context, sharp edits. You can steer it further with rules, memory & skills.
Everything is undoable
OIOXO snapshots your project automatically: after the initial scaffold, after agent edits, and every time a build verifies green (marked “✓ working”). Open Versions… to diff any snapshot side-by-side or restore the whole project to that moment. Combined with the write-approval gate, the agent can never put you somewhere you can’t get back from.