Kodro - Kids can code robots. Astra debugs. Evidence decides.

Kodro is an offline robot studio for UK Computing, KS1 to KS4. Pupils design a rover, write real code, and run it through simulated worlds, getting pass/fail evidence with telemetry on every run. NEW: Genesis races up to 4 controllers through seeded contracts and proves the winner. Built for reasoning agents: Astra can drive the full loop over MCP, reading code, running it, grading it, iterating, while Kodro holds verdicts no model can fake. Honest kinematic sim, MIT open source.

Add a comment

Replies

Best
Hey hunters, Vaibhav here. I build learning tools, and I kept seeing the same sadness: kids typing code into a void, pressing run, getting "wrong" with no reason why. So I built Kodro, where code becomes a robot you can watch, and every run ends with evidence, not a shrug. Give it 60 seconds on the live site: open "Drive to the Flag" and press Run. The rover stops after 1 metre. Flag is at 3. Big honest NOT YET, with telemetry showing why. Change the 1 to a 3, rerun, and enjoy the little rush of a 100 with zero failures. That loop (try, see, understand, fix) is the whole product, across 24 curriculum lessons, KS1 to KS4. What's new today is Genesis. Hand it up to 4 controllers and it races them all through the same seeded simulation contracts, ranks them by evidence, and tells each loser exactly what to fix. Cursor for machines, except the code has to survive the simulator first. Now, Astra, honestly told. This challenge changed how I architected Kodro. I stopped thinking "add an AI chatbot" and started thinking "what would a reasoning agent need to actually drive this?" The answer became 8 deterministic MCP tools: read the lesson, read the program, run it, grade it, prove it over seeded contracts. Same verdicts the pupil sees, so the agent can never hallucinate a pass. Astra reasons, Kodro proves, and the proof keeps everyone honest, including the agent. How you can use it today: if you have ChatGPT with Codex, attach the kodro-mcp server over stdio (the app itself shows you how: pick Astra in the provider list and it hands you the setup with one-click copy). Then ask Astra to inspect your program, run it, read the telemetry, fix the failure, and re-prove it. In the webpage itself Astra is listed as not connected, because I would rather show you that sentence than fake a live call. The verification doc (docs/ASTRA_VERIFICATION.md) maps every claim to code and tests. The simulator is kinematic with openly disclosed limits, no hardware claims, no magic. Free, offline, MIT open source. Roast it, break it, ask me anything. The skeptical questions are my favorite part. 🔧
What became possible in your product with Astra that was not practical before?
The Astra challenge did what a deadline plus a hard question does: it focused everything. Designing for a reasoning agent forced the architecture Kodro has now: machine-readable verdicts, 8 deterministic tools, the Genesis compare loop, and an honest provider story. In this window the product moved from a good simulator to a complete proof system: Genesis, the Connect-Astra panel, EarthProof surfacing, 1908 passing tests, and evidence docs for every claim. The vision was always agents operating and Kodro proving. The challenge made us build the second half.