Baton

Baton: Meta splits it. Three AIs build it at once. Problem: people who build with AI subscriptions hit one AI's usage limit halfway through a job. Switching AIs loses the plan, and the next AI breaks what the first one made. Each AI is also better at different work, yet one premium plan gets spent on everything. What it is: a Mac menu bar app. You type one idea and pick a folder. Meta's model first writes a shared contract (stack, folder layout, every shared name, style rules), then splits the work by strength: Claude gets the main code and UI, Codex gets setup, scripts and tests, Gemini gets web research and copy. Each AI gets files only it may change (Baton also enforces this in code) and a complete, ready-to-paste prompt. "Run all at once" starts all three at the same time in the same folder through their official CLIs (claude -p, codex exec, gemini -p), each signed in with the user's own subscription: no paid API keys, no screen scraping. Cards show live progress read from each tool's JSON output. If a tool is missing or its plan runs out, drag the card into the Claude, ChatGPT or Gemini app, or press Copy & open. "Check the work" sends the contract, each AI's report and the file list back to Meta's model, which answers with what's done, what's broken and the best next prompt. Everything stays on the Mac; the Meta key lives in the Keychain. Honest status: native Swift app, 19 tests passing, Claude and Codex verified running in parallel in one folder. Muse Spark on Meta Model API is wired in as the default (the models call works), but my account couldn't verify a payment method, so requests were blocked. A last-minute offline run with Meta's open-weights Llama 3.2 3B didn't finish the planning step on my 8 GB Mac before the deadline, so the demo video shows the real interface with a clearly labelled sample plan.