Build, verify, ship. Documented.

The agent loop

Every braincell turn runs five phases: describe → design → build → verify → deploy. The agent owns each phase; you only step in to confirm the result.

• Describe. You write a single prompt. We parse intent, choose a stack, scaffold a project plan.
• Design. The agent drafts pages, components, and copy with the brand kit baked in.
• Build. Code is written, dependencies installed, and the dev server boots inside a Vercel microVM.
• Verify. Typecheck, lint, build, and a screenshot run before any change reaches you. Five strikes per turn.
• Deploy. One click ships to *.vercel.app with snapshot history attached.

Sandboxes (Vercel microVM)

braincell does not render previews in a browser stub. Every project boots a real Vercel microVM. Your install script runs, your env vars resolve, your routes hit real Node. If something fails in production, it fails in the sandbox first.

Snapshots and rollback

The agent captures the full project state on every assistant turn. Rewind any decision with a single click. Snapshots are diffable, named, and free on every paid plan.

Universal Key

One key proxies to Anthropic, OpenAI, Google AI, and OpenRouter. Billed against your braincell credits, no separate provider accounts required. Drop-in compatible with the OpenAI SDK.

Supported models

• • Claude Opus 4.7: default, deep multi-step reasoning
• • Claude Sonnet 4.6: balanced quality and cost
• • Claude Haiku 4.5: fast tool-calls, low credits
• • GPT-5 and OpenRouter fallbacks for niche tasks

Command line

A CLI ships alongside the workspace: clone a project locally, run the sandbox commands, push back to braincell. Targeting public beta in 2026.