How Team-AI works

Sign up, paste your Anthropic API key under Settings, and start a new project with a one-sentence brief. The eight specialists plan, code, build, and test in sequence — the Integration Validator runs npm install and your test suite before the pull request opens on your GitHub. Build failures trigger up to three automatic rework rounds, root-cause first.

Team-AI uses a bring-your-own-key (BYOK) model. You add your Anthropic API key under Settings, and your AI usage is billed directly to Anthropic at their published Claude pricing. With prompt caching enabled, a typical 15-task project runs about $0.10 in API costs. The platform itself is free to use during the current preview.

Yes. Everything the agents write is committed to your own GitHub repository, on a feature branch, as a normal pull request. Edit, revert, rebase, or merge as you would any branch. Each commit is attributed to the agent that wrote it, so the diff is auditable.

Project generation is fast. Simple projects (CRUD app, landing page) take roughly 3-5 minutes end-to-end — from your description to a pull request on GitHub. Medium projects with multiple integrations take 5-10 minutes. Complexity compounds mostly through rework rounds when build checks fail, not raw LLM time.

Projects are stored in private GitHub repositories under your own account — the Team-AI platform creates them private by default and you remain the owner. Your Anthropic API key lives only in your own user record and is used exclusively for your project's LLM calls. The platform does not claim formal certifications (SOC 2, HIPAA, etc.) at this stage, so please evaluate whether that's acceptable for your data before using it with anything sensitive.

Generated code lives in your own GitHub repository, so you can iterate normally — commit new changes, open your own PRs, or describe modifications in a follow-up project. The Integration Validator also triggers up to three automatic rework rounds when build checks fail, without any input from you.

Sign in to console.anthropic.com, add a payment method (Anthropic bills you directly), then create an API key under Settings → API Keys. Copy the key — it starts with sk-ant- — and paste it into Team-AI under Settings → API Keys. We never see, log, or proxy your key. It's encrypted at rest with Fernet (AES-128-CBC + HMAC-SHA256) before being written to the database, decrypted only at the boundary of a Claude call, and never leaves the process in plaintext.

On first project creation we ask for a GitHub OAuth grant with the repo scope so the platform can create a new private repository under your account, push the agents' commits to a feature branch, and open a pull request. We don't read or modify any of your existing repos. You can revoke the grant any time at github.com/settings/applications.

Today, projects are one-shot — agents run from scratch and ship a PR. To add a feature you commit changes yourself in the generated repo, or start a new project that includes the new requirement and copy the relevant files across. In-place iteration ("add password reset to my Todo App") is on the roadmap; until it lands, the generated repo is yours to evolve directly.

Reliable: React + Vite, Vue 3, vanilla JS frontends; Node.js (Express), Python (FastAPI), and Django backends; Postgres, MongoDB, SQLite for storage; Tailwind CSS; Jest/Vitest/pytest for tests. Less reliable today: native mobile (iOS/Android), microservice architectures, anything requiring custom build tooling beyond npm install && npm test. The PM agent picks the stack from your description — you can also nudge it ("React + Postgres", "use Express not Django").

Two layers: a per-IP global cap of 240 requests/minute (returns 429 with a one-minute Retry-After) and a per-user project-creation cap of 5 projects/minute. You'll never hit either in normal use. On top of that, your Anthropic key has its own organisational rate limits — if you see a Claude error, check the Anthropic dashboard's usage page first.

Go to Settings → Account → Delete account — it removes your user record, all project metadata, tasks, and your stored Anthropic API key. GitHub repositories created during projects are NOT deleted (they live in your own GitHub account; manage them there). If you want to revoke our GitHub OAuth grant separately, head to github.com/settings/applications.

If a single LLM call fails, the agent retries with exponential backoff. If a build or test check fails after the agents finish, the Integration Validator triggers a rework round — up to three — that asks the relevant agents to fix the failure with the failing output included as context. Rework prompts are structured root-cause first, so the third round (if needed) targets the underlying issue rather than the symptom. If the validator still can't get a green build after three rounds, the PR is opened anyway with the failing checks documented in the PR body so you can finish the fix yourself. Hard crashes (rare) are caught by Sentry; you can also abort a run from the project page.

The full REST API that powers the dashboard is documented at /ai_team_api_reference.html and uses the same JWT cookie / Bearer token auth as the UI. There isn't a dedicated CLI yet; if you want one, sign up and post in the feedback channel — it's high on the roadmap if there's pull.