← Build on your apps

Build with Claude Code or Codex

Build the whole app in Claude Code or Codex, push it, and let your agent read Onvoi’s results and fix them — Onvoi checks, previews and publishes.

Your agent owns the code

Build 100% of an app in Claude Code, Codex or any coding agent, with your own subscription. Onvoi is where it is checked, built, previewed and — once a person approves — published. Onvoi’s own coding agent is never involved and never charged for, and Onvoi never rewrites your code.

1. One command to set up

Create the app with Apps → + Create an app → Develop with Claude Code or Codex. Onvoi gives you a command with your app in it. Run it in your project folder:

terminal
curl -fsSL https://api.onvoi.ai/api/apps-dev/cli/setup.sh | sh -s -- <your-app-id>
  • Signs you in once in your browser (a short code to confirm — no password or token in the terminal). Revoke it any time in Settings → Connected dev tools.
  • Adds a Git remote named onvoi. Your origin (GitHub) and your history stay exactly as they are.
  • Makes Git sign in to Onvoi through the CLI with short-lived tokens — nothing is stored in the repo.
  • Writes Onvoi’s guide to .onvoi/AGENTS.onvoi.md, a Claude Code skill in .claude/skills/onvoi/, and a short section in AGENTS.md / CLAUDE.md — only if you don’t already have those files. Yours are left alone unless you run onvoi setup --merge.
  • It never commits for you, never switches your branch, and is safe to run again.

2. The loop your agent runs

terminal
onvoi doctor            # check with the same rules Onvoi’s builder uses
git commit -am "…"
onvoi push              # push, then wait for THIS commit’s result
onvoi diagnostics --json # if it failed: code, file, line, what was expected, the fix

Ask your agent: *“Build this app for Onvoi and get a working preview.”* It pushes, reads the result, fixes what the diagnostics say and pushes again — you don’t copy logs around.

Accepted ≠ liveA push is never a deployment. Statuses are: received → analyzing → building → failed / needs action / preview ready → awaiting approval → published. onvoi status --wait waits (with a time limit) and its exit code tells the agent what to do next.

Who acts next

  • source — fix the code and push again (exit code 1).
  • transient — Onvoi had a temporary problem; push again without changing code, at most twice (exit 3).
  • action_required — a person must act in Onvoi: add a secret’s value, change the plan, confirm a license, or approve publishing (exit 4). Agents stop and ask; they never guess a secret or approve their own release.
  • Plans and docs only (exit 6) — accepted and kept, nothing to build yet.

Over MCP instead

Agents connected to Onvoi over MCP get the same answers from app.platform_guide, app.check, app.status and app.diagnostics — the same services the CLI calls.

Onvoi → your computer → Onvoi

Any app can continue on your computer: open it and choose Continue on my computer. It’s one repository, so you can switch back any time. Onvoi’s agent works on its own branches; onvoi setup lists any in progress so you pull before editing the same files, and Git refuses a push that would overwrite someone else’s work.