Build production apps with an autonomous agent

Sign in, click New project, then describe what you want in the chat panel. The agent will plan, write files, lint them, and self-correct until the preview builds.

Every project starts empty. Your first prompt should describe the whole app — e.g. "A pomodoro timer with light/dark mode, keyboard shortcuts, and a session history."

The agent runs a multi-step loop (up to 50 tool calls per turn): plan → read → search → write → lint → fix. You can watch each step as it streams.

• Diffs: every file write shows a unified diff with a lint badge.
• Images: drop screenshots into the composer — the agent reads them.
• Stop early: cancel mid-stream if it goes the wrong direction.

The Preview tab renders the current files in an in-browser Sandpack runtime. The Code tab shows the file tree, editor, and per-file version history with restore.

Every successful write is checkpointed — roll back to any previous version with one click.

Open the Knowledge panel to add URLs, notes, or files. Foundry chunks and embeds them with pgvector and retrieves the top results on every chat turn.

Project files you write are auto-embedded in the background, so the agent always knows the current shape of your codebase.

The retrieval pipeline pulls the top 20 vector hits, then re-ranks with a diversity pass so a single file can't dominate the context window.

Click Agent memory in any project to set free-form notes that follow you across every project — preferred stack, tone of voice, accessibility rules, etc.

These notes are injected into the system prompt on every chat turn.

Click Deploy to publish your current build to /p/your-slug. Every deploy gets a numbered version available at /p/your-slug/v3, etc.

Rollback is instant — pick any prior version and click restore. Deploys are rate-limited to 3/min and 30/day per user.

Invite teammates with viewer, editor, or owner roles. All access is enforced by row-level security.

• Presence: live avatars show who's looking at the same project.
• Comments: thread discussions on the project or on a specific file, with @mention notifications.
• Activity: every write, deploy, and invite is logged.

Publish any project as a public template from the workspace header. Other users can fork it from /marketplace and rate it 1–5 stars.

Ratings are averaged by a database trigger so listings stay fast even under load.

The Integrations panel installs pre-wired Stripe Checkout, Resend email, OpenAI chat, Google Maps, and PostHog snippets into your project with one click.

The agent can also call installIntegration autonomously when you ask for "Stripe payments" or "send email when X happens".

Ask the agent to "make this a mobile app" and it will scaffold a Capacitor wrapper plus /docs/MOBILE.md with the exact npx cap add ios / android steps.

Mobile-first rules (44px tap targets, safe-area insets, no hover-only states) are baked into the system prompt.

Every public table has explicit GRANTs and RLS policies scoped to auth.uid() or has_project_role. CI runs an RLS audit that fails the build on any new TO anon policy without an explicit security-review marker.

Mutating endpoints (chat, deploy, invite, ingest, publish-template) are rate-limited per user via a security-definer RPC.

Server logic uses createServerFn + requireSupabaseAuth. The admin client is only loaded behind explicit role checks. AI calls go through the Lovable AI Gateway — LOVABLE_API_KEY never reaches the browser.