The in-app AI

Code: convex/ai.ts. Key: the space's own Claude key from its vault ("Connect your AI" in HQ, or the Connections page). The platform's own key is optional and off for now: when it's set, spaces without a key use it under the platform's limits (docs/build/AI-LIMITS.md); when it isn't, they see "Connect your AI".

How it works #

1. The person sends a message (ai.send), which schedules ai.respond.
2. respond sends Claude the block list, the rules and the current blueprint, plus the last 20 messages.
3. Claude calls propose_change({ summary, ops }). The server applies the ops to a copy and runs checkBlueprint. Errors go back to Claude to fix (up to 6 rounds).
4. A valid proposal is stored as pending. The person approves it in the chat or in HQ.

Limits #

• The AI never writes the blueprint directly; only blueprint.approve does, after a person taps it.
• It can't remove the HQ page.
• Mini-app HTML runs sandboxed (WebView on phones, iframe on web) with a strict content policy (no network).

Next tools #

• read_blueprint for big workspaces (send only the part it needs)
• Capability tools: list connected MCP servers and keys by name, call them on the server
• save_template: turn a page or the whole blueprint into a reusable template