@AGENTS.md
Claude Code
- MCP (
.mcp.json):mcp__supabase__*= DEVlbefrzvyznaoaglldluk, read + write (apply_migrationfor schema changes).mcp__supabase-prod__*= PROD, read-only: reference data only, never write. - Guards:
.claude/hooks/guard.py(PreToolUse) + deny rules in.claude/settings.jsonblock secret paths (.env.dev*) and whole-tree staging. Best effort, not a sandbox. Each hook block names what matched: a real path, file or history the command touches → stop and ask the user; a word that is only text (message, pattern) → collateral, rephrase (AGENTS.md). Settings deny rules don’t explain themselves: treat those as stop and ask. - Path-scoped rules load from
.claude/rules/(frontend.md,edge-functions.md,database.md). - Agent testing: sign in test users with the password API (
docs/live/dev-environment.md), not magic links. The magic-link recipe needs the service-role key: a human runs it.