Many small teams learn about a production problem when a user complains. Supabase’s new Agent prompts are ready-made instructions that send your coding agent to check your live Supabase project on a schedule and report what it finds.

Four prompts, each with its own schedule

Each prompt covers one area, and Supabase suggests how often to run it:

  • Health (hourly): API and Auth server errors, error-rate spikes, connection pressure.
  • Security (daily): Security Advisor findings, authentication and authorization failures.
  • Performance (hourly): slow queries, lock waits, long-running sessions, Performance Advisor findings.
  • Resources (daily): growth in requests, errors, storage, tables and connections.

Supabase says each finding includes the measured value against its threshold and one next step. For example, an API error finding can show that the error rate passed 1% and doubled from the previous hour, along with the log source and time range to investigate.

The agent has to save its alert state between runs. Without that, it cannot tell new findings from ones it already reported, and you get duplicate alerts.

Setup and scheduling

The prompts run through Supabase’s MCP server, the connection that lets an agent use Supabase’s tools. Supabase’s example for Claude Code scopes the server to one project and turns on read-only mode, so every query runs as a read-only database user:

claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?project_ref=abc123&read_only=true"

Replace abc123 with your project reference. Then copy a prompt from Supabase’s documentation, run it once and read the report before scheduling it. Supabase names Claude Code routines, Codex scheduled tasks and Cursor automations as scheduling options.

A scheduled run cannot stop to ask you for approval, so Supabase advises pre-approving only the read-only tools the prompt needs. If the agent cannot fetch the data for a check, it reports that it cannot assess the project. An empty report is therefore not proof that nothing is wrong.

Other tools in the same release

Supabase shipped related pieces alongside the prompts. query_logs is an MCP tool that runs read-only SQL against project logs for up to 24 hours per request. The supabase agent skill now covers debugging. Advisors in the dashboard gained a Health tab that flags unusually high error rates across the Data API, Auth, Storage and Edge Functions.

If you have already set up Supabase’s agent confirmations, keep the same habit here: a monitoring agent should only be able to read. Cloudflare takes a similar approach for Workers, where grouped errors go to a coding agent for triage.

End of article