# Supabase's New Monitoring Tool Is a Prompt You Run Yourself > Source: > Published: 2026-09-16 00:00:00+00:00 # Supabase's New Monitoring Tool Is a Prompt You Run Yourself Supabase's Observability on Auto-Pilot ships four prepared prompts for AI agents to monitor a database, run through Claude Routines, Codex or Cursor. Supabase shipped four prepared prompts on Sept. 13, 2026 that turn a general-purpose AI agent into a database monitor, but the company isn’t running any of them for you ([Supabase changelog, “Observability on Auto-Pilot,” Sept. 13, 2026](https://supabase.com/changelog/observability-autopilot)). The product is the prompt, not the schedule that runs it. ## What shipped The four roles are Health and Performance, checked hourly, and Security and Capacity, checked daily. Health watches API and Auth error rates and connection pressure; Performance tracks slow queries, lock waits and long-running sessions; Security reviews Security Advisor findings and RLS gaps; Capacity follows request, storage and connection growth against a baseline ([Supabase changelog, Sept. 13, 2026](https://supabase.com/changelog/observability-autopilot)). Each one connects through Supabase’s own MCP server with a `project_ref` and `read_only=true`, so the agent can read a project’s state but can’t change it. Supabase frames the problem it’s solving in blunt terms: “Most attempts to use agents for monitoring fail for the same reason: the brief is too vague,” the changelog entry says. Its fix is a specific, scoped prompt per role, meant to be copied into whichever agent harness a team already runs: Claude Routines, Codex scheduled tasks, Cursor Automations, or, per the docs, “any environment you choose” ([Supabase changelog, Sept. 13, 2026](https://supabase.com/changelog/observability-autopilot)). The same release expands the MCP server itself with a new `query_logs` tool, giving an agent structured access to logs across the API, Auth, Storage, Edge Functions and Postgres services in one call, instead of stitching together separate log sources by hand ([Supabase changelog, Sept. 13, 2026](https://supabase.com/changelog/observability-autopilot)). Supabase’s setup docs add a caution that matters here: because each scheduled run starts with fresh context, the prompts have to be fully self-contained, and because logs and query results can carry secrets or personal data, teams are told to redact before anything leaves the project ([Supabase docs, “Hire an agent”](https://supabase.com/docs/guides/observability/automate-with-agents)). ## The take Calling this “observability” oversells it. Health checks, slow-query alerts and growth tracking are table stakes for a hosted database, and most competitors bundle them as a dashboard with built-in alerting, not a prompt a developer has to paste into their own scheduler. Supabase is betting that its customers already run an agent harness for other work, so handing them a well-scoped prompt is cheaper to build than a new monitoring product, and cheaper for the customer than standing up Datadog. That’s a fair bet for the teams already living in Claude Routines or Cursor. It’s a worse deal for anyone who wanted monitoring that works out of the box: they now own the schedule, the harness costs and whatever breaks when a run silently stops firing. Read-only access limits the blast radius if a prompt goes wrong, but it also means the agent can flag a problem, never fix it, and depends entirely on someone actually reading its output. The `query_logs` addition is the more durable piece. A single structured entry point across five services is useful infrastructure whether or not anyone hires an agent to sit on top of it, and it’s the kind of primitive Supabase’s MCP server was missing. ## Key Takeaways - Supabase’s Sept. 13, 2026 release adds four prepared monitoring prompts — Health, Performance, Security, Capacity — run through a customer’s own agent harness, not a Supabase-hosted service. - Agents connect via the Supabase MCP server with `project_ref` and`read_only=true` , so they can read project state but not modify it. - The same release adds a `query_logs` MCP tool for structured log access across API, Auth, Storage, Edge Functions and Postgres. - Supported harnesses include Claude Routines, Codex scheduled tasks and Cursor Automations, with prompts portable to any other agent environment.