Uptime and schema drift monitor for MCP servers

The idea

MCP servers have a uniquely nasty failure mode: when one dies or degrades, nothing errors. The agent’s tool list just quietly shrinks, Claude or Cursor stops being able to do something it did yesterday, and the user experiences it as the model getting dumber. Teams wiring MCP servers into real workflows have no way to see this happen; generic uptime tools return 200-OK on an endpoint that can no longer complete an MCP handshake.

Schema drift is the second silent killer: a server update renames a tool or changes a parameter, and every agent integration downstream degrades without a single error. The category is brand new with no incumbent, the check itself is a weekend of work (initialize, tools/list, diff against snapshot), and being first to own mcp monitoring as a term is most of the opportunity.

Who pays

The buyer is a team that has moved MCP servers from toys to workflow dependencies, internal servers wrapping company APIs, or third-party servers their agents rely on. At 9 to 29 dollars a month it is priced for individual engineers and small teams to expense without asking. Distribution: the MCP ecosystem lists and directories, Claude and Cursor community Discords, and SEO on mcp server monitoring while the term has zero competition.

MVP scope

• Register servers with transport and auth config
• Cron checks that perform a real MCP handshake: initialize plus tools/list
• Up/degraded/down states with uptime history and latency
• Tool-schema drift alerts with before/after diff and acknowledge-to-rebaseline
• Email and Slack alerting, Stripe-billed tiers
• Skip for v1: stdio-transport servers, invoking tools with test payloads, public directory of server statuses, multi-region probes, SLA reports

Build MCPWatch, an uptime and schema-drift monitor purpose-built for MCP servers. Stack: Next.js on Vercel, Supabase for auth and data, Vercel Cron for scheduled checks, Stripe checkout for paid tiers.

Core flow: user signs in with magic-link auth and registers MCP servers to watch: URL, transport (Streamable HTTP or SSE), optional auth header, check interval. On each cron tick a checker speaks actual MCP, not a bare HTTP ping: open the transport, send an initialize request with a current protocol version, confirm a valid initialize result, then call tools/list and store the full tool list with each tool's name, description and JSON Schema. Health states: up, degraded (initialize succeeds but tools/list fails or takes over a threshold), down.

Drift detection is the differentiator: diff the returned tool schemas against the stored snapshot and alert when a tool disappears, a parameter is added or removed, a type changes, or a description changes materially. A renamed or vanished tool means agents silently lost a capability, which is exactly the failure users cannot see. Alerts go to email and Slack webhook with a rendered before/after schema diff; the user acknowledges a drift to accept the new snapshot as baseline.

Pages: landing explaining the silent-degradation problem, dashboard with per-server status cards and uptime bars, server detail with check history and drift timeline, public status page per server (optional, shareable), billing.

Store snapshots normalized (sorted keys) so cosmetic reordering never fires an alert.

Pricing: free 2 servers hourly; 9 GBP/month 10 servers; 29 GBP/month 50 servers at 5-minute intervals.