mcpcheck · weekly or nightly checks for MCP publishers
Your server broke on Windows, in Cursor, or on OAuth refresh last night. We tell you before your users do.
Enroll a remote Streamable HTTP URL or an npx/uvx package. On Server, every week we clean-install it on three operating systems and diff the tool catalog. On Server Pro we do that every night, and also connect with the clients your users run, check OAuth end to end and score error messages. On both plans, a remote server's tasks and subscriptions/listen support is checked against MCP 2026-07-28, with the date each check has passed since. A badge on your README says exactly what passed.
Badge example: 
What the checks cover
Fresh npm/uv cache on GitHub-hosted Windows, macOS and Linux runners; spawn over stdio (or connect to your URL), initialize, tools/list. The Windows npx ENOENT that broke six official plugins shows up here first.
Official TypeScript and Python SDK clients at the current and previous minor, plus a scripted Claude Code connection check. Cursor and VS Code embed the TypeScript client, so the badge says what actually works.
401 challenge, protected-resource metadata, authorization-server metadata, PKCE S256, dynamic client registration, refresh grant, resource indicator — each failure with the fix in plain English. Live refresh when you give us a refresh token.
tools/list hashed every run and diffed against your baseline: removed tools, renamed arguments, tightened schemas, changed descriptions. Rug-pull watch for your own server, alerted by email/Slack/Discord.
10 malformed inputs per tool, every error graded on code · cause · fix · retryable. The one place we use an LLM (≤ $0.02 per server per run); the score is on the report.
For remote servers on MCP 2026-07-28: 9 tasks-extension checks and 3 subscriptions/listen checks, run when server/discover advertises them. Each quotes its rule from the spec, names the official conformance check it mirrors, and shows “conformant since” a date that carries over from run to run. No tool is called and nothing is created on your server.
CLI paths only — no GUI automation. Results on a no-JavaScript report page, report.json, trust.json and the badge. Runner: public GitHub Actions repo agentwares/runner.
Pricing — per server
- one full baseline run
- public report + badge
- no schedule, no alerts
- weekly runs
- clean install matrix
- catalog diff
- badge
- tasks and subscriptions conformance
- 30-day history · alerts
- nightly runs
- clean install matrix
- catalog diff
- badge
- client compat matrix
- oauth conformance
- error ergonomics
- tasks and subscriptions conformance
- 90-day history · alerts
Machine-readable: /mcpcheck/pricing.json · for agents: /mcpcheck/llms.txt.