Can agents actually use it?
Readiness scanners check whether a site has the right files. We send four agents (Claude, GPT, Gemini and an open-weight model) through the same real tasks and rank by how often they finish the job. Last updated October 30, 2026.
Public MCP servers
Servers anyone can connect to without signing in. Agents use the server’s own tools.
| # | MCP server | Score | Sessions passed | Agents | Biggest problem |
|---|---|---|---|---|---|
| 1 | Shopify Dev mcp.shopify.com | 91 | 20 of 24 | Best: Gemini 4 Argon 96 Worst: DeepSeek V4.1 Flash 84 | generate-business-names and generate-domain-names embed model-directed instructions and promotional prose in the result payload |
| 2 | Microsoft Learn learn.microsoft.com | 90 | 18 of 24 | Best: GPT-6.1 Sol 99 Worst: Gemini 4 Argon 81 | microsoft_docs_search ranking: wrong or non-canonical pages surface first for task queries (MFA admin, support request, quickstarts) |
| 3 | Hugging Face huggingface.co | 90 | 18 of 24 | Best: GPT-6.1 Sol 99 Worst: DeepSeek V4.1 Flash 73 | hub_repo_search returns empty or near-empty results for text-classification + language:en filter combos, with no hint or documented filter syntax |
| 4 | DeepWiki mcp.deepwiki.com | 86 | 17 of 24 | Best: DeepSeek V4.1 Flash 95 Worst: GPT-6.1 Sol 73 | ask_wiki_question returns stripped or malformed citations and no source file paths |
| 5 | GitMCP gitmcp.io | 84 | 20 of 24 | Best: Claude Sonnet 5.5 88 Worst: DeepSeek V4.1 Flash 78 | search_generic_code returns bare file paths with constant score 1, no snippets, opaque errors, and empty results on large repos |
| 6 | Cloudflare Docs docs.mcp.cloudflare.com | 84 | 17 of 24 | Best: Claude Sonnet 5.5 100 Worst: Gemini 4 Argon 69 | search_cloudflare_documentation ranks troubleshooting, changelog and tangential pages above canonical how-to/reference pages |
| 7 | AWS Knowledge knowledge-mcp.global.api.aws | 76 | 12 of 24 | Best: GPT-6.1 Sol 81 Worst: DeepSeek V4.1 Flash 67 | aws___search_documentation ranks blogs, FAQs and old SDK pages above canonical how-to pages (S3 versioning) |
| 8 | Context7 mcp.context7.com | 72 | 11 of 24 | Best: Gemini 4 Argon 81 Worst: DeepSeek V4.1 Flash 67 | query-docs returns too few, partial snippets for broad or multi-part queries (Tailwind upgrade guide, Spring Boot Actuator and server.port) |
| 9 | Exa mcp.exa.ai | 65 | 9 of 24 | Best: Claude Sonnet 5.5 78 Worst: Gemini 4 Argon 55 | Server instructions advertise `agent_run` (and sometimes `web_fetch_exa`) but tools/list does not expose them; no bulk enrichment path (T6, T5) |
| 10 | Kiwi.com Flights mcp.kiwi.com | 64 | 15 of 24 | Best: Gemini 4 Argon 96 Worst: GPT-6.1 Sol 25 | search-flight: no current date and no relative-date input, so agents guess past dates or stop to ask the user |
Websites
Agents use each site in a real browser (headless Chrome), signed out.
| # | Site | Score | Sessions passed | Agents | Biggest problem |
|---|---|---|---|---|---|
| 1 | Supabase supabase.com | 82 | 18 of 24 | Best: Gemini 4 Argon 87 Worst: Claude Sonnet 5.5 76 | Sign-up page (/dashboard/sign-up) gives no feedback after clicking Sign up |
| 2 | Stripe stripe.com | 81 | 16 of 23 | Best: Claude Sonnet 5.5 99 Worst: GPT-6.1 Sol 64 | open_page reports harmless background POSTs as BLOCKED and as a CAPTCHA/human check on stripe.com and docs.stripe.com pages |
How these are run
- Six tasks per target, drafted from its public tools or pages: the jobs a real user would ask an assistant to do there. Every agent gets the same tasks, with no hints.
- Read-only and signed out. Nothing that would change data runs: no sign-ups, no form submissions, no purchases. Pages disallowed by robots.txt are not fetched.
- A separate model grades each session against the task’s success criteria. Score: 70% completion, 20% efficiency against the ideal number of steps, 10% safety. Each report shows every step.
- Rankings show a target’s latest complete run. Runs where too many sessions failed to start are left out.
Is one of these yours? Run your own tasks, including signed-in flows, or tell us if something here looks wrong.