I run a free MCP server that checks whether an AI tool is still alive before your assistant recommends it. Built after I found 191 tools in my own catalog that had quietly become something else.

Sharing my work, with scope disclosed up front: this is a hosted remote server (streamable HTTP), free, no auth, read-only. The GitHub repo is documentation, not source.

The story behind it: I maintain a continuously verified catalog of 8,000+ AI tools. Two days ago I scanned all of them for cross-domain redirects and found 191 whose website now silently redirects to a different company. One of them was scoring 73/100 on my own liveness system while its domain redirected to an online gambling site. HTTP 200 does not mean alive. That scan became a public dataset with a DOI, and it is exactly the failure mode assistants hit when they recommend tools from training data.

So the server lets any MCP client check before recommending. Seven tools: check_tool_status, pricing_reality, market_sentiment, find_alternatives, compare_tools, viability_score, deadpool_digest. Every answer carries the date that tool was last verified, so the model can judge staleness itself.

Connect (Claude Code):

claude mcp add –transport http rightaichoice https://rightaichoice.com/api/mcp

Any other client (Cursor, Windsurf, etc):

{ "mcpServers": { "rightaichoice": { "url": "https://rightaichoice.com/api/mcp" } } }

Setup for every client: https://rightaichoice.com/mcp

Honesty rules I hold the server to: unknown tools return "not in our catalog, which is NOT evidence it is dead" instead of a guess. Fuzzy name matches are disclosed so the model can reject a wrong match. Answers are never for sale.

Context for the mortality angle: in my survival study of 2,291 top Product Hunt launches (2023 to 2025), 24.4% of the 2,066 with a determinable outcome are already dead.

Would genuinely value feedback from people running agents: what would you want an assistant to check about a tool before it builds on it?