Must read for anyone working with MCP still: here are my condensed learnings from crawling the entire MCP corpus for @toolplex_ai: (Stats as of Dec 2025) State of the MCP Ecosystem 1. Total of 36,039 MCP servers, across 32,762 unique GH repos. 2. The median MCP server has 0 stars. 51% have zero. 77% have less than 10. The ecosystem is mostly experiment projects, tutorials and personal tools. 3. MCP growth exploded in Spring 2025, peaked in June, and cooled off since then. New MCP servers went from 135 / month at launch (Nov 2024) to 5,069 / month in June 2025, and now to 2,093 for Nov 2025. 4. TypeScript dominates. 43% are TS, 20% Python, 16% Python, 5% Go, and remaining are rust and other languages. 5. Half the ecosystem is package managed: 32% published to npm, 13% to PyPi, 4% to docker. The other half require cloning and manual setup. 6. 61% of MCP servers are solo projects with zero forks. 16% have no READMEs. Most are personal projects. 7. The top 50 repos account for 60% of the GH stars. modelcontextprotocol (Anthropic), Microsoft, AWS, and CloudFlare lead the top. But 83% of publishers have only one server. It's a long tail ecosystem. 8. Stdio transport dominates with 85% share. SSE is growing (9%) for remote/hostest. Detected via README declaration. 9. 29% of servers haven't been updated in 6 months. Only 27% touched in the last 30 days. Expect consolidation as abandoned projects fade and winners emerge. 10. Chrome DevTools MCP (github.com/ChromeDevTools/ch…) gained 15k stars in 3 months. The hottest category right now is browser automation and dev tooling. 11. Other growing categories: Memory/Context, Security, Finance / Business. 12. Category breakdown of MCP servers. Note, servers can be tagged in multiple domains so percentages exceed 100%: AI / LLM tooling: 49.3%, Automation / Workflow: 18.9%, Cloud / DevOps 12.8%, Analytics 9.3%, Security 8.1%, Database 6.5%, Messaging 4.2%, Finance 4.1%, Browser automation 3.6%, Blockchain / Crypto 3.0%. #mcp

Will MCP survive? My personal thoughts: Probably yes, but it's still in "promising but unproven" territory. MCP suffers from a few key issues: - MCP month-over-month growth is in decline: Probably saturation from an early builder euphoria. - Protocol is not prescriptive about certain things like auth, discovery, or transport (SSE vs Stdio). Leading users to choose their own best practices. An OAuth spec has been implemented but only applies to SSE (modelcontextprotocol.io/spec…). Most users are on stdio and many stdio servers make network calls. - High abandonment rate: 29% haven't been updated in 6 months, 27% updated in the last 30 days. This is not alarming, it just means most builders probably don't have a project that catches on and they move on to other things. - Abandoned by Anthropic?: Anthropic seemed to have released MCP but moved on to other things like Skills, developer tooling or 1P tool integration in Claude Desktop. IMO the Claude Desktop MCP integration is very ad-hoc -- adding tools is confusing, error prone, and all your added tools still bloat the context. - Poor discoverability: Very few are trying to accurately catalogue and separate the 90% poor quality servers. An official registry was created by anthropic, but IMO it's more of a grassroots / community effort. IMO it still lacks standardizations, safety metrics, classification and categorization and thus will lead to MCP marketplaces still needing to curate a lot of their own signals.

The Model Context Protocol is one of the most powerful new ideas in AI. But it has a major problem. Anyone can publish a tool. Some are malicious, many are broken. There's no trust, no feedback. ToolPlex is my solution. Now in open beta! 🧵 toolplex.ai #ToolPlexBeta #mcp

The ToolPlex beta is open, as of today! 🔍 Curated tools 📈 Evolving playbooks 🤝 Shared agent signals 🤖 A protocol built for real-world execution Try it today → toolplex.ai