The State of the MCP Ecosystem (2025)

Or: Why finding the right AI tool just got a lot harder

The Model Context Protocol launched with a promise: give AI agents a standard way to connect to tools. It worked. Maybe too well.

There are now 600+ MCP servers in the wild. The official registry alone lists 100-200+ production integrations from major companies. And that number is growing every week.

If you're building with AI agents, this is both incredibly exciting and completely overwhelming. There's an MCP server for almost anything you can think of — databases, cloud infrastructure, development tools, smart home devices, crypto wallets, payment processors. The problem isn't lack of tooling anymore. It's discovery.

What We Found

We spent the last few days mapping the MCP landscape. Here's what the ecosystem actually looks like right now:

Top Categories by Volume

1. Development Tools (20%+): GitHub, GitLab, CI/CD pipelines, testing frameworks, code quality tools
2. Databases (15%+): Every major SQL and NoSQL option, plus vector databases and data warehouses
3. Cloud Infrastructure (15%): AWS, Azure, Cloudflare, Alibaba Cloud — full service coverage
4. AI/ML Tooling (10%): Observability platforms, prompt management, evaluation tools
5. Data & Analytics (10%): Business intelligence, metrics, logging systems
6. Financial Services (8%): Payment processors, trading platforms, crypto exchanges
7. Productivity (8%): Project management, docs, calendars (Jira, Notion, Cal.com)
8. Browser/Web Automation (5%): Scraping, testing, automation (Puppeteer, Apify with 6000+ scrapers)
9. IoT/Smart Home (3%): Device control and home automation
10. Crypto/Web3 (6%): Blockchain APIs, DeFi protocols, wallet integrations

The Meta-Connector Trend

Something interesting is happening: companies like Composio (100+ tools) and Paragon (130+ SaaS integrations) are building MCP servers that aggregate other APIs. They're becoming "Zapier for MCP" — one connection gets you dozens of tools.

This is both clever and concerning. Clever because it solves the connection sprawl problem. Concerning because it creates new gatekeepers in what's supposed to be an open ecosystem.

The Gap

Here's what nobody is doing yet: making sense of all this.

The official registry is just a list. If you're building an agent and you need to "let users manage their calendar," you have to:

1. Know MCP servers exist
2. Search the registry manually
3. Compare 3-5 calendar integrations
4. Read docs for each one
5. Test compatibility
6. Pick one

That's not a developer experience. That's archaeology.

What's Missing

The ecosystem needs:

• Category-based browsing — "show me all database integrations"
• Quality signals — which ones actually work? Which are maintained?
• Use-case-driven discovery — "I need my agent to read email" → here are your options
• Agent-type recommendations — coding agents need different tools than research agents
• Dependency mapping — what does this server actually require to run?

What We're Building

This is why we're building forAgents.dev. Not another MCP server. Not another meta-connector. A directory — like npm, but for AI agent tooling.

We're starting with:

• Clean categorization of the MCP ecosystem
• "Best in class" picks for common use cases
• Search that actually understands what you're trying to do
• Real usage data (Phase 1 is live now, measuring what people actually use)

The goal isn't to own the ecosystem. It's to make the ecosystem usable.

Why This Matters

If MCP succeeds, we'll have thousands of servers within a year. Maybe tens of thousands. That's the vision — a rich, open ecosystem where anyone can build integrations.

But richness without discoverability is just noise. We've seen this movie before: npm has 2 million packages, but you don't browse npm, you search Stack Overflow for "best React date picker." PyPI has similar problems. The WordPress plugin directory has 60,000+ plugins and finding good ones is mostly vibes and download counts.

The MCP ecosystem is small enough right now that we can build discovery infrastructure before it becomes overwhelming. That window won't stay open long.

Phase 1 of forAgents.dev is live: https://foragents.dev

We're measuring real usage for the next 2 weeks. If you're building with MCP, try the directory and let us know what's missing. We're learning in public.

— Team Reflectt