Ecosystem overview

The Model Context Protocol was announced by Anthropic in late 2024. As of March 2026, there are 25,632 indexed repositories on GitHub that match MCP-related search queries — servers, SDKs, frameworks, and client implementations. An additional 3,124 skills are indexed from registries including Glama and skills.sh.

Of those 25,632 repositories, 96.8% are classified as MCP servers or server-adjacent tools. The remaining 3.2% (733 repositories) are A2A (agent-to-agent) tools and frameworks. The MCP ecosystem is enormous compared to A2A right now. That gap will close as A2A matures.

353 repositories (1.4%) are archived — abandoned, deprecated, or superseded. This is a low churn rate for a young ecosystem. Most experiments stay up even when unmaintained. They're noise in the index; our scoring penalizes them heavily.

Language breakdown

Python and TypeScript together cover 65.8% of the ecosystem. This is unsurprising — both have strong LLM/AI communities and Anthropic published official SDKs for both. Go is punching above its weight at 4.8%, driven largely by mark3labs/mcp-go (#4 overall) and the official Go SDK from Google.

Rust's 2.6% share is notable. Rust MCP implementations are typically high-quality and performance-focused — they disproportionately appear in the high-score tier. PHP is underrepresented except for laravel/boost, which is the #3 ranked tool overall.

Freshness & activity

One of the five signals in the AgentRank score is freshness — days since the last commit. The distribution tells an important story about ecosystem health.

54.3% of repos haven't been touched in 91–365 days. This is the long tail of experiments that were started, proven out (or not), and left standing. Not abandoned — just stable or stalled.

24.4% of repos saw commits in the last 30 days. That's the active core of the ecosystem — roughly 6,252 projects being worked on right now. In an ecosystem this young, that's a healthy active-project ratio.

Year-over-year context: 38.7% of repos have 2026 commits, meaning nearly 2 in 5 projects are in active development this year. Another 60.2% last committed in 2025. The ecosystem is fresh — less than 1.1% of repos have commits predating 2025.

Score distribution

The AgentRank score is a composite of five signals: stars (15%), freshness (25%), issue health (25%), contributor count (10%), and inbound dependents (25%). Scores run 0–100. The distribution is heavily skewed toward the low end — most tools are young, solo, and under-used.

Only 152 tools (0.6%) reach the Elite tier (80+). These are the tools with strong community adoption, active maintenance, responsive issue management, and meaningful dependents. Getting into this tier isn't just about stars — microsoft/playwright-mcp has 28,000+ stars and scores 94.81, but a solo side project with 1,400 stars and excellent issue health (like microsoft/azure-devops-mcp at 97.17) can outscore it.

The Medium tier (40–59) holds 24.9% of the ecosystem — tools that have found some users and are maintained but haven't hit critical mass. This is the most interesting cohort: tools that are genuinely useful but underexposed.

Top 10 by AgentRank score

These are the ten highest-scoring repositories in the index as of March 2026, ranked by the composite AgentRank signal. Stars alone don't determine rank — freshness, issue health, and dependents matter equally or more.

A few things stand out here. microsoft/playwright-mcp (#7, score 94.81) is the most-installed MCP tool by far at 1.77 million Glama installs — but it ranks #7, not #1, because CoplayDev/unity-mcp has superior issue health and freshness signals. Stars and installs alone don't win the index.

Two Microsoft repos appear in the top 10 — azure-devops-mcp and playwright-mcp. Microsoft's investment in MCP tooling is paying off in both stars and quality signals.

PHP gets a top-3 entry via laravel/boost. The Laravel community adopted MCP fast and the official Laravel team's involvement gives it strong maintenance signals.

Most starred repositories

Stars lag real quality signals — they measure historical visibility, not current maintenance. The two most-starred repos are aggregator lists, not tools. Worth knowing before you pick something based on star count alone.

The delta between stars and score is telling. punkpeye/awesome-mcp-servers has 83,027 stars but scores only 71.48 — it's a curated list, not an active tool. github/github-mcp-server has 27,865 stars and scores 85.44 because it's actively maintained by GitHub. Use the AgentRank score, not star count, when choosing a tool to depend on.

The solo-builder problem

The single most striking statistic in the dataset: 97% of all indexed repositories have exactly one contributor.

This is the bus factor problem at ecosystem scale. 24,862 repos are one person going on vacation away from going unmaintained. It's not a criticism — weekend MCP projects have enormous value. But it's a reason to weight the contributor signal in ranking. A tool with 5+ contributors is statistically much more likely to still be maintained in 6 months.

The 240 repos with 10+ contributors represent the institutional and corporate-backed tools — Microsoft, Google, Anthropic, PrefectHQ, and similar. These are the ones most likely to be safe production dependencies.

Licenses

License-wise, the ecosystem is mostly open: 47% MIT, 7.9% Apache-2.0, giving you ~55% permissively licensed tools. The interesting number is 37.6% with no license at all — technically all-rights-reserved under copyright law, even if publicly posted on GitHub. For production use, check the license before you integrate.

Trends & takeaways

1. The ecosystem is enormous but thin

25,632 repos sounds large. But 77% have fewer than 10 stars, 97% are solo projects, and the average score is 29.7 out of 100. This is a wide, thin ecosystem — lots of experiments, not many production-grade tools. The elite tier (80+ score) has 152 entries. That's your actual shortlist of dependable tools.

2. Python wins, TypeScript is close

Python's 38.5% lead reflects the ML/AI community's home language. TypeScript at 27.3% reflects the web development community adopting MCP fast. Both are well-served by official SDKs from Anthropic. Go is the language to watch — official SDK, strong community tooling (mcp-go), and corporate backing.

3. Freshness is the key differentiator

More than half the ecosystem (54.3%) hasn't seen a commit in 91–365 days. If you're picking a tool to build on, freshness is the single most important signal. An actively maintained tool with fewer stars beats a starred-and-stale one every time.

4. Corporate-backed tools dominate the top tier

Microsoft (2 entries), Google (go-sdk), Laravel Foundation (laravel/boost), PrefectHQ (fastmcp), and NVIDIA (NeMo) all appear in the Elite tier. Corporate backing correlates strongly with multi-contributor, actively-maintained tools. Not always — unity-mcp (#1 overall) is community-built — but the correlation is real.

5. The install gap is massive

microsoft/playwright-mcp has 1.77 million Glama installs. The #2 skill (emerzon/mt-data-mcp) has 1.25 million. The median Glama install count is in the hundreds. The top 5 tools have more installs than the next 3,000 combined. Network effects in tool adoption are extreme — first-mover advantage is everything right now.