Headline stats

The index crossed 25,000 repositories this month. That number represents every GitHub repository that matches searches for mcp server, model context protocol, agent tool, and a2a agent — crawled nightly, scored on five signals, ranked daily.

The average score of 29.6 tells the real story: 63% of indexed repos score under 30. Most of what's on GitHub is experimental, abandoned, or early-stage. The top 1% — 40 Elite-tier tools with scores above 90 — captures the bulk of real ecosystem activity.

Top 10 leaderboard

The full leaderboard at agentrank-ai.com updates daily. These are the top 10 as of March 18, 2026, ranked by composite AgentRank score (stars 15%, freshness 25%, issue health 25%, contributors 10%, inbound dependents 25%).

CoplayDev/unity-mcp holds the top score at 98.67 — a community-built Unity Editor bridge with 43 contributors and 7,003 stars. No corporate backing. Highest score in the entire index.

PrefectHQ/fastmcp is the most popular tool in the top 10 by raw star count (23,659) and has the deepest dependency graph — 7,718 other repositories depend on it. That inbound-dependent signal is the strongest single indicator of ecosystem adoption in the scoring model.

mark3labs/mcp-go is the infrastructure pick: 8,353 stars, 170 contributors, 2,938 dependents. If you're building Go tooling on top of MCP, this is the default SDK.

Biggest movers this week

The AgentRank index takes daily score snapshots. These are the five tools with the largest score improvement from March 16 to March 17 — measured by the composite score delta, not just star count.

homeassistant-ai/ha-mcp is the standout mover: it jumped 56 rank positions in a single day (rank 72 to 16), driven by a surge in issue closures and a contributor count increase. Home automation is an underappreciated MCP category — this is the tool to watch in April.

rekog-labs/MCP-Nest (NestJS framework for building MCP servers) also gained 57 rank positions. The NestJS community is a late but fast adopter — TypeScript-native, opinionated, high issue-health baseline.

Weekly movers are posted every Monday. Follow @AgentRank_ai for the live feed, or check the trends dashboard for the full moving-window chart.

Skills leaderboard

Beyond GitHub repos, AgentRank indexes 3,124 agent skills from three registries: skills.sh (2,029 skills, 64.9%), Glama (1,023 skills, 32.7%), and ClaWhub (72 skills, 2.3%). Total recorded installs across all tracked skills: 17,816,525.

Microsoft's Playwright MCP has 1,777,315 Glama installs — that's more installs than the next 20 skills combined. Browser automation through MCP is the dominant use case for production agent deployments right now. This isn't surprising: web interaction is the most common task a general-purpose agent needs, and Playwright is the production-grade tool for it.

MongoDB's MCP server ranks #5 with 386,047 installs despite only 959 GitHub stars — a stark reminder that install counts and star counts measure different things. Enterprise tooling gets deployed quietly.

Language distribution

Python leads at 38.5% but TypeScript is gaining. The top tier of the leaderboard is increasingly TypeScript-heavy — 6 of the top 10 repos are TypeScript or C# (both compiled, typed languages). Dynamically-typed Python still owns the experimentation layer.

Go's 4.8% share understates its influence. The two most-depended-on tools in the index (mark3labs/mcp-go and github/github-mcp-server) are both Go. When production teams want a stable, low-overhead MCP server, Go is the answer.

Rust's 2.6% produces the highest-quality tools per repo. Rust MCP servers are statistically more likely to appear in the Elite score tier than tools in any other language. See our best Python MCP libraries and FastMCP tutorial for language-specific guidance.

Ecosystem health signals

The 15.6% freshness rate (repos with a commit in the last 30 days) is the number to watch month over month. The absolute count matters more than the percentage: 3,995 actively maintained tools is substantial production coverage for an ecosystem that didn't exist 16 months ago.

The 54.4% of repos last committed 91–365 days ago represents the bulk of the "long tail" — tools built for specific use cases, used quietly, and rarely updated. This is normal. Not every MCP server needs weekly commits. The signal that matters is whether issue health is high (problems get resolved when they arise) and whether the core functionality is stable.

Archived repos: 353 (1.4%). The graveyard rate is remarkably low. Most abandoned MCP projects are left in a limbo state (last commit 6–12 months ago, no archived flag) rather than formally archived. The 1.4% number undercounts true abandonment but the ecosystem churn rate is genuinely lower than most developer tooling ecosystems at this age.

Score tier breakdown

The score distribution is heavily right-skewed. That's healthy: a small number of well-maintained tools doing the work for a much larger number of experiments is exactly what a maturing ecosystem looks like. The question for April is whether the Medium tier (2,357 repos) can graduate more tools into the High tier as maintainers respond to issues and push commits.

What to watch in April 2026

1. Home automation MCP goes mainstream

homeassistant-ai/ha-mcp's 56-rank jump in a single day signals that the home automation community is arriving in force. Home Assistant has a large, highly engaged developer base. Expect more home automation MCP tools to enter the top 50 as the HomeAssistant community formalizes its MCP integration patterns.

2. NestJS / framework-native MCP gains traction

rekog-labs/MCP-Nest entering the top 25 confirms that enterprise TypeScript teams want opinionated framework integration, not raw SDK use. Watch for similar libraries for Spring Boot, Django, and Rails as framework-native MCP adapters become the preferred enterprise entry point.

3. A2A tool count crosses 200

Currently 128 repos in the index match A2A agent queries — up from near-zero at the start of Q1. The pace is accelerating. April will likely see A2A cross 200 indexed repos as the spec stabilizes and early adopters publish reference implementations. See our A2A vs MCP comparison for context on where the two protocols diverge.

4. The skills install concentration problem

Microsoft Playwright MCP's 1.77M installs represent more than 10% of all skill installs across the entire index. As MCP moves into production, this concentration will increase. The April question: does the second tier (MongoDB at 386K, browser-use at 49K) close the gap, or does Playwright extend its lead further? Browser automation and database access are the two highest-frequency agent tasks — the tools that win those categories will dominate the install curve.