The problem with sorting by stars

Open GitHub and search for "MCP server." Sort by stars. The top result has 83,000 stars — it's a curated list of MCP servers, not an MCP server itself. The second result is the official reference implementation repository with 81,000 stars. Both are genuinely useful. Neither tells you which individual tool you should actually build on.

Stars accumulate over time. They don't decay. A tool that was popular in early 2025, went unmaintained, and has been effectively abandoned still carries all those stars. When you're choosing a production dependency, that star count is actively misleading.

The MCP ecosystem has 25,000+ indexed repositories. Of those, 54% haven't seen a commit in the past 90 days. More than 97% are solo-contributor projects — one developer who might have moved on to something else. Sorting by stars surfaces the historically popular; it doesn't surface the currently healthy.

That's the gap the MCP server quality score is designed to close.

The five scoring signals

AgentRank computes a composite score from five GitHub signals. Each is normalized to a 0–1 value, weighted, and summed to produce the final 0–100 number. Here's what each signal measures and why it was chosen.

1. Freshness (20% weight)

How many days since the last commit? This is the fastest proxy for "is anyone home."

A commit within the last 7 days scores 1.0. From days 7 to 90, the score decays linearly. Beyond 90 days, the decay accelerates exponentially — a tool that hasn't been touched in six months can score near 0 on this signal alone. One exception: repos with 200+ stars or 5+ dependents get a freshness floor of 0.3, because stable, widely-adopted tools don't need daily commits to be reliable.

Freshness gets 20% weight because inactivity is often an early signal of abandonment — but it's not the whole story. A tool can be stable and complete without needing frequent commits. That's why it shares weight with four other signals.

2. Issue health (25% weight)

Issue health is closed issues divided by total issues. A maintainer who closes bug reports and resolves feature requests scores well here. One who accumulates open issues without responding scores low.

This is the strongest proxy for maintainer responsiveness. Stars tell you a tool was once useful. Issue health tells you whether the maintainer is still paying attention. It gets the highest single weight (25%) because unresponsive maintainers are the most common source of MCP tool rot.

3. Inbound dependents (25% weight)

How many other GitHub repositories import or depend on this tool? This signal comes from GitHub's dependency graph and is the closest thing to real-world adoption data we have.

A tool with 500 inbound dependents isn't just popular — it's trusted by other developers enough to build on. That's qualitatively different from a star, which requires a single click. Inbound dependents also weight freshness differently: tools with many dependents are unlikely to vanish overnight, so their freshness floor is protected.

PrefectHQ/fastmcp — currently #9 overall with a score of 94.73 — has among the highest inbound dependent counts in the Python MCP ecosystem. Developers building their own MCP servers use fastmcp as a foundation, which creates a compounding signal of trust.

4. Stars (15% weight)

Stars still matter — but they only get 15% of the weight. Normalized against the full distribution across 25,000+ repos, a high star count contributes meaningfully to the final score without dominating it. The most-starred repos in the index (microsoft/playwright-mcp at 28,849 stars) score well on this signal. But a less-starred tool with an active maintainer and healthy issues can outscore a viral-but-abandoned one.

5. Contributors (10% weight)

The number of unique contributors to a repository. A solo project has a bus factor of one — if the maintainer disappears, the tool is effectively dead. A project with 5+ contributors is statistically much more likely to survive maintainer turnover.

In the full index, 97% of MCP repos are single-contributor. The 3% with multiple contributors cluster strongly in the high-score tiers. mark3labs/mcp-go (#4 overall, score 96.29) and microsoft/playwright-mcp (#7, score 94.81) both have multi-contributor teams — and it shows in their sustained quality and update cadence.

Minor signals (5% combined)

License quality, documentation presence, and release cadence together contribute 5%. These are tie-breakers more than drivers — but a tool with no license, no README, and no version tags starts the scoring process with a small penalty.

For the full technical specification — exact decay curves, normalization formulas, and how missing data is handled — see the scoring methodology page.

Before and after: what scoring changes

Here is a concrete illustration. Suppose you're looking for a database MCP server. GitHub search sorted by stars returns results based on historical popularity. The AgentRank database category returns results sorted by the composite score.

The difference matters most in the middle of the distribution — tools with between 100 and 2,000 stars. At the top (Microsoft, Anthropic, PrefectHQ), stars and score agree: these are well-maintained flagship tools backed by organizations with incentives to maintain them. At the bottom, both signals agree the tool is probably experimental.

The disagreement happens in the middle. A tool with 800 stars and its last commit in January 2024 will appear high on a stars sort. On the AgentRank score, it falls — because its freshness decayed, its issue backlog grew unclosed, and it never attracted a second contributor. Meanwhile, a tool with 150 stars that ships weekly, closes issues within days, and has three contributors rises in the ranking despite its lower star count.

That reordering — the surfacing of healthy tools that stars bury — is the core value of the score.

Real examples from the index

Why unity-mcp is #1

CoplayDev/unity-mcp holds the top position with a score of 98.67 and 7,003 stars. It's not the most-starred tool in the index (that's punkpeye/awesome-mcp-servers at 83,000), but it earns the top score because it has near-perfect marks across all five signals: committed within days, an issue close rate above 90%, significant inbound dependents, high stars for its category, and multiple active contributors. Every signal fires. That's rare.

Why a 10,000-star tool can score 40

There are tools in the index with 5,000–15,000 stars that score in the 30–50 range. The pattern is consistent: viral launch in 2024, strong star growth, then maintainer moves on. Last commit: 8+ months ago. Issue queue: 40+ unresolved bugs. Contributors: 1. The stars are a historical artifact. The score reflects the current state. If you build on that tool today, you're inheriting those 40+ unresolved bugs with no one to fix them.

The DevOps category surprise

In the DevOps category, the top-scoring tool by AgentRank score is not the one with the most stars. microsoft/azure-devops-mcp (#2 overall, score 97.17) scores higher than several more-starred competitors because it has Microsoft's engineering team maintaining it, ships releases on a regular cadence, and has a growing dependent count as developers build Azure pipelines on top of it. Stars favor early movers. The score favors sustained quality.

What your score means

The average score across all 25,000+ indexed repos is 29.7. That average reflects the reality of the MCP ecosystem: it's wide and thin, full of experiments that haven't yet built a health record. A score above 60 genuinely stands out.