How do I know if an MCP server is well-maintained?

Check four signals on GitHub: (1) last commit date — anything over 90 days idle is at risk of drift; (2) issue close rate — divide closed issues by total; above 70% indicates an active maintainer; (3) open issue count and age — many old open issues signals backlog neglect; (4) contributor count — single-maintainer repos carry higher bus factor risk. AgentRank computes these automatically for 25,000+ servers.

What is a good AgentRank score for an MCP server?

The average score across all 25,000+ indexed tools is 29.6 out of 100. Anything above 70 is well above average. The top 10% start around 60. The top 1% start around 85. For production use, targeting servers with scores above 70 and individual issue health above 80% is a reasonable baseline.

Why is freshness an important signal for MCP servers?

MCP is a protocol that evolves. Servers that haven't been touched in six months may not track protocol updates, API changes in upstream services, or security patches. A server with a recent commit is more likely to work correctly with current versions of Claude, Cursor, and other MCP clients than one that went stale in 2024.

Should I trust a high-star MCP server?

Stars measure historical interest, not current quality. A server can accumulate stars from an initial viral moment and then go unmaintained for a year. Before installing any high-star server, check freshness (last commit) and issue health (close rate). In the AgentRank index, the most-starred server (microsoft/playwright-mcp, 28,849 stars) ranks 8th by composite score — not first.

How to Evaluate MCP Server Quality Before Installing

March 19, 2026 · 9 min read · Developer Guide

Most developers pick MCP servers by star count or recency on a discovery page. Both are weak signals. A server can go viral on Reddit, collect 3,000 stars, and then sit untouched for 14 months while the MCP protocol moves on. This guide explains the five real signals that predict whether an MCP server will work reliably in your stack six months from now — and how to check them in under two minutes.

Why stars alone fail you

Stars measure a moment in time — the moment a repo went viral, got shared on Hacker News, or was featured in a newsletter. They don't update as the project decays. A server with 4,000 stars from a 2024 launch that hasn't seen a commit since August 2024 will rank at the top of any star-sorted list — and will break against current MCP client versions, upstream API changes, and protocol updates that happened in the intervening months.

In the AgentRank index of 25,000+ servers, the correlation between stars and composite quality score is weak. microsoft/playwright-mcp has 28,849 stars and ranks 8th by composite score. CoplayDev/unity-mcp has 7,003 stars and ranks 1st. Stars capture awareness. Quality signals capture reliability.

The question you actually need answered before installing an MCP server is not "how popular was this when it launched?" It's "will this still work three months from now?"

The five quality signals

These are the signals AgentRank weights in its composite score. They're also the signals you should check manually when evaluating a server not in the index.

1. Freshness 25% of AgentRank score

Days since the last commit. This is the most actionable signal. A server with a commit in the last two weeks is almost certainly tracking protocol changes. A server last touched 14 months ago is not.

Good Commit in the last 30 days

Acceptable 30–90 days, stable protocol-only server

Red flag 90+ days, especially for servers wrapping external APIs

Why it matters: External APIs change. The MCP protocol itself evolves. MCP clients release updates. A server that hasn't been touched since before the latest protocol version may have silent compatibility issues you'll only discover in production.

2. Issue health 25% of AgentRank score

Closed issues divided by total issues. This is the best proxy for maintainer responsiveness. A 90% close rate means the maintainer actively triages their issue queue. A 30% close rate on 200 issues means there are 140 unaddressed problems.

Good >80% close rate

Acceptable 60–80%, larger scope projects often run lower

Red flag <50% with more than 20 open issues

Why it matters: If you file a bug, will anyone respond? A low close rate doesn't just mean the project is slower — it means the maintainer isn't actively engaged. When things break (and they will), you'll be on your own.

3. Inbound dependents 25% of AgentRank score

How many other repositories depend on this one, via GitHub's dependency graph. This is the strongest signal of real-world adoption beyond starred interest. Stars are passive — someone liked the repo. Dependents are active — something is actually running that depends on this.

Good 10+ dependents (for a framework-type server)

Acceptable 0–10, normal for most specialized servers

Red flag 0 dependents on a server claiming widespread use

Why it matters: High dependent counts mean other developers and projects have integrated this server into their own software. That creates accountability — breakage affects downstream users, increasing the incentive to maintain. Low dependent counts don't disqualify a server, but combined with low freshness, it signals lower stakes for the maintainer.

4. Contributors 10% of AgentRank score

Number of unique contributors. This is a bus factor proxy. A server with one contributor goes dark the day that person loses interest. A server with 20+ contributors has distributed maintenance responsibility.

Good 5+ contributors (broad community ownership)

Acceptable 2–5, still distributed enough to reduce single-point risk

Watch 1 contributor — not disqualifying, but increases abandonment risk

Why it matters: Single-maintainer projects are fragile. If that person changes jobs, loses interest, or goes on a long vacation, the project stalls. 10% weight reflects that this is real but secondary to maintenance activity signals.

5. Stars 15% of AgentRank score

The raw GitHub star count, normalized against the index. Stars get only 15% weight because they measure past awareness, not current quality. They still matter — popularity correlates with more eyes on the code, more bug reports getting filed, and more community investment. They just shouldn't dominate.

Good 500+ stars in a 25,000-server index signals real traction

Acceptable 50–500, normal for niche but legitimate tools

Weak signal Stars alone, without checking the other four signals

Why it matters: Popularity correlates with visibility of defects — more users means more bug reports and fixes. But it's a lagging indicator. Don't use it alone.

Red flags to walk away from

These patterns, individually or together, predict unreliable or abandoned servers:

Last commit 6+ months ago — especially for servers wrapping external APIs or the MCP protocol itself
Issue close rate below 40% with more than 15 open issues — indicates an overwhelmed or disengaged maintainer
Open issues older than 6 months with no response — maintainer has stopped engaging
No tests in the repository — higher risk of undetected regressions during updates
README mentions "coming soon" features from a year ago — development has stalled
Multiple open issues about the same bug — indicates the maintainer knows about a problem and hasn't addressed it
Last version published to npm/PyPI was pre-MCP 1.0 — may have silent protocol incompatibilities

Green flags that predict reliability

These patterns indicate a server worth trusting in production:

Issue close rate above 85% — the maintainer actively triages and resolves reports
Commits in the last 14 days — active development, will track protocol changes
Official backing — the organization that owns the underlying platform maintains the server (e.g., microsoft/playwright-mcp, laravel/boost, perplexityai/modelcontextprotocol)
100 or more GitHub stars with 90%+ close rate — strong combination of traction and maintenance discipline
Changelog or release notes updated recently — maintainer communicates changes
Published TypeScript types or Python types — investment in DX and compatibility
Listed in the official MCP registry — explicit quality bar from the protocol team

Using AgentRank to evaluate instantly

Rather than checking each signal manually on GitHub, AgentRank computes the composite score nightly for every indexed server. Search by name, browse by category, or filter by score threshold.

Each tool page shows:

Composite score (0–100)
Individual signal breakdown: freshness score, issue health score, dependent count, star count, contributor count
Last crawl date and raw GitHub metadata
Category, language, and license

For any server in the index, you can see at a glance whether the quality signals justify installation — without opening GitHub, calculating close rates, or checking commit dates manually.

If a server isn't in the AgentRank index yet, you can still check the five signals manually on its GitHub repo in under two minutes. The calculation is:

Issue close rate = closed_issues / (open_issues + closed_issues)
Freshness = days since last commit
Dependents = GitHub "Used by" count on repo page
Contributors = GitHub "Contributors" count
Stars = raw star count

No formula required — just a quick scan of the repo page tells you most of what you need.

Worked example: two servers head-to-head

Say you're looking for an MCP server to provide web search to your agent. You find two candidates: Server A has 4,200 stars. Server B has 800 stars. Most developers would pick A.

Signal	Server A (4,200 stars)	Server B (800 stars)
Last commit	11 months ago	6 days ago
Issue close rate	38% (62 open, 38 closed)	94% (3 open, 47 closed)
Contributors	3	8
Dependents	12	4
Stars	4,200	800
Recommendation	Skip — stale, unresponsive maintainer	Install — active and well-maintained

Server A won the star race once. Server B is being actively maintained now. For a server that wraps a live search API — where the upstream API can change, rate limits can change, and MCP protocol updates affect behavior — Server B is clearly the safer dependency. Stars got it wrong. The other four signals got it right.

This pattern repeats throughout the AgentRank index. The highest-starred server (microsoft/playwright-mcp, 28,849 stars) ranks 8th. The top scorer (CoplayDev/unity-mcp, 7,003 stars) has less than a quarter of Playwright's star count but better maintenance signals across every other dimension.

Check any MCP server: Search agentrank-ai.com — composite scores, signal breakdowns, and nightly-updated data for 25,000+ servers.

Scoring methodology: Freshness 25%, issue health 25%, dependents 25%, stars 15%, contributors 10%. Full docs here.

API access: Query the index programmatically with score filters. API docs.