Scoring Methodology
Every score on AgentRank is computed from eight verifiable GitHub signals. No black boxes. No pay-to-play. Updated nightly.
Why a score?
GitHub search sorted by stars shows you the most popular tools — not the most trustworthy ones. A tool with 10,000 stars and its last commit in 2023 is not a better choice than one with 500 stars that shipped yesterday. AgentRank separates popularity from health.
The score is a single number — 0 to 100 — that answers the question an agent or developer actually cares about: is this tool actively maintained and likely to work?
The Eight Signals
Each signal is normalized to a 0–1 value, then weighted and summed to produce the final 0–100 score. When a signal has no data (e.g., no registry package to measure downloads), its weight is redistributed proportionally across the remaining signals.
How recently was the repository updated?
Freshness is the #1 predictor of reliability. Stale repos accumulate unpatched bugs and bitrotting dependencies. The MCP protocol itself is evolving fast — a tool that hasn't been touched in six months may not work with current clients.
Does the maintainer respond to bug reports?
A high closed-to-total ratio means the maintainer actively triages reports. A repo with 200 open issues and 10 closed is a red flag — problems are piling up, not getting resolved.
How many other repos depend on this one?
When other real projects import your package, that's real-world validation. It's hard to fake and hard to buy. High dependent counts are the strongest signal that a tool actually works in production.
Weekly installs from npm or PyPI.
Download counts reflect actual usage, not just curiosity. A tool being installed 50,000 times a week is being used in real workflows.
Raw GitHub star count.
Stars are the weakest signal — easy to inflate, often reflecting hype rather than utility. We include them because broad recognition does correlate weakly with quality, but we deliberately weight them lowest among the popularity signals. A tool shouldn't rank #1 just because it went viral on Twitter.
How many distinct contributors has the repo had?
A single-contributor repo has single-maintainer risk. If that person disappears, the project dies. Two or more contributors means bus-factor risk is reduced. We cap the benefit at 20 — beyond that, additional contributors don't meaningfully increase reliability for a tool of this type.
Does the repo have a meaningful description?
A maintainer who can't write three sentences about what their tool does is either in stealth mode or not thinking about users. It's a weak proxy for maintainer investment, but it's measurable and consistent.
Is the tool usable by others?
Permissive licenses maximize the likelihood that the tool can be integrated into any project without legal friction.
The Formula
Base weights (when all signals have data):
When Dependents or Downloads data isn't available for a tool, those weights are redistributed proportionally across the remaining signals. A missing signal is not a penalty — it's just fewer data points.
Live Examples
Here's how the scoring works in practice, using real tools from the index as of March 2026.
Scores high on every health signal. The 97% issue resolution rate is exceptional — most tools sit at 40–60%.
The dependents signal is the key driver here. Nearly 3,000 other repos build on top of mcp-go — real adoption by real developers.
The most widely adopted Python MCP framework. 7,718 dependents across the ecosystem is a strong signal that this has become infrastructure.
301 stars but a score of 39. This is exactly what AgentRank is designed to surface: a project that got initial traction but shows no signs of ongoing maintenance. Stars don't tell you this. The scores do.
Update Frequency
The crawler runs nightly. Every morning, all 25,000+ tools are re-scored from fresh GitHub data. Score and rank changes from the previous day are tracked and surfaced in the Movers feed.
Scores reflect reality as of the previous night's crawl. A tool that ships a critical fix today will see its scores improve within 24 hours.
How AgentRank differs from alternatives
| Source | Approach | What's missing |
|---|---|---|
| GitHub search (stars) | Sort by raw star count | No health signals. Stale viral repos rank above active maintained ones. |
| MCPMarket | Engagement-based ranking (clicks, installs) | No composite score. New tools with no traffic are invisible regardless of quality. |
| PulseMCP | Popularity-only (GitHub stars + social) | No maintenance signals. No issue health. Popularity ≠ reliability. |
| Glama | Security scans per tool | No public composite score. Security-only lens misses freshness, adoption, maintainer responsiveness. |
| Awesome-MCP-Servers | Curated flat list | No scoring. Human-curated means latency. New tools take weeks to appear. |
| AgentRank | 8-signal composite score, updated nightly | No competitors publish a transparent 0-100 composite score. This is the differentiator. |
Open source
The scoring engine is open source. You can read the exact implementation, propose weight changes, or run your own scoring pass against the dataset.