Methodology: the public-signal score (v0.1)
The score is a transparent summary of public signals. It is not a hands-on test and does not say a tool works, is safe or is right for you. Every number is computed by a script from the data on each listing page, and the same script runs for every listing.
What we measure
| Criterion | Max | How points are earned |
|---|---|---|
| Maintenance | 30 | Last push to the repository: ≤30 days = 30, ≤90 = 22, ≤180 = 12, ≤365 = 5, older = 0. Archived or deprecated repositories get 0. |
| License clarity | 15 | GitHub detects a standard open-source license (MIT, Apache-2.0, BSD, MPL, GPL family, ISC and similar) = 15. A license file GitHub can't classify = 8. None detected = 0. |
| Documentation | 20 | README length (≥6,000 characters = 10, ≥2,000 = 7, ≥500 = 4) + a working official site or docs link (5) + an install, quick-start or usage section (5). |
| Adoption | 20 | GitHub stars: ≥50k = 20, ≥10k = 17, ≥3k = 14, ≥1k = 10, ≥300 = 6, otherwise 3. Stars measure attention, not quality, and can be inflated. |
| Client compatibility | 15 | Number of agent clients named in the README (Claude Code, Claude Desktop, Cursor, VS Code, Codex, Gemini CLI, Windsurf, GitHub Copilot, Cline, Zed, OpenCode): ≥5 = 15, 3–4 = 11, 2 = 8, 1 = 5, none = 0. Not applied to agent clients, frameworks and specifications themselves; their score is computed out of the remaining 85 points. |
The final score is points earned ÷ points available × 100, rounded. Bands: 90+ Strong signals, 75–89 Solid, 55–74 Mixed, below 55 Thin.
What the score does not tell you
- Whether the tool actually installs and works. That needs a hands-on test.
- Whether it is secure. We do not audit code or permissions. Read what an MCP server or skill can access before you install it.
- Whether it is the best fit for your workflow.
“Works with” badges
A badge means the project's own README names that client. It is a claim by the project, found by keyword match, and is not verified by hand. Keyword matching can produce false positives, for example a README that says a tool is not for a client.
Hands-on agent tests (planned, not yet live)
The next layer is a sandboxed test: install the tool from its own instructions, trigger it from an agent, run a small task, and compare with a no-tool baseline. Results and logs will be published on the listing and labeled separately from the public-signal score. No listing on this site has had a hands-on test yet, and each listing page says so.
Data and freshness
GitHub data (stars, license, last push, language) and README text were pulled on Oct 30, 2026. Official-site links were checked for a successful response on that date. Commercial products without a public repository are listed but not scored, and their descriptions come from the vendor's own site. The full dataset is at /api/listings.json.
Independence and conflicts of interest
Sponsors can buy a clearly labeled placement. They cannot buy, change or remove a score, and sponsored placements sit outside the ranked lists. At launch there are no sponsors and no affiliate relationships. If that changes, each affected listing will say so. Spotted an error? Corrections are welcome and will be logged here once the contact channel is live.