How we score

Scores are relative: each signal becomes a percentile rank among all tracked tools (0 = lowest, 100 = highest). A score of 40 means below the median of an already strong set, not “bad”.

Ranks are shown as “#x of N overall · #y in category”. We say “highest scored”, never “best”: a score summarises public signals, not fitness for your project. Tools are ordered by the unrounded score; equal scores are ordered by safety label (No issues found, then No dependencies scanned, then Caution), then stars, then name. In the Top 10 AI use cases, ties are broken by direct evidence, then publishers, then cited findings, then title A to Z, so no rank is shared.

All data comes from public sources: the GitHub REST API, raw README files, the Hugging Face API and official model configs. Every tool page lists its sources and fetch time. We don't test tools hands-on.

Separate from the score. Each repo is shallow-cloned and scanned automatically every week. Each label shows the commit and date scanned. This is an automated scan, not a manual audit.

• Runtime dependency vulnerabilities: lockfiles and manifests checked against the OSV database with osv-scanner. Dev, test and docs dependencies are excluded (manifests under docs/, examples/, tests/, website/ and similar folders, requirements-dev/test/docs files, npm packages marked dev in package-lock.json and direct devDependencies). Transitive dev dependencies in non-npm lockfiles may still be counted.
• Install hooks: npm preinstall/install/postinstall scripts, flagged if they download remote content.
• Remote code execution patterns like eval(fetch(...)).
• Agent instruction files (SKILL.md, skills/, CLAUDE.md, AGENTS.md, .cursorrules): hidden or bidi unicode, long encoded blobs, instructions to pipe a remote script into a shell, and injection-like phrases (listed for review only).
• MCP server code: whether it spawns processes, reads or writes files, or uses the network.

✕ Risk found only on concrete evidence: a dependency flagged malicious in OSV (MAL- advisory) or a human-confirmed malicious finding. Tools with this label are never ranked. ▲ Caution: critical runtime dependency vulnerabilities, downloading install hooks, remote-code patterns, hidden content in instruction files, instructions to pipe remote scripts into a shell, or an archived repo. ✓ No issues found: none of these were found by the automated scan, and runtime dependencies were checked. □ No dependencies scanned: none of the code and instruction-file checks found anything, but osv-scanner found no runtime lockfile or package manifest (or only dev, test or docs ones), so dependency vulnerabilities could not be checked. It does not change the score.

Notes (shown separately, never change the label): license not machine-detectable by GitHub, telemetry or analytics references in code, and vulnerabilities that only affect dev/test/docs dependencies.

1. We collected findings from published usage studies and surveys, preferring 2025 and 2026. Every finding below links to its source, and a finding without a link was left out.
2. Each finding was mapped by hand to the AI use cases on this site. A finding is "direct" when it measures that very job (for example debugging code) and "adjacent" when it measures a nearby job (for example AI writing in general, for our research-backed article use case). That mapping is our editorial judgement and the one place where we could be wrong.
3. Scoring is deliberately simple: 2 points for direct evidence and 1 point for adjacent evidence, counted once per publisher, so one organisation cannot add points by publishing twice. Ties are broken by direct evidence, then number of publishers, then number of cited findings, then title A to Z, so every rank is unique. The cited-findings step only orders items with equal points; points still count each publisher once.
4. Only use cases with at least two points are eligible. We do not convert shares from different studies into one number, because the studies measure different things (messages, conversations, developers, companies).