How grading works
Every listing is evaluated by a program, not by its publisher. The grade is the share of checks passed, out of 100. It's run again every week and whenever a new version appears.
What we check
MCP servers
We read the server.json in the official MCP Registry's format. For hosted servers we connect for real: we open a session, time the handshake, list the tools and prompts, and check that each tool is described and says whether it changes data. A server that asks for sign-in counts as up; its tools are listed once its publisher gives us access. For servers you run yourself we check the package exists at the declared version in npm, PyPI or Docker Hub. We never run them.
Skills
We read SKILL.md, check its frontmatter against the Agent Skills format, check that the description says when to use it, that linked files exist, and list any scripts it could run.
Plugins
We read the marketplace entry and the plugin's plugin.json, and list what it adds: commands, agents, skills, MCP servers and hooks. Hooks run shell commands on your computer automatically, so we read every hook and the scripts they call.
For everything
We scan all text the model will read (tool descriptions, server instructions, skill instructions) for prompt injection: attempts to override instructions, hide actions from you, reach for credentials or send data elsewhere. We scan scripts and hooks for dangerous commands, like piping a download into a shell. We also look at the source repository: is it public, licensed and maintained.
Critical findings
Some checks are critical: prompt injection or dangerous commands, a hosted server on plain http, or a manifest we can't read. One critical failure caps the grade at F and holds the listing for review: it disappears from search and rankings until a moderator or the publisher resolves it. The finding and the exact text that triggered it are shown on the listing's page.
How ranking works
Each kind (servers, skills, plugins) is ranked on its own. The overall score blends four things:
| Grade | 40% | The latest evaluation. Listings not yet evaluated start low. |
|---|---|---|
| Reviews | 25% | Average rating, counted as if every listing started with 5 reviews of 3.6 stars, so a few early 5-star reviews can't beat many 4.6s. |
| Popularity | 25% | Installs and saves here, GitHub stars and package downloads, on a log scale, as a percentile within its kind. |
| Maintenance | 10% | How recently it had a commit or a release. |
Trending is separate: installs and views in the last 7 days against the 4 weeks before. Installs and views count once per visitor network per hour, so they can't be pumped by refreshing.
Where listings come from
MCP servers are imported hourly from the official MCP Registry. Publishers also submit servers, skills and plugin marketplaces directly. To publish under a name, you prove you own it, the same way the official registry works: sign in with GitHub for io.github.you/…, or add a DNS record for your own domain.
Waypoint serves its MCP servers in the registry's own API format too. See the API.