Analytics
Turn the MCP traffic you already serve into a map of what agents use, miss, and ignore.
ContextPie's measurement loop is free of instrumentation work because the requests are the data. Every time an agent calls find_skills, get_skill, or get_skill_file, serving that request records an event. No proxies, no wrappers, no SDK hooks in your agent. The query is the telemetry.
Because progressive disclosure splits behavior into distinct search → load → file_load stages, ContextPie can see not just what agents used but exactly where they stopped.
The four views
1. Used vs dead skills
Which skills agents actually load versus those that sit idle. Action: retire or rewrite dead skills so the registry stays signal, not noise.
2. Discovery gaps
Searches that matched nothing well — low top similarity scores. This is your to-do list of skills agents wanted but you don't have. Action: author the missing skill.
3. Surfaced-but-not-loaded
Skills find_skills returned but agents never loaded with get_skill. The skill is discoverable but the descriptor didn't earn the click. Action: sharpen the description or trigger phrases.
4. Dead reference files
Supporting files bundled with a skill that agents never fetch via get_skill_file. Action: prune the file or fix the SKILL.md reference that should point to it.
From insight back to the editor
| View | Tells you | Drives |
|---|---|---|
| Used vs dead | What's idle | Retire / rewrite |
| Discovery gaps | What's missing | Author new skill |
| Surfaced-not-loaded | Weak descriptors | Edit description |
| Dead reference files | Unused files | Prune / fix links |
Every insight links straight back into the editor for the relevant skill or version, so measuring and fixing are one continuous loop. Not connected yet? See Connecting an agent.