Five different Jev guard packages showed up in a week. A guard decides whether your agent can do the destructive things, and part of how it decides is just text in a tool description.
They're all on their first version right now, so this is me writing down what they say before any of it changes.
https://t.co/l8tS0iodfR
Building Small Print. It reads the text MCP servers, skills and plugins hand to AI agents, every night, and flags when that text changes. Agents follow it, and a version number won't tell you it changed.
Where it's at this morning:
91,671 servers, skills and plugins listed across 12 sources
88% of them read so far
84,554 changes to that text on record
24 people came by yesterday, most so far
Nobody's signed up yet, so that's this week's problem.
https://t.co/LRxaSBB1UH
@deeepakbagada Makes sense. It's the same reason we keep flow logs on the network side and don't just look at firewall denies.
The deny tells you something tried, the flows tell you what it was doing for the ten minutes before that.
How long do you keep the turns around?
@its_meseba It's called Small Print, https://t.co/3VULO2xmij. Every MCP server and skill hands your agent text it just follows, and that text can change on any update. It tells you when yours changes.
npx [email protected] check --no-upload
shows what you've got installed and sends nothing.
Honestly the thing that's worked best so far surprised me. ChatGPT sends me more people than X does, because it reads the pages and quotes them when someone asks a question.
So I'm mostly writing pages that answer what people actually type, and I put a small MCP server version of it in the official registry so agents can find it too.