Safe is half the question. What does the server cost your agent?
Every MCP server is started over stdio during its honeypot run. We time the handshake (initialize to tools/list) and measure what its tool list adds to every prompt: the names, descriptions and schemas the agent re-reads on each turn, in estimated tokens. Two servers that do the same job can differ by thousands of tokens per turn. Context per tool is the fairest single number; the rest is beside it.
| MCP server | Tools | Context / prompt | Per tool | Handshake | Undocumented | Installs | Verdict |
|---|---|---|---|---|---|---|---|
| Benchmarks appear as servers are run; the crawler is working through the index most-installed first. | |||||||
Tokens are estimated at four characters each from the JSON the server returns for tools/list; your model's tokenizer will differ by a constant. Handshake is wall-clock in the sandbox from launch to the tools/list reply, dependencies pre-installed. Undocumented counts tools whose description is under 20 characters. Skills and plugins carry an instruction-token figure on their own verdict pages. Benchmarks have been collected since November 5, 2026; older verdicts show none until their next run, most-installed servers first.