AI@kitten never used them but it's apparently a program that processes the output of an LLM, allowing it to contain special commands which may trigger an automatic reply from the MCP to the LLM.
for example you could make a "memory MCP" that processes {%store key value %} and {%load key %} commands in the output (the latter one would cause the MCP to reply with the contents of the key)
or you could make an MCP that allows the LLM to execute shell commands on your project, replying with the result. or one that allows the LLM to query the language server on your codebase...
AI@mildsunrise@kitten That shifts the threat level from annoying to potentially dangerous. I think I prefer the cramps and the chauvinists, tbh (see my other post for context).
re: AI@cinnamon@kitten yes, this explains the horror stories I see about LLMs deleting databases. although as long as the MCPs don't allow write access to stuff (or request approval before actually executing modifications) you should be fine right?
re: AI, weapons mentioned, fatal consequences implied@mildsunrise@kitten "As fine as letting a monkey play with a submachine gun with the safe disengaged"
re: AI, weapons mentioned, fatal consequences implied@cinnamon@kitten not that I'm a fan of LLMs or anything but I think it's far easier for a monkey to engage a safe, than for an LLM to exploit some bug in the MCP
re: AI, weapons mentioned, fatal consequences implied@cinnamon@kitten unless I'm the monkey in your analogy, in which case yeah i guess i could misconfigure the MCP or approve a dangerous modification, but in that case I'd kinda deserve it