Most people already run software like that on their computer, even if it's just your coding harness that could just as easily run ssh and use your keys in ~/.ssh to connect to your servers. In my case I'd get a permission prompt from 1Password as I have my keys in there, but with all the prompts that a modern macOS throws at you the chances of it getting approved mistakenly are not zero.
Anyone who gives an LLM unfettered access to their ssh keys/agent socket deserves what they get IMO. I sandbox Claude with my own bubblewrap-based script which denies it access to anything I deem unnecessary.
There's many ways to shoot yourself into the foot, most things that run in your userspace can do evil things and it's not an LLM exclusive feature. Any npm package you install on your machine can put you in danger already if it executes some scripts.
Comments
Most people already run software like that on their computer, even if it's just your coding harness that could just as easily run ssh and use your keys in ~/.ssh to connect to your servers. In my case I'd get a permission prompt from 1Password as I have my keys in there, but with all the prompts that a modern macOS throws at you the chances of it getting approved mistakenly are not zero.
I'd not run OPs software either though.
Anyone who gives an LLM unfettered access to their ssh keys/agent socket deserves what they get IMO. I sandbox Claude with my own bubblewrap-based script which denies it access to anything I deem unnecessary.
There's many ways to shoot yourself into the foot, most things that run in your userspace can do evil things and it's not an LLM exclusive feature. Any npm package you install on your machine can put you in danger already if it executes some scripts.