Hey! I started bot-bottle about 6 months ago as a way to run agents securely on my local machine. Specifically, I wanted something that was agent agnostic, and which could provide a security layer that would be easy to set up and allow me to try new harnesses and agents. I also wanted something local first, and which would support harnesses like pi that would allow me to run models without reliance on the big providers.
The main contribution of bot-bottle is the idea of a "bottle". It's an easily defined, composable definition for how an agent can interact with the outside world, and doesn't depend on anything harness or model specific.
Bottles help solve the "exploit through trusted channel" problem by reducing the size of that trusted channel to be as small as possible, keeping credentials away from agents, and monitoring everything that comes in and out so leaks can be detected (either in real time when easy, or later via forensics if a more complex exploit gets by so you know what happened)
Mobile support was also pretty important for me and how I use coding agents, so I provided a way to interact with and unblock agents on my phone regardless of the harness or model.
Let me know what you think! Feedback on the threat model and the bottle abstraction are especially welcome.
Hey! I started bot-bottle about 6 months ago as a way to run agents securely on my local machine. Specifically, I wanted something that was agent agnostic, and which could provide a security layer that would be easy to set up and allow me to try new harnesses and agents. I also wanted something local first, and which would support harnesses like pi that would allow me to run models without reliance on the big providers.
The main contribution of bot-bottle is the idea of a "bottle". It's an easily defined, composable definition for how an agent can interact with the outside world, and doesn't depend on anything harness or model specific.
Bottles help solve the "exploit through trusted channel" problem by reducing the size of that trusted channel to be as small as possible, keeping credentials away from agents, and monitoring everything that comes in and out so leaks can be detected (either in real time when easy, or later via forensics if a more complex exploit gets by so you know what happened)
Mobile support was also pretty important for me and how I use coding agents, so I provided a way to interact with and unblock agents on my phone regardless of the harness or model.
Let me know what you think! Feedback on the threat model and the bottle abstraction are especially welcome.
this is cool, will try this out
Thanks! Let me know how it goes! Polishing up the console and then planning some more phone related features, would love whatever feedback you have