Motivation
What problem does this solve? Who runs into it and when?
This solves the inherent security risks of running coding agents directly on workstations.
A key feature missing currently from almost any solution is protecting the host system where agents run. So called "guardrails" are a joke, and the truth is no coding agent, whether Claude Code, Codex, Goose, or otherwise, can be trusted not to potentially swing wide and damage the host OS it sits on.
There are many stories of agents going rogue, deleting files/directories on people's computers outside the scope of the project being worked on. This is why so many folks went out to buy Mac Minis when OpenClaw and other such projects popped up. Along with having a headless unit that could run 24/7, they would run these agents on separate small boxes, not their main desktop/laptop, where these tools "ran free" and could wreak havoc, but only on the Mac Mini and not their actual work machines.
If Buzz were able to let users create agents which each ran in their own container, the agents would have free reign within their container to install/run whatever they wanted (i.e., they would think they had full control over their host) but not be able to cause damage to the actual host. Given the option to share a host directory subtree inside the container, an agent could manipulate source code/etc. that was on the actual host but still not be able to do harm to the actual host OS. You could, in short, contain the potential "blast radius" of coding agents which go rogue.
Buzz already has such neat features for spinning up agents using different harnesses. This would be like icing on the cake, allowing developers to run their workspace on their main system without the security concerns it currently raises (and which limits its use).
Proposed solution
Describe what you'd like to see.
This might be a tad challenging to do cross-platform as a "built-in" feature. But on some platforms, such as macOS, options exist. The use of Apple's Container tech would be one possibility. However, more likely would be if Buzz simply had the ability to interact with a Docker daemon (e.g, Docker Desktop, Podman, OrbStack (macOS)), then the only requirement would be that the user had to make sure Docker functionality was installed on the system. And Buzz could check to see if the Docker daemon was available. If available, then the option could exist to create an agent either as is currently done (directly on the host running Buzz) or to run it within a container, with possibly the option to mount a host system directory within the container (e.g., some shared directory all the agents might have access to in order to work on project code).
Alternatives considered
What other approaches did you consider?
The only alternative option I see if having to build custom Docker Compose files to spin up containers, one per agent, where the user has to configure everything manually, then get the agents to hook into the Nostr relay being used by Buzz. But this is anything but efficient. At that point we're kind of back to how folks have to setup tools like OpenClaw with Telegram, etc., as far as how much work it takes. And one of the key benefits of Buzz (beyond the cryptographic audit trail, etc.) is that it makes creating teams of agents so easy.
Well that or risk the security implications of running agents directly on your host machine.
Additional context
Anything else that helps — links, screenshots, prior art.
Before opening: please search open issues and PRs for duplicates — link the closest one, or say "none found".
none found
Believe me, I have looked, both here and online for any tool which could do something like this. Buzz is the closest I have found in that you make it easy to create agents and leverage the different harnesses, along with all the Slack-like and Git features/etc. But the security risks are the same, if not higher, since folks would be installing/running more than one such coding agent harness in many cases, directly on their main workstations. That is, where most users might only install one such as Claude Code, Buzz makes it easy to create different agents using different harnesses. But that means the host system is exposed to multiple coding agents (e.g., Claude Code, Codex, etc.) and their respective security risks combined.
Currently if I do so, I tend to run a coding agent on my host system where I have to confirm each step (which is frankly exhausting), as there is no way I am unleashing a coding agent in "YOLO" mode and walking away. But if I had each coding agent contained within its own container/VM/microVM similar, where I KNEW that it could, at most, only wreak havoc on its own little world and, at worst, whatever shared directory subtree that was mounted into it, but otherwise could do no harm on my workstation, I would gladly run this on my main system.
Motivation
What problem does this solve? Who runs into it and when?
This solves the inherent security risks of running coding agents directly on workstations.
A key feature missing currently from almost any solution is protecting the host system where agents run. So called "guardrails" are a joke, and the truth is no coding agent, whether Claude Code, Codex, Goose, or otherwise, can be trusted not to potentially swing wide and damage the host OS it sits on.
There are many stories of agents going rogue, deleting files/directories on people's computers outside the scope of the project being worked on. This is why so many folks went out to buy Mac Minis when OpenClaw and other such projects popped up. Along with having a headless unit that could run 24/7, they would run these agents on separate small boxes, not their main desktop/laptop, where these tools "ran free" and could wreak havoc, but only on the Mac Mini and not their actual work machines.
If Buzz were able to let users create agents which each ran in their own container, the agents would have free reign within their container to install/run whatever they wanted (i.e., they would think they had full control over their host) but not be able to cause damage to the actual host. Given the option to share a host directory subtree inside the container, an agent could manipulate source code/etc. that was on the actual host but still not be able to do harm to the actual host OS. You could, in short, contain the potential "blast radius" of coding agents which go rogue.
Buzz already has such neat features for spinning up agents using different harnesses. This would be like icing on the cake, allowing developers to run their workspace on their main system without the security concerns it currently raises (and which limits its use).
Proposed solution
Describe what you'd like to see.
This might be a tad challenging to do cross-platform as a "built-in" feature. But on some platforms, such as macOS, options exist. The use of Apple's Container tech would be one possibility. However, more likely would be if Buzz simply had the ability to interact with a Docker daemon (e.g, Docker Desktop, Podman, OrbStack (macOS)), then the only requirement would be that the user had to make sure Docker functionality was installed on the system. And Buzz could check to see if the Docker daemon was available. If available, then the option could exist to create an agent either as is currently done (directly on the host running Buzz) or to run it within a container, with possibly the option to mount a host system directory within the container (e.g., some shared directory all the agents might have access to in order to work on project code).
Alternatives considered
What other approaches did you consider?
The only alternative option I see if having to build custom Docker Compose files to spin up containers, one per agent, where the user has to configure everything manually, then get the agents to hook into the Nostr relay being used by Buzz. But this is anything but efficient. At that point we're kind of back to how folks have to setup tools like OpenClaw with Telegram, etc., as far as how much work it takes. And one of the key benefits of Buzz (beyond the cryptographic audit trail, etc.) is that it makes creating teams of agents so easy.
Well that or risk the security implications of running agents directly on your host machine.
Additional context
Anything else that helps — links, screenshots, prior art.
Before opening: please search open issues and PRs for duplicates — link the closest one, or say "none found".
none found
Believe me, I have looked, both here and online for any tool which could do something like this. Buzz is the closest I have found in that you make it easy to create agents and leverage the different harnesses, along with all the Slack-like and Git features/etc. But the security risks are the same, if not higher, since folks would be installing/running more than one such coding agent harness in many cases, directly on their main workstations. That is, where most users might only install one such as Claude Code, Buzz makes it easy to create different agents using different harnesses. But that means the host system is exposed to multiple coding agents (e.g., Claude Code, Codex, etc.) and their respective security risks combined.
Currently if I do so, I tend to run a coding agent on my host system where I have to confirm each step (which is frankly exhausting), as there is no way I am unleashing a coding agent in "YOLO" mode and walking away. But if I had each coding agent contained within its own container/VM/microVM similar, where I KNEW that it could, at most, only wreak havoc on its own little world and, at worst, whatever shared directory subtree that was mounted into it, but otherwise could do no harm on my workstation, I would gladly run this on my main system.