AI Sandboxes by Buddy - Isolated cloud machines for your AI coding agents
by•
Buddy Sandboxes give AI coding agents their own isolated Linux machines to build, run, test, and fix apps without touching your local setup. Each Sandbox has a persistent disk, root access, Docker, SSH, and a public URL. Create one per task, branch, or PR, then let your agents manage it through MCP, CLI, SDK, or API.
Replies
Hey PH, I work on Sandboxes. Some details from the engineering side:
You can set up a Sandbox in the UI, or write the config in YAML and keep it in your repo like any other IaC. Setup commands, apps, env, all in one file. If a Sandbox gets messy, recreate it and it runs the setup again from zero
Snapshots are probably my favorite part. Install deps, seed the DB, warm the caches once, snapshot it, and start new Sandboxes from there. The agent doesn’t spend the first 5 minutes of every task watching npm install
Long-running stuff (dev server, worker, DB) runs as an app with its own status and logs, separate from one-off commands. So the agent can actually check if the server is up, not just hope it is
Same goes for debugging. Logs, command output and status are all available through CLI and MCP. Give your agent access and it can run the app, read the error, fix it, and run again without you copy-pasting stack traces into the chat
Once an app is running you get a URL right away. Send it to teammates who have access to the Sandbox and they can check the result without pulling the branch or setting anything up locally
Plus the usual: SSH, root, Docker, and an SDK/API if you want to build your own tooling on top.
If you have questions about how it works, ask here, I'll be around all day.
What's the cost?
@bartshoot You pay by the minute while your Sandbox is running. We have a pricing calculator on our website to estimate the cost based on the workload you run:
https://buddy.works/sandboxes#pricing
I like that it’s a full machine and not just another short lived container we've seen a bunch of recently. How fast they spin up in practise?
@marcin_kudla1 Start time is around 70 ms.
This is very cool!
Which regions do you guys support?
@lev_perlman1 Hi Lev, US, EU and Asia
Had a chance to try sandboxes and must say that they are super-fast. Ready in like few seconds, with all those preinstalled tools must be a masterclass in caching or sth. EC2 instances launching can't even compare.
@michal_bigos Great to hear you like it!
Can we use Buddy Sandboxes as a standalone tool e.g. spinning them up directly from GitHub Actions or do we need to migrate our whole CI/CD pipeline to Buddy?
@tprus No migration is needed. Sandboxes work standalone. In GitHub Actions, install the bdy CLI, authenticate with a token from secrets, and run bdy sandbox create in a step. You can then run commands, expose endpoints, and tear it down after the job. There’s also a REST API and an MCP server. This way, you can keep GitHub as your CI, and Buddy provides the machines.
I’ve seen a bunch of “sandboxes for agents” on PH lately but the persistency is what makes this one interesting. Will try it out!
@mjakobsche Thank you! Give it a try and let us know what you think. We’d love to hear your feedback.
Which coding agents do you support?
@pawel_grobelniak you can install any popular coding agents like Claude Code, Cursor, Codex, Antigravity, Grok, and others.
How is this different from running an agent in a Docker container?
@marcin_kudla1 Containers come with limits: no systemd, no background services, no Docker without hacks. A Buddy Sandbox works like a regular Linux machine, so agents can install packages, run services and spin up apps with Docker, then expose them via a public URL. You can also snapshot it and roll back when an agent breaks something.