Quickspin Update
I added a control plane to Quickspin over the past couple of weeks. Quickspin grew from a CLI that drove a "runtime" interface to start Docker containers into a platform, with a control plane that owns each sandbox’s lifecycle. Quickspin creates sandboxes so that AI agents can run untrusted code in them. I rewrote the CLI to be the first client of the control plane.
Control Plane
I built the control plane in chi for HTTP routing, and it saves sandbox state in SQLite. The platform is single tenant today, but adding tenants to the schema later won’t be hard. I am focusing on hardening the control plane and creating SDKs that talk to the control plane before adding in multi-tenancy. The control plane tracks each sandbox’s lifecycle state itself, apart from what the runtime (Docker, today) says the container is doing. and it has an API contract for every action it supports. The server can list sandboxes, create sandboxes, execute commands in sandboxes, list files in sandboxes, and write files in sandboxes.
Control Plane Next Steps
Next, the control plane needs to behave like a real-world platform. The control plane’s job is to reconcile what you asked for with what the runtime is doing. It will also treat TTL as a first class primitive, which makes the platform cheaper to run. Sandboxes will expire, and users have to plan around it. Snapshots will loosen that, since a reaped sandbox can come back. For now, the reconciliation logic needs the TTL primitive to act on.
Better Isolation With gVisor
I watch a lot of YouTube videos about agent infrastructure, containers, and isolation techniques for agent sandboxes. Today, I watched Accelerating AI Securely with GVisor - Lucas Manning, Google. I updated Quickspin’s Linux development VM to use gVisor’s application kernel to give client sandboxes better isolation.
Web Application in the Works!
I’ve built full stack web apps for a decade. When the control plane hardens, I will create a web application client to pair with the CLI. I like the UIs of modern AI platforms like E2B, Vercel, Daytona, and Inngest (even though that is a workflow platform, not a sandbox platform). Quickspin needs a good UI. Later, I will add observability data, and charts in the ui to read it.