AI Agent Sandboxes
An AI agent sandbox limits what an agent can reach: which files, which hosts, which credentials. In Cosmonic, an agent runs as a WebAssembly component that starts with none of those, and each tool it calls is a separate component with its own grant. The host makes every model call and holds the keys, so a key never enters the agent.
Build an agent
- Quickstart: build a sandboxed agent, run it and give it a tool.
- Agent SDK overview: what the SDK is and how its parts fit together.
- Write a typed tool and give an agent a tool.
Run a ready-made harness
Run agents where you need them
- Cosmonic Desktop on your own machine.
- Cosmonic Control on your Kubernetes clusters.
Understand the model
- Anatomy of an agent
- Security model and limits
- Sessions
- Security design principles for agent sandboxes
- How to choose an AI agent sandbox
Frequently asked questions
What is an AI agent sandbox?
An AI agent sandbox limits what an agent can reach: files, network hosts, credentials and processes. In Cosmonic’s sandbox the agent starts with none of them, and each tool it calls is a separate component granted only what its job needs, such as one folder or one host.
Which agents can run inside the Cosmonic sandbox?
Pi, OpenAI Codex and Goose run inside the sandbox today, on Cosmonic Desktop’s Agents page and on Cosmonic Control. Any harness built on the Cosmonic Agent SDK runs the same way.
Do I need Docker to sandbox an AI agent?
No. Cosmonic Desktop runs agents and their tools as WebAssembly components on macOS, Windows and Linux, with no containers or virtual machines. Each component gets only the grants it needs, rather than one wall around the whole process.
Can I run sandboxed agents on Kubernetes?
Yes. Cosmonic Control runs the same agent images on Kubernetes. The workload manifest declares each component’s grants, and runs start over HTTP with the OpenAI Responses API or over NATS.