Cosmonic

Founded 2021

Run untrusted AI agent code and tools inside secure isolated WebAssembly sandboxes

You can use Cosmonic to run untrusted code and Model Context Protocol servers generated by coding agents in isolated sandboxes. It prevents programs from touching your local filesystem, network connections, or secret credentials until you explicitly grant access. You can test agent-written scripts and tools on your local desktop for free, or deploy thousands of lightweight WebAssembly sandboxes per node across Kubernetes clusters in production. This lets you inspect application permissions safely and contain prompt injection risks without using heavy containers.

What Cosmonic does

Download and install the desktop sandbox applicationConnect coding agents via the Model Context ProtocolSet network access rules and allowed API endpointsInspect permissions for running agent componentsDeploy sandboxed workloads to Kubernetes clustersRun agent-written scripts in isolated memory spacesDeny-by-default execution environment blocking files, network, and secretsNative support for Model Context Protocol servers used by coding agentsSub-millisecond startup times with minimal memory footprintDesktop application for running sandboxes locally across macOS, Windows, and LinuxBuilt-in catalog of pre-configured applications and toolsKubernetes-native control plane for running sandboxes at production scaleFine-grained egress policies and granular permission grants per workload