Cosmonic Desktop

Cosmonic Desktop is a sandbox for your agent's code: a free, cross-platform desktop app for creating, running, and managing WebAssembly (Wasm) component workloads securely on your own machine. It is in public beta for macOS, Windows, and Linux, and runs entirely locally: no account, no cloud, no Kubernetes required.

If you've used desktop container tools, this should feel familiar. Cosmonic Desktop gives you a visual interface and a local runtime daemon: here, the runtime is the open source wasmCloud host. Everything you run is a sandboxed Wasm component—a portable, isolated unit of software compiled to WebAssembly. (New to these terms? The glossary has you covered.)

A desktop app for sandboxed code

AI assistants and coding agents now produce working code in minutes. Running that code safely hasn't sped up at all; the work has shifted from writing software to running it securely.

Cosmonic Desktop gives you a place to run untrusted code inside secure-by-default sandboxes, and it runs on hardware you already own. Workloads and data stay on your machine. That makes Cosmonic Desktop a simple and secure environment for:

• AI-generated code
• Agents
• MCP servers
• Third-party components

Everyday isolation tools run on ambient authority: a program inherits the permissions of whatever launched it. WebAssembly components start with no authority at all, no filesystem, no network, not even a clock, until the host explicitly grants each capability.

For more on why WebAssembly sandboxes are right for AI-era workloads, see Security and Non-Deterministic I/O.

How it works

Cosmonic Desktop has two parts: the graphical interface, and a daemon (cosmonicd) that embeds the wasmCloud host and does the under-the-hood work of pulling and verifying images, resolving configuration and secrets, and reconciling your declared workloads against what's actually running.

Because the daemon runs as an OS service (on a per-user basis), workloads keep running when you close the app and come back automatically at login. On Linux package installs they keep running even while you're logged out.

You run a component by pasting a link—an OCI image reference or a GitHub/GitLab repository URL. Desktop inspects the component, drafts a declarative Workload spec, and hands it to you for review before anything runs. Two more paths are enabled under Settings → Labs: a curated Catalog, and local projects with a scaffold and watch-rebuild-restart dev loop that promotes to a durable Workload.

The Workload specification is the same schema that Cosmonic Control uses on Kubernetes, so a workload you shape on your laptop deploys to a cluster unchanged, making Desktop the local on-ramp to production.

Secure by default

Every workload runs in a WebAssembly component sandbox designed for untrusted code:

• Deny-by-default egress. A component can only reach network hosts you explicitly allow—an application of the Principle of Least Authority that the whole platform is built on.
• Verified images. Component images are digest-pinned when applied and checked against a cosign signature policy on every start. Strict signing can be required globally or per registry.
• Secrets as references. Workload specs carry secret references (OS keychain, 1Password, AWS Secrets Manager)—plaintext values never appear in YAML, the API, the UI, or logs.
• No network API. The app talks to its daemon over a Unix socket (macOS/Linux) or named pipe (Windows). The only network listener is the optional workload HTTP ingress, bound to 127.0.0.1.

Get started

• New to WebAssembly? WebAssembly for Platforms and the Platform Concepts explain the foundations.
• Install on macOS, Windows, or Linux.
• Run your first workload in the Quickstart.
• Prefer to learn by doing? The built-in Academy (in the app, under Explore) is the guided, interactive path—first flights, sandboxing agent code, and the core concepts, all offline. These docs are the reference companion to it.
• Get oriented with the app, view by view—or go deeper on the daemon and the tray.
• Coming from containers? Containers to Components maps what you know.
• Take a workload to production in From laptop to cluster.
• Connect a coding agent with Connect Coding Agents, or run tool servers safely with Sandbox MCP Servers.
• Try in-progress surfaces like Projects and the Catalog under Settings → Labs.
• See what data the app does (and deliberately doesn't) collect in Privacy and telemetry.

note

These pages are growing quickly—if you hit something undocumented, troubleshooting covers the most common issues, and we'd love to hear what's missing.