Your project may need a desktop's tools, a Linux server or a particular development environment. Canopy lets you run coding agents there and follow the same work from another device. Viewing and execution have separate locations.
Put execution beside the files and tools
A worker runs coding agents on a computer you control. Add a project workspace pointing to its directory, then start a Thread there. The agent uses that computer's installed tools and configured access, which can matter for builds and platform-specific work.
One project can have workspaces on several workers. Choose the right copy when starting work. These directories remain independent: Canopy does not push or pull project files for you. Keep your normal Git process for exchanging code between computers.
Use one gateway or link local desktops
A server setup keeps state in one gateway and connects workers on other computers. Workers initiate their connection, so they do not need an inbound listening port. The gateway routes instructions and retains the transcript for the web and desktop interfaces.
Alternatively, local Canopy Desktop installations can link their computers. Each keeps its own services and can access the other's work. This is a desktop feature, rather than a way to federate independent server gateways. Choose the arrangement before adding machines.
Stay involved without sitting at the worker
Open a Thread to see recorded progress, add a follow-up or answer an approval. The agent continues on its original worker. A laptop can become your control surface while a desktop handles a long task in the environment it needs.
Your phone's browser can use a reachable web app when web access is configured. Native mobile apps are coming soon. Network access still needs a working route, such as your configured network or VPN; Canopy does not create universal internet reachability.
Plan for availability and model access
Keep the gateway and execution computer available for work that must continue while you leave your desk. A dropped viewing connection does not move the task. Local worker commands can wait for reconnection, while writes to an offline linked computer are rejected.
Install the relevant agent CLI and configure its login or compatible endpoint on the worker. Hosting Canopy yourself does not make model inference local. The configured provider processes model requests, and connected content or project context can be part of those requests.