Run a cloud agent

Run the agent on a machine you already have, or order a new one. The list shows how much memory each machine has free and whether its CPU, memory and disk are healthy. A machine without room for another agent is listed as unavailable.

Order new asks for a region in the EU or the US, the type, the architecture and the size. A shared machine shares a physical core and can burst to full speed, but isn't built for continuous heavy load. A dedicated one keeps its cores to itself. Arm gives you physical cores and roughly 1.5 times the performance. Sizes run from 512 MB to 8 GB, and each shows its monthly price before you order.

Machines are paid up front. You are charged for the month when one starts, prorated to the time you use it, and anything unused returns to your balance.

Sharing a machine is your choice. Everything on one machine shares one security boundary. Give an agent a machine of its own when its credentials should stay apart from the rest.

Give the agent a name, add any variables it needs, and set how much of the machine's memory it gets. Then click Create agent.

Variables hold the keys the agent needs for its own work, such as GH_TOKEN for GitHub, or an API key for your model provider if you sign in with one. The agent gets them when it starts, so a change applies at its next start.

An agent needs at least 512 MB, and OpenCode 1 GB. Give it more for builds and long sessions. An agent that runs out of memory stops, and its page says why.

Once the agent is ready, click Connect next to its terminal. The agent is already running. Sign it in the way you would on your own computer, with your own plan or key.

The agent is yours alone. Only the person who creates an agent can open its terminal, so only you sign it in and type into it. Teammates see that it exists. Owners and admins can stop, restart or delete it, but never see its variables.

On the agent's page, pick the workspaces it may use, one at a time. Pick one and Meandr handles the rest. The agent sees every tool those workspaces allow, and every call it makes goes through their policies and approvals, like any other agent.

A new connection applies when the agent restarts. Restarting keeps the conversation, so the agent carries on from where it was.

Ask the agent for something in its terminal and close the tab. It keeps working, around the clock, for as long as it exists. Nothing stops it for being idle.

Come back from any browser, on your laptop or your phone, and connect to its terminal again. It is the same session, right where it is. Opening it on one device takes it over from the other, which then offers to reconnect, so you can switch devices whenever you like.

Its state lives on its machine. Repositories, sign-in and conversations stay in the agent's home folder on the machine's disk. They survive restarts and stops, and last as long as the machine does. If the machine is deleted or has to be replaced, they go with it, so push work you want to keep. Deleting the agent deletes its home too.

Every tool call the agent makes shows in the workspace's traffic under its name, with the tool, the rule that decided and how long it took. Calls your policies mark for approval wait for a person, wherever the agent runs.

The agent's page shows its machine too, so you can see when it needs more room. From there you can restart or stop the agent, give it more memory or a bigger machine, or delete it. A new size or machine restarts it at once.