MCP tools
Everything a connected agent can do, and the two things it cannot.
The tools
Deploying a real project
An agent can pass files inline, which is fine for a single script or a one-page site. Beyond that it stops being practical: every byte has to travel through the agent’s own context, so a build output or a folder of images is effectively undeployable that way.
For anything larger the agent uploads a zip instead, and the bytes go straight to storage without passing through the conversation:
- Zip the project, skipping node_modules, .git and .env — dependencies are installed here, and secrets belong in the environment.
- Call spocket_upload for a URL and an upload_id.
- PUT the zip to that URL. The URL is the credential, so there is no auth header.
- Call spocket_deploy with upload_id instead of files.
The URL lasts fifteen minutes and one upload, and the ceiling is your plan’s disk allowance. Five uploads can be open at once.
Moving an app
Rarely needed — a crash restarts on its own, and a machine that fails is evacuated automatically. But if an app is on a machine that is misbehaving, it can be moved:
The app is rebuilt from its current version somewhere with room, and the old container is removed only once the new one exists — so a failed move leaves it exactly where it was.
What agents cannot do
- Read an environment value back. Write-only, with no exception for automated callers.
- Delete an app. Deletion is permanent, so it stays a human decision in the panel.
- See anything about the machines beyond which ones your own apps are on — not how many exist, nor how full they are.
- Choose which machine an app is deployed or moved to.
Both limits hold regardless of what the agent is asked to do.
Revoking access
Every connected tool is listed in Settings with when it last called us. Revoking one takes effect immediately — the token stops working on the next request, and reconnecting requires your approval again.