If you built an app by talking to Claude or Cursor, the deploy step is usually where the conversation stops and a tutorial starts. It does not have to. MCP lets the editor you are already in do the deploying.
What MCP is, briefly
The Model Context Protocol is a standard way for an AI assistant to call tools that live outside itself. The assistant discovers what a server can do and calls it when the conversation calls for it. You do not learn an API; you say what you want.
Practically: instead of reading deployment docs, you say "deploy this app" and the editor does it, then reads the logs back to you when something goes wrong.
Connecting once
Adding an MCP server is a one-time step per editor. Spocket uses OAuth rather than an API key. Your browser opens, you approve the connection, and there is no secret to copy into a config file or leak into a screenshot.
- 01Add the serverPaste the setup prompt into your assistant, or add the URL to your editor’s MCP config. Both routes end in the same place.
- 02Approve it in the browserA window opens once. Click allow. No key is generated for you to store.
- 03Ask it to deployPoint it at the folder and name the app. The editor bundles the source and ships it.
What it can do once connected
Deploying is the first thing, not the only thing. The tools cover the loop you actually run: ship, watch, fix, ship again.
- Deploy a folder and start it
- Read and search logs without leaving the chat
- Set environment variables, encrypted and write-only
- Restart, stop and start
- Roll back to any earlier deploy
- List what is running and how much memory it is using
The useful part is the combination. "It went offline, why?" makes the assistant read the logs, spot the missing variable, set it and restart. Three tool calls that would have been three context switches.
What gets uploaded
Your source files and your manifest. Not node_modules, not .git, and not your .env. Dependencies are installed from package.json or requirements.txt on the other side, and secrets go in environment variables where they belong.
Keeping the assistant in bounds
Letting a model deploy for you is only comfortable if the destructive actions are gated. Restart and stop ask before they run. Delete always asks. Every deploy is kept, so a bad one is a rollback rather than an incident, and every action is logged with who did what and when.
That is the actual design constraint: an agent should be able to do the routine things freely and the irreversible things never, without you reading a diff first.
Which editors work
Anything that speaks MCP: Claude Code, Claude Desktop, Cursor, Windsurf, Zed and VS Code among them. The protocol is the point: there is no Spocket plugin to install per editor, and an editor that adds MCP support later works on the day it ships.
If you do not use an AI editor at all, the panel still exists. Upload a folder, hit deploy. MCP is the fast path, not the only one.