AI Agents (MCP)

Let an AI agent such as Claude Code, Claude Desktop or Cursor check on, deploy and fix your projects. It works through zenavo-mcp, an open-source Model Context Protocol server, with an API key that can do no more than you allow.

Set up

1. Install zenavo-mcp from its releases (the README shows how to check the download) and sign in:

   zenavo-mcp login

2. The console opens /device and says that an AI agent wants to connect. Check that the code matches your terminal. Choose the permission: Deploy to let the agent deploy, roll back, restart and change environment variables, or Read only to let it look. Choose Only some projects and tick the projects it works on. Then approve.

3. Add it to your agent. For Claude Code, install the Zenavo plugin, which adds ready-made skills and starts the server for you:

   claude plugin marketplace add Zenavo-App/zenavo-skills
   claude plugin install zenavo@zenavo

   Or add only the server: claude mcp add zenavo -- zenavo-mcp. Other clients start it with {"command": "zenavo-mcp"}.

The key is named MCP on <computer>, expires after 90 days, and is saved on your computer, not in the agent's settings. Every change it makes (deploys, rollbacks, restarts, variable changes) is in your audit log under that name.

What an agent can do

Look	Projects, services, live state, deployments, jobs and the newest logs (with variable values replaced)
Change (Deploy permission)	Deploy, roll back, restart a service, set or remove an environment variable
Never	Delete anything, open a shell, read files or database rows, download backups, change servers or firewalls, billing, team members

It sees the names of environment variables, never their values. Tools that change something ask you first in your agent. Output from your services is passed to the agent marked as untrusted data. That makes it much less likely that text a service prints is taken as instructions, but it cannot rule it out, which is one more reason to keep the key narrow.

An agent with the Deploy permission can still deploy code that does anything your services can, so limit its key to the projects it works on and read what it proposes before you approve. Delete the key under API Keys to cut it off at once.