AI and MCP

Connect an AI agent without giving it the keys to everything

Start VMBook’s built-in stdio MCP server as a named, scoped team agent.

Connection model

The AI client starts vmbook --mcp (or vmbook-server --mcp) and exchanges MCP JSON messages over standard input and output. The server authenticates as a team agent using VMBOOK_TOKEN; it is not given the computer-wide API token.

Create a scoped agent

vmbook --team agent "Sponsor name" "Agent name" viewer "Lab/Build"

The command asks for the sponsor’s password and authenticator code, then prints the agent token once. The equivalent UI path is Team → People, Grants, then Tokens. Start with viewer and the narrowest useful group, host, site or machine scope.

Configure a stdio client

claude mcp add vmbook --env VMBOOK_TOKEN=vmbt_... -- "C:\Users\you\AppData\Local\Programs\VMBook\VMBook.exe" --mcp

Use the executable’s actual installed path; a system-wide installation may use another folder. Use VMBOOK_TOKEN_FILE instead when a protected file is safer than an environment variable. Agent tokens normally expire after 90 days unless configured otherwise.

Tools describe themselves

After connection, MCP tools/list returns every tool’s current description and JSON input schema. resources/list exposes vmbook://docs/api and vmbook://docs/agents. Treat this runtime schema as authoritative for the installed version.

Make capabilities the first call

{ "name": "get_capabilities", "arguments": {} }

The result identifies the agent, its grants and scope. Supplying a machine ID also reports which actions this agent would be allowed to request for that machine.