Agent

Overview

This page introduces the operator agent, @tetherto/mdk-agent, as a stack component: what it owns, how it reaches fleet data, and the two ways to run it. Read this before choosing how to deploy it.

What the agent is

A conversational operator interface: it answers plain-language questions about the fleet, calls fleet tools over MCP, and gates every write behind human approval. A local model routes and narrates; the tools compute, and the agent never invents fleet data.

How it fits the stack

Like any AI agent, it reaches fleet data through the standalone @tetherto/mdk-mcp server, the same MCP endpoint every other agent uses, not a special path of its own. What makes it distinct from a generic AI agent connecting to that server is that it ships as a complete, opinionated package on top: a tool-authoring contract, a pluggable session store, and an eval battery that scores routing, the answer, the result contract, and the approval gate against a live fleet.

Two ways to run it

• Standalone CLI: a small library and CLI for local development and evaluation. One conversation lives in a process variable and exits with it; no Gateway, no auth, no multi-user session store.
• Behind the Gateway: @tetherto/mdk-plugin-agent mounts the same library as a chat API, sessions, SSE message streams, and approval round-trips, so multiple operators hold independent conversations through one running Gateway.

The Gateway plugin is a deployment mode, not a different product: enabling it brings the same session, tool, and approval loop the CLI runs, with a SessionStore behind it instead of one variable in a process.

Key packages

Next steps

• Choose a guide to run the agent standalone or behind the Gateway
• Understand AI agents and the MCP server, the mechanism this agent is one instance of
• Understand the Gateway as a development surface, if deploying behind it