Skip to guide

Connectors and tools

The model proposes a tool call

Suppose you ask an agent to find a document and summarise it. The model can request a search. The harness runs the permitted search tool and passes its results back. The model can then request the relevant document and use its contents in the answer.

A connector supplies access to a system. The tools exposed through that connection determine which operations the agent can request.

Understand the ways tools connect

A command-line interface, or CLI, lets software run commands. Model Context Protocol, or MCP, is a way to expose tools and data to an agent application. An application programming interface, or API, lets software request data or actions from another system.

These can work together. A connector may expose an MCP tool that calls a service's API. An agent may also reach a service through its command-line tool.

My connections in practice

My AI Tools & Skills page records the connections I use.

Printing Press builds API clients and optional MCP tools. Git worktrees keep concurrent repository work in separate checkouts.

Give the task the access it needs

Distinguish operations that read information from those that change it. Searching documents, editing a file and sending a message require different permissions. Check which operations a connection exposes and when the harness asks for approval.

Inspect the result of an action. A tool's response can report an error or incomplete work. The agent needs that result before it can decide what to do next.

Tools are one part of the harness. They supply context and carry out steps in a workflow. See AI Tools & Skills for examples.

Say hello

Draft a message about the work you need to complete.

Local draft preview. No message is sent. Review the draft, then copy it for use elsewhere.