Agentic workflows
Let an AI agent operate your desks, workspaces, and environments through the Kode skills and tools.
Last updated
Overview
An AI agent works in a desk. The desk gives the agent its tools: one tool for each action on the desk, its workspaces, and its environment. The skills tell the agent what the Kode words mean and which tool does which job. The agent acts with the account, rights, and limits of the developer it works for. A quota or a paused team access that stops you also stops your agent. See Limits.
The sandbox of an agent is a workspace: the agent reads, edits, and runs code there. A capsule runs one command when the agent needs a clean, short-lived run.
| Page | Contents |
|---|---|
| Skills | The Kode skills, what each skill gives the agent, and how to install them. The Claude Code plugin delivers them to Claude Code. |
| Connectors — MCP | Connect any MCP client to the tools of a desk with kode mcp. |
| Kode harness (preview) | Direct an agent in kd-tui, the harness that kode opens. |
Choose the page by the agent that you use.
| Agent | Tools | Skills |
|---|---|---|
| Claude Code | The plugin starts kode mcp. | The plugin delivers them. Install the plugin. |
| Another MCP client | Add kode mcp to the client. | Copy the kode skill to the client, if it reads skills. |
| The Kode harness | Built in. | Built in. |
Every tool, with its arguments and its API route, is in Agent tools. The HTTP API of the tool server of a workspace is in Tool server API. The kd command inside a workspace is in kl.