> For the complete documentation index, see [llms.txt](https://docs.pipekit.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pipekit.io/ai.md).

# AI

Pipekit ships an MCP (Model Context Protocol) server inside its CLI. When you run `pipekit mcp`, the CLI exposes Pipekit's capabilities as MCP tools and resources that any compatible LLM client (Claude Desktop, Cursor, Claude Code, Goose) can call on your behalf.

That means you can use an AI agent to do the work you'd otherwise do through the UI, CLI, or Python SDK:

* **Debug a failed Run**: the agent fetches the Run, pulls the workflow node graph, reads the pod logs, and tells you what broke. See [Patterns > Debug a Failed Run](/ai/debug-failed-run.md).
* **Author a Workflow**: the agent reads the workflow-authoring guides shipped with the MCP server, drafts a Workflow YAML, validates it offline, and submits it. See [Patterns > Author a Workflow](/ai/debug-failed-run/author-workflow.md).
* **Automate operations**: schedule recurring work, suspend a cron, update a cluster, import an Argo Workflow Archive. See [Patterns > Automate Operations](/ai/debug-failed-run/automate-ops.md).

## What the MCP server exposes

The server registers **19 tools** covering Pipekit's read and write surface: list and inspect clusters/pipes/runs, fetch run logs, submit/stop/terminate/restart workflows, manage cron schedules, import from an Argo workflow archive. Each tool is annotated with the standard MCP `readOnlyHint` and `destructiveHint` so compatible clients can render warnings or apply approval policies. For the full inventory and per-tool hints, see [Tool Inventory](/ai/tool-inventory.md).

The server also publishes reference resources (Argo workflow schemas, examples, and Pipekit how-to guides) that the LLM can list and read on demand. This lets an agent learn how Pipekit and Argo work without you pasting docs into the chat.

## Get started

1. [Set up the MCP server](/ai/setup.md): install Pipekit's CLI, log in once, then add the MCP server to your LLM client.
2. Try a pattern: [Debug a Failed Run](/ai/debug-failed-run.md), [Author a Workflow](/ai/debug-failed-run/author-workflow.md), or [Automate Operations](/ai/debug-failed-run/automate-ops.md).
3. If something doesn't work, see [Troubleshooting](/ai/troubleshooting.md).

## Security model

The MCP server inherits your CLI's identity (the token at `~/.pipekit/token`). The agent can do exactly what you can do, no more. For the auth model, transports (stdio and HTTP), and the rules around binding to non-loopback addresses, see [Security & Transports](/ai/security-and-transports.md).