Connecting an AI Assistant

This guide connects Claude to a Runner project over MCP and runs a first few prompts against it, so you can see the whole chain — client, OAuth consent, tool call — work end to end before trying anything more ambitious. The same steps apply to any other MCP-compatible assistant, only the client-side setup screen looks different.

Get the MCP address

Open Integrations from the left sidebar (or search for it) and find the Connect an AI Assistant card. It shows your Runner instance's MCP address, in the form https://<your-runner>/mcp.

Connect Claude

Click the dropdown next to the address and choose Claude — this opens claude.ai's connector settings in a new tab with the address already filled in. Confirm adding the connector there.

If you'd rather add it manually, or you're connecting a different client, copy the address instead and paste it into that client's own connector or integration settings.

This only works if your Runner instance is reachable from the public internet over HTTPS — claude.ai's connector settings call that address from Anthropic's own cloud, not from your machine. If your instance is internal-only, see Self-hosted Runner without internet access for how to connect a locally-run client instead.

Approve the connection

The first time Claude actually calls a tool, Runner shows a consent screen: "Claude wants to connect to your NodePit Runner account." Sign in if you aren't already, then choose a scope:

• Read Only — enough for the prompts below, and a safe default while you're still finding your footing.
• Full Access — needed once you want the assistant to create or change anything, see Building a Schedule From a Prompt.

Click Allow. The connection now shows up under Connected Applications on the same Integrations page, where you can review or revoke it later.

Run your first prompts

Back in Claude, try:

List my Runner projects.

Claude calls the list-projects tool and replies with the projects your account can see. Pick one from the reply and continue:

Show me the schedules in [project name], and tell me if any of them are currently disabled.

This chains two tools — listing schedules and reading each one's active state — into a single answer. If a schedule recently failed, ask:

Any failed runs in [project name] in the last week? Show me the log for the most recent one.

Once these come back correctly, the connection is working end to end and you're ready for a prompt that actually builds something — see Building a Schedule From a Prompt.