Recipes for building agents that actually do things — connect to real accounts, act on behalf of real people, and stay inside the permissions those people already have.
Every recipe is self-contained and runnable. Clone the repo, cd into one, follow it
start to finish.
| Recipe | What you'll build | Time |
|---|---|---|
| Connect Claude Code to Gmail and Slack | A fresh Claude Code install reading your inbox and your Slack, through one gateway you describe in a sentence | ~5 min |
| Build a Gmail agent with Muse Glimmer | An agent that triages, drafts, and sends mail — with a 30B open-weight model, running locally or in the cloud, and the credential nowhere near it | ~15 min |
If you're new to Arcade, run Connect Claude Code to Gmail and Slack first. It's the shortest path from nothing to a working agent, and it introduces the two ideas the rest of the recipes build on:
- Tools are scoped by description. You describe the job in English, Arcade resolves it to a fixed tool list, and what isn't on that list can't happen — no matter how you prompt the agent later.
- Authorization happens at the point of need. No setup wizard collecting every permission up front. The consent screen appears when a tool is first called, and each service is a separate grant.
Recipes say what they need, but most assume:
- An Arcade account — the free tier covers everything here
- An MCP client (Claude Code, Cursor, Claude Desktop, or another)
- Accounts for whatever the recipe connects to
- Recipes start from a state you can verify yourself, usually
claude mcp listreturning nothing. If a recipe needs setup, it says so before the first step. - Every recipe ends with cleanup — how to remove what you added and revoke what you authorized.
- Folder names are the thing you're building, so you can find one from a search result and run it without reading the rest of the repo.