Install the Arcade agent plugin

The Arcade agent plugin bundles a connection to Arcade’s hosted MCP gateway with two skills, three commands, and the arcade-operator subagent. The gateway gives your agent thousands of tools for email, calendar, chat, issues, docs, and dozens of other apps. The rest of the plugin helps your agent choose and use those tools. You authorize each app in the browser the first time it comes up, with no API keys to create and no gateway to configure.

Install

Each client installs plugins differently. Use the path for yours.

claude plugin marketplace add ArcadeAI/arcade-plugin
claude plugin install arcade@arcade

codex plugin marketplace add ArcadeAI/arcade-plugin
codex plugin add arcade@arcade

git clone https://github.com/ArcadeAI/arcade-plugin ~/.cursor/plugins/local/arcade-plugin

https://github.com/ArcadeAI/arcade-plugin

copilot plugin install ArcadeAI/arcade-plugin

After installation, reload your agent or start a new session. Using another client that does not load plugins? Connect to the same gateway with the tools-only Arcade bundle.

Sign in

Authentication happens in two layers:

1. Into the gateway. Your client will ask you to authenticate when it first connects. You sign in to Arcade in the browser, and the client can reach the gateway from then on.
2. Into each app. When a task needs Gmail, Slack, Linear, or anything else, your agent hands you a link that grants the scopes that task requires. Approve it once and Arcade holds the token for that app. You won’t need to authenticate that service again unless a later task requires greater scopes.

You never give your agent an API key or a password for either layer. Ask your agent to run /arcade-status to see what applications are currently enabled, or /arcade-connect google to authorize an app with the widest possible scope before you need it.

All of your tool calls, tokens, and secrets route through your own Arcade project, isolated from every other user of that gateway. No one else’s agent can ever reach your apps, tools, or tokens.

Every Arcade gateway works this way, not just the one in the plugin. Server-level and tool-level authorization goes deeper on the split.

What you get in each agent

Every install connects to the same Arcade gateway and gets the same tools. What differs is how much of the rest of the plugin each agent can load.

• ¹ Cursor IDE lists the commands on the plugin page but does not show them in the / menu.
• ² Cursor IDE and Cloud Agents do not run plugin hooks. The always-apply rule carries the routing guidance instead.
• ³ Cursor CLI runs the session hook but does not load the always-apply rule.
• ⁴ Cowork runs the prompt and subagent hooks but does not add the session-start text to the main conversation.
• ⁵ Copilot CLI runs session and subagent hooks, but drops prompt-hook output from configuration files.
• ⁶ VS Code reads the Copilot hook file but cannot run its Agent Plugins hook commands yet.
• ⁷ Codex does not run Agent Plugins lifecycle hooks yet.

Here is what each of those is for.

Commands

Cursor, Claude Code, and Cowork load all three commands. Claude Code and Cowork namespace them as /arcade:arcade-status, /arcade:arcade-connect, and /arcade:arcade-apps.

/arcade-status

Reports whether the gateway is reachable, whether you have signed in, and which apps you have authorized.

/arcade-connect <app>

Authorizes an app on demand, so you are not interrupted mid-task.

/arcade-apps

Lists the apps you have connected and lets you disconnect one.

Skills

Skills are instructions your agent reads on its own, so you never have to invoke them. try-arcade teaches it to turn a request like “reply to that thread” into the right sequence of tool calls across your apps. scale-arcade handles the other kind of question, the one about moving a workflow onto a team gateway, and it walks you through the decisions on this page against your own project.

Subagent

The arcade-operator subagent runs an app task in its own context and reports back just the result. Arcade exposes a lot of tools, and searching them fills a conversation with noise that pushes out your actual work. The operator absorbs that. Cursor, Claude Code, Cowork, Copilot CLI, and VS Code load it. Codex and Claude Desktop Chat do not.

Rule and hooks

The rule and hooks tell your agent that Arcade exists and when reaching for it is the right move, which is the difference between a tool that sits unused and one your agent actually picks up. Cursor IDE and Cloud Agents use an always-apply rule. Cursor CLI, Claude Code, Cowork, and Copilot CLI run the hooks their hosts support.

The columns thin out because the Agent Plugins  1.0 specification only makes skills and MCP servers portable. Commands, subagents, hooks, and editor rules are specific to each agent, so the plugin ships them where the agents that support them will find them.

Try it

This gateway includes every optimized toolkit in the catalog. Try asking your agent:

• “Send a summary of tomorrow’s calendar events to me on Slack.”
• “Read my Granola notes from last week, check them against Linear, and file anything I promised that is not a ticket yet. Wait for me before creating anything.”
• “Find unread mail that needs a reply. Look the sender up in HubSpot if you can, draft a response, and Slack me the draft before sending.”
• “Pull my open GitHub pull requests, my Linear issues, and tomorrow’s calendar. Write a standup update and send it to me on Slack.”
• “Search my Drive and Notion for the latest product spec, then drop a one-paragraph summary in Slack.”

Or ask “What can Arcade do?” and pick a workflow from the answer.

Next steps

• From personal use to team rollout: move a workflow onto your own gateway
• Connect to MCP clients: configure a client against a gateway you control
• Integrations catalog: every app and toolkit Arcade supports
• Agentic development: point your agent at these docs with llms.txt