Your playbooks in every Claude — one plugin, no API key

Until today, the honest answer to "how do I use my playbook in Claude Desktop?" was: you can't, not properly. Our own docs said otherwise. They told people to put a URL and an Authorization header into claude_desktop_config.json — a file that only describes local programs. Claude Desktop ignored the entry, said nothing, and the server never appeared. Four documentation pages and the dashboard repeated it.

That is fixed.

What changed

Every playbook endpoint is now an OAuth protected resource. Point any client that speaks MCP authorization at https://agentplaybooks.ai/api/mcp/ — or at /api/mcp/manage for the whole account — and it discovers our sign-in, registers itself, and asks you to log in. No key to create, copy, paste, or rotate. That works in claude.ai, Claude Desktop, Cowork, the mobile apps, Claude Code, Cursor, and VS Code.

There is a Claude plugin. It bundles three things:

• the account connector, so every playbook you own or have been given is one

connection away — including one a colleague shares with you next week;

• a playbooks skill, which teaches Claude the habits that make a shared

playbook work: find the right playbook before acting, adopt its persona when you ask, put durable facts into memory instead of repeating them, and ask before changing anything other people rely on;

• /agentplaybooks:doctor and /agentplaybooks:sync for Claude Code, which

audit and synchronize the agent configuration files in a project.

A plugin you add on claude.ai lands on your account, so it is in chat, in Cowork, and — at the next session start — in Claude Code without installing it again.

How to get it

Today, from the repository. In claude.ai or the desktop app, open Customize → Plugins → Add → Add marketplace, enter matebenyovszky/agentplaybooks, and add AgentPlaybooks. Then open the plugin's Connectors tab and select Connect. In Claude Code:

/plugin marketplace add matebenyovszky/agentplaybooks
/plugin install agentplaybooks@agentplaybooks

From Anthropic's directory. The plugin is built to the directory's rules and goes into its review now. Once it is listed, it appears under Discover on Pro, Max, Team, and Enterprise plans.

Just one playbook? Add it as a custom connector under Customize → Connectors with its URL from the playbook's Integrations tab. That route also applies the playbook's persona and instructions as the server's own prompt. The Integrations tab now has the Claude steps and one-click buttons for Cursor and VS Code.

All of it is in the new guide: AgentPlaybooks in Claude.

What Claude can and cannot do with it

The connector acts as you, and no more than you: it reaches what your account can reach. Every tool declares whether it only reads, so Claude can run reads — listing, searching, opening a playbook — without asking, and asks before a tool that changes something, unless you have chosen to always allow that tool. Deleting and calling out to connected services are marked destructive.

Credentials stay where they are. list_secrets returns names. use_secret asks the AgentPlaybooks server to make the request with the value injected there, and only the response comes back. Claude never sees the key, and neither does the conversation transcript.

Building it to the directory's rules

Two things in the existing plugin would have kept it out of claude.ai entirely, and they are worth knowing if you build one:

• A top-level bin/ folder makes claude.ai and Cowork refuse the whole plugin.

Our CLI plugin shipped its executable that way. The Claude plugin carries no executables; its commands run the published CLI through npx, pinned to the exact version it was released with, and the release check fails if a pin falls behind.

• Every tool needs a title and its read-only and destructive hints. The

directory's portal reads them to decide what Claude may run without asking. All fifty of our tools declare both, and a test fails the build if one stops doing so.

The plugin is fourteen kilobytes of markdown and JSON. Everything it does is in the plugin's README.