Run WordPress playbooks in ChatGPT or Claude, limited access

To run a WordPress playbook from ChatGPT or Claude on limited access, approve the app with the Content editor preset, switch on the two playbook tools and add a starter. The agent loads your steps with List playbooks and Get playbook. The preset decides which of this plugin’s tools those steps can reach.

That setup works from version 1.10.2, released on 5 November 2026. On 1.10.1 a connection on Content editor or Shop manager could not read playbooks at all.

This post is about running playbooks, not writing them. For what a playbook is and how to write one, read give an AI agent step-by-step instructions for WordPress. Here you get the setup on a limited connection, what each of the five starters needs, what happened when we ran one from Claude and ChatGPT, and what a pause does.

Takeaways:

  • Match the preset to the playbook: Read only for the reports, Content editor for Publish checklist and Draft a new post, Shop manager for Shop tidy-up.
  • A playbook is guidance, not a permission. The switches on the Abilities tab and the connection’s preset still cap what the agent can do with this plugin’s tools.
  • Three starters lose steps on Content editor: Weekly content report (no comments), Draft a new post (no site info) and Shop tidy-up (no WooCommerce tools). Check the starter table before you add them.
  • A paused connection answers each tool call with a plain message, and Resume needs no new approval.
  • Our test covered one playbook, Publish checklist, on a Content editor connection from claude.ai and ChatGPT. It is one data point.

Why a preset sits under a playbook

A playbook is a request, and the preset is a limit. The steps tell the agent what to do, and an agent can misread them. The preset caps which of this plugin’s tools the connection can call, whatever the steps say.

A preset only narrows. It never adds a tool you have not switched on under Abilities. The plugin describes Content editor as “Posts, pages, media and categories or tags. No deletes, users, settings or comments.” Custom is “Everything you have switched on, limited only by the approving user’s role”, so an administrator who approves a Custom connection hands the agent every tool you have enabled, at that user’s reach, apart from the high-risk tools you have left locked.

Since 1.11.0 (2026-11-06): the plugin’s wording for two presets now names playbooks. Content editor reads “Posts, pages, media, categories or tags and playbooks. No deletes, users, settings or comments.” Shop manager reads “Everything in Content editor, playbooks included, plus WooCommerce, except deletes and the nine money and authority tools.” What they reach has not changed. See the changelog.

Since 1.11.1 (2026-11-08): the Shop manager preset no longer reaches Create WooCommerce order, which joined the high-risk lock, so its summary counts ten money and authority tools instead of nine.

A site that has never set a default treats new connections as Custom. Choose the preset yourself the first time, on the approval screen or in Quick Connect.

Preset The plugin’s own summary Fits these starters
Read only Can look at your content, never change it. Weekly content report, Content health check, Shop tidy-up, the checking half of Publish checklist
Content editor Posts, pages, media and categories or tags. No deletes, users, settings or comments. Publish checklist, Draft a new post
Shop manager Everything in Content editor plus WooCommerce, except deletes and the nine money and authority tools. Shop tidy-up, plus everything Content editor fits
Custom Everything you have switched on, limited only by the approving user’s role. A playbook that needs a tool no other preset reaches

Content editor does not stop publishing. It covers content writes, and publishing still depends on the tool being on and the user being allowed. To hold new posts as drafts, turn on Force draft on create, covered in review what your AI agent wrote before it goes live. It does not hold edits to posts that are already live.

Agent Abilities for MCP is best for site owners who want Claude or ChatGPT to follow a routine they wrote, on a connection that stays on a preset such as Content editor. The steps live on your site, and the preset keeps the connection inside the plugin’s tools you chose.

Set it up on a limited connection

Update to 1.10.2, then set up the playbook itself as in the guide: turn on List playbooks and Get playbook, add a starter and set it Active. Quick Connect never switches those two tools on. Then connect the client.

  1. Add your site to the client. Both clients need the MCP endpoint, the address AI clients connect to, and sign in with OAuth, the browser sign-in method MCP clients use.
    • ChatGPT on the web: open Plugins, press the plus button, choose Add custom MCP server, enter a name and the endpoint URL, set authentication to OAuth, tick “I understand and want to continue” and choose Create as a plugin.
    • Claude on Pro or Max: open Customize, then Connectors, press + Add, choose Add custom connector, and enter a name and the URL.
  2. On the approval screen, pick Content editor, or Read only for the report starters. Already approved? Change it in the Access column on the Connection tab, under OAuth.
  3. Make the client reload its tools. In ChatGPT, open Settings, then Integrations, then Plugins, pick the app and choose Refresh. In Claude, start a new chat and switch the connector on from the + button.
  4. Ask in plain words: “Check my draft before I publish it.”

The ChatGPT and Claude.ai guides show every screen. Check your plan first. OpenAI’s help article says the current Plugins experience needs no Developer Mode (“You don’t need to turn on Developer Mode first”), lists full MCP, writes included, for Business and Enterprise/Edu, and says Pro connects with read and fetch permissions (OpenAI Help Center, Connecting custom MCP servers in ChatGPT).

Claude lists custom connectors on Free (one connector), Pro, Max, Ultra, Team and Enterprise. On Team and Enterprise an owner adds the connector under Organization settings first, then members connect it (Claude Help Center, Get started with custom connectors using remote MCP).

Our test used a Plus account, which the OpenAI article does not mention either way. On a read and fetch plan, expect the reading starters to work and the write steps to be the part that fails, though we have not tested that.

What each starter needs

The Playbooks library shows each starter in full. What a limited connection changes is which steps can run. This table comes from how the presets are defined in the plugin’s code, and the live test below covered only Publish checklist.

Starter Changes anything? Read only Content editor Shop manager
Publish checklist Only after you say yes to that exact change. Never publishes. Checks run, the two write steps are refused All six tools in reach Same as Content editor
Weekly content report No All three tools in reach No comments row No comments row
Content health check No All four tools in reach All four All four
Draft a new post Saves a draft, nothing else Cannot save, and the starter says to paste the post into the chat No Get site info Same as Content editor
Shop tidy-up No All three No WooCommerce tools All three

Publish checklist and Content health check run in full on Content editor. The other three need a decision from you.

  • Weekly content report: Get pending comments is out of reach on Content editor, because the preset leaves comments out. Step 5 of the starter tells the agent to say exactly what is missing instead of filling the gap with a guess. For the full report, use Read only on a connection whose user can moderate comments, since that tool requires it.
  • Draft a new post: Get site info is out of reach on Content editor, and the starter has no line for it. Edit step 2 to say “skip this if the tool is unavailable”. On Read only, Create draft is a write and is refused, so the starter’s fallback applies and the post goes into the chat.
  • Shop tidy-up: Content editor reaches none of its three WooCommerce reads. Use Shop manager or Read only.

What happened when we ran one

On 5 November 2026 we ran Publish checklist against our own demo site, demo.agentabilitieswp.com, on a build made just before the 1.10.2 release, once from claude.ai and once from ChatGPT. It was one run each on a demo that resets every night. Treat it as a smoke test, not a benchmark.

We switched on List playbooks and Get playbook (36 tools were enabled in total), added the Publish checklist starter and set it Active. Both clients connected with the Content editor preset, which showed as “up to 20 of 36” on the claude.ai connection.

  • claude.ai listed the playbooks, loaded publish-checklist, read the page with ID 3 and walked the checklist. It changed nothing.
  • ChatGPT, on a Plus account, was added through Plugins, Add, Add custom MCP server, with no Developer Mode step. It listed and loaded the same playbook on Content editor.
  • ChatGPT’s list result named Add post terms and Set featured image as unavailable. That field tells the agent which of a playbook’s tools this connection cannot use. We checked the demo afterwards: both were switched off there, and Content editor does reach them.
  • In our run the agents used Get page where the starter says Get post, because the item was a page.

What the test did not cover: the other four starters, the Shop manager preset, any write step, and clients other than these two.

What a pause does

On the Connection tab, under OAuth, press Pause on the connection’s row in Active grants. The connection keeps its sign-in and still lists its tools, and each tool call returns “This connection is paused by the site owner. It keeps its sign-in and works again as soon as they resume it.”

Both clients showed us that message in our test. A paused connection also gets no playbook index in the server instructions, which are the text a server hands a client at the start of a session, and it could not load a playbook anyway.

Resume brought it back on claude.ai in the same chat, with no reconnect, and ChatGPT’s list call returned its normal result. Since 1.10.2, resuming leaves no “paused” note behind in the instructions. Before that, a note put there when you paused could outlive the pause, because a client may keep the instructions for the whole session.

A pause stops further calls and does not undo what the agent already did. The Activity Log shows what it did, and how to find what an AI agent did on your WordPress site walks through reading it. Pausing one connection does not stop the same person’s Application Password, so pause, limit or expire an AI agent connection covers Pause all access too. Since 1.11.2, edits made with the tools the Changes tab covers can be reverted from there: undo what an AI agent changed.

What a playbook still cannot do

The guide’s what a playbook cannot do covers access, following the steps and editing. Two limits matter more on a limited connection.

  • A playbook cannot widen a preset. If a starter needs a tool the preset excludes, edit the steps or move the connection to a wider preset. Get playbook reports which of the playbook’s tools this connection cannot use.
  • A playbook cannot cover other plugins’ tools. A plugin that publishes tools straight to the default server of the MCP Adapter, the component this plugin builds on, skips this plugin’s switches and controls, and neither a preset nor a playbook changes that.

The run a playbook one step at a time recipe asks the agent to report after each step, which helps while you are still testing a playbook.

Questions people ask

Can I run a playbook in ChatGPT without Developer Mode?

Yes, in the current Plugins experience, as the plan note under the setup steps explains. We added our demo that way.

Why does the agent say a tool is unavailable?

Get playbook returns, for each tool a playbook lists, whether this connection can use it. A tool counts as unavailable when it is switched off on the Abilities tab, its plugin is not active, or the connection’s user or preset does not reach it. The Playbooks docs explain the fields.

Does Content editor mean the agent cannot publish?

No, publishing is covered under the preset table above. Use Force draft on create if you want new posts held as drafts.

Where do I find more playbooks?

The Playbooks library has the five starters and more written the same way. The release note for the version that made this work is what 1.10.2 changes.