Version 1.10.1 adds Playbooks: step-by-step instructions you write once on a new Playbooks tab, which your AI agent loads when a task matches. It came out on 4 November 2026.
The same release makes refusals say why, adds an origin check to this plugin’s MCP endpoint (the address AI clients connect to over the Model Context Protocol, or MCP), and moves the catalog from 182 to 184 abilities, 90 in core and 94 from integrations.
Most of it waits for you: the two new tools are off, no playbook exists until you add one, and the starters arrive switched off. Three changes apply the moment you update, which are the origin check, a no-store header and the new refusal wording.
Takeaways:
- Five starter playbooks come with the plugin. A playbook is guidance the agent may follow, not a permission: the switches on this plugin’s Abilities tab and the connection’s access still decide which of its tools run.
- The two new abilities, List playbooks and Get playbook, are reads. They stay off until you enable them, and Quick Connect never turns them on.
- A paused connection, a blocked address, a rate limit and an unknown ACF field now tell the agent why the call was refused.
- This plugin’s MCP endpoint now answers 403 to a request from a web page on another site. Desktop apps, command-line tools and the hosted ChatGPT and Claude apps are unaffected.
Playbooks
A playbook is plain text with a title, a one-line description and steps. You write it on the Playbooks tab, tick the abilities it uses, and switch it to Active. Only Active playbooks reach agents, and the tab enforces these limits.
| Limit | Value |
|---|---|
| Playbooks per site | 50 |
| Title | 1 to 80 characters |
| Description | Up to 200 characters |
| Steps | 1 to 8,000 characters |
Why write one at all? An agent given a one-line request fills in the rest itself, and what it fills in is not always your routine. A playbook for “check a draft before it goes out” says what to look at, what to leave alone, and what to report when a tool is missing.
Five starters ship with the plugin. Each one arrives switched off, so you can read and edit it before any agent sees it.
| Starter | What it does | Writes? |
|---|---|---|
| Publish checklist | Checks one draft for a weak title, placeholder text, a missing excerpt, featured image, categories or tags, and lists what to fix. | Only after you say yes to that exact change. It never publishes. |
| Weekly content report | Counts posts published, drafts started and comments waiting in the last 7 days, as one short table. | No |
| Content health check | Scans recent posts, images and terms for gaps such as no featured image, missing alt text or unused tags. | No |
| Draft a new post | Writes a post from your brief and saves it as a draft. | It creates a draft and nothing else. It never publishes or schedules. |
| Shop tidy-up | Finds WooCommerce products with no price, no image, no stock or stuck in draft. Needs WooCommerce. | No |
The description matters more than it looks. It is the line the agent reads at the start of a session, so it is what the agent uses to decide whether a playbook fits the task.
How an agent finds and loads one
Playbooks reach the agent through the server instructions, a short block of text that an MCP server hands to a client when a session starts. The plugin adds an index to it when at least one playbook is Active, List playbooks and Get playbook are both on, and the connection can use both. A paused connection gets a one-line paused notice there instead.
Since 1.10.2 (2026-11-05): a paused connection gets nothing added to the server instructions. Clients can keep those instructions for the whole session, so a paused line could outlast Resume. The pause message now comes only with each tool call.
The index holds one line per playbook: its slug, which is the short name made from the title, such as weekly-content-report, and its description (the title if you left the description blank). It is cut at 600 characters, with a closing line that says how many more there are. Clients such as claude.ai and ChatGPT open with a server discovery request (server/discover) instead, and they get the same index.
From there the agent works in two steps. List playbooks returns each Active playbook’s slug, title, description and the abilities it uses, without the steps. Get playbook takes a slug and returns the steps, plus every ability the playbook lists with its tool name and whether this connection can use it right now.
The steps name tools by the labels you see on the Abilities tab, such as “the Get posts tool”. Because Get playbook returns the label next to the tool name, the agent can match one to the other and tell you when a tool the playbook needs is switched off.
What a playbook does not do
A playbook grants nothing. If the steps tell an agent to publish a post and the publish tool is off, or the connection’s user cannot publish, the call is refused like any other call to this plugin’s tools.
Agents cannot change playbooks. There is no ability, REST route or post type that reaches them, so only an administrator can edit them on the tab. The flip side is that an agent reads the text word for word, so keep passwords and private details out of it.
Adding, editing or deleting a playbook is recorded in the Activity Log by its slug, never by its title or steps. If the playbook list cannot be read, the plugin reports an error instead of treating it as empty, so a failed read cannot overwrite what you wrote.
The Playbooks library at /playbooks/ has more ready-made playbooks, and the Playbooks docs cover every field. For a worked example with a full playbook, read give an AI agent step-by-step instructions for WordPress.
Refusals that say why
Four refusals used to be vague. They now tell the agent, and through it you, what stopped the call.
| Situation | What the agent now sees |
|---|---|
| The connection is paused | The connection still signs in and lists tools. Each tool call returns an ordinary tool error: “This connection is paused by the site owner. It keeps its sign-in and works again as soon as they resume it.” |
| The address is blocked by your IP allowlist | A 403 with the code aafm_ip_blocked and the message “Your network address is not allowed to use this endpoint.” |
| The rate limit stopped the call | A message that names which limit refused it, the account’s or the app’s, and how many calls a minute it allows. |
| An ACF write names a field that does not exist | “These are not ACF fields on this site, so nothing was written”, followed by the names it did not find. If ACF’s field group tool is on and visible to the connection, the message points to it. |
The paused connection is the one that changed shape. In 1.10.0 it got a 403, and Claude showed a firewall or sign-in error instead of the reason. A paused connection now signs in as normal, and the refusal arrives where the agent can read it.
A blocked address used to get WordPress’s generic rest_forbidden code. Other status codes are unchanged.
For the controls behind the first and third rows, see pause, limit or expire an AI agent connection and the connection controls docs.
The origin check
The MCP specification says a server that uses Streamable HTTP, the transport that carries MCP over ordinary web requests, must validate the Origin header and answer 403 when it is present and not allowed. The current revision puts it this way: “Servers MUST validate the Origin header on all incoming connections to prevent DNS rebinding attacks” (Model Context Protocol, Streamable HTTP, revision 2026-07-28; the previous revision, 2025-11-25, says the same). As of 1.10.1 this plugin’s MCP endpoint follows the rule.
Browsers attach an Origin header when a web page sends a request. A request with no Origin header passes untouched, and the check leaves desktop apps, command-line tools and the hosted ChatGPT and Claude apps alone. A request that does carry one is allowed only when it names one of these:
- the site’s own address
https://claude.aiorhttps://claude.comhttps://chatgpt.comorhttps://chat.openai.com
Anything else gets a 403 with the code aafm_origin_not_allowed and the message “This site does not accept MCP requests from that web page.” The refusal is also written to the Activity Log as a denied transport row, though the log caps how many of those it writes.
If you run a browser-based MCP client you trust, add its origin with a filter:
add_filter( 'aafm_mcp_allowed_origins', function ( $origins ) { $origins[] = 'https://client.example'; return $origins;} );Use the form scheme://host, with a port only if it is not the default. The check covers this plugin’s endpoint only. The same release also sends Cache-Control: no-store on every response from that endpoint, refusals included.
Two new abilities, 184 in all
| Ability | What it does |
|---|---|
aafm/list-playbooks |
Lists your Active playbooks: slug, title, description, the abilities each uses and which of those this connection cannot use. |
aafm/get-playbook |
Loads one Active playbook by slug, with its steps and the details of each ability it lists. |
Both are reads, both stay off until you turn them on in the Site section of the Abilities tab, and neither can write a playbook. A connection with the Read only or Custom preset can use them. Content editor and Shop manager connections cannot.
Since 1.10.2 (2026-11-05): Content editor and Shop manager connections can use both tools too, so every access preset now reaches them. See the changelog.
The catalog is now 90 core abilities and 94 from integrations: WooCommerce, ACF, the four SEO plugins and the rest.
Smaller changes
- An unauthenticated MCP request no longer writes four lines to the PHP error log.
- A connection row shows “Saved” again after you save it, and a refused grant save shows its message on one line under the grants table.
- Adding a starter that is already in your list is refused, and the refusal is logged.
- The FAQ gained answers on refreshing a client’s tools, which ChatGPT plans can connect, what playbooks are, and why a browser-based client can be refused.
Updating
Updating does not switch any ability on. After it, you decide whether to enable the two playbook tools and whether to add a starter.
Clients keep the tool list they saw when they connected. In ChatGPT, after you enable the playbook tools, open Settings, then Integrations, then Plugins, pick the app and choose Refresh. In Claude, start a new chat or reconnect the connector.
On ChatGPT plans, OpenAI’s help article says that in the current Plugins experience you can add a custom MCP server directly from Plugins without turning on Developer Mode first. It lists full MCP support for Business and Enterprise/Edu, and says Pro can connect custom MCP servers with read and fetch permissions (OpenAI Help Center, Connecting custom MCP servers in ChatGPT). Our ChatGPT guide has the steps, and the Claude.ai guide covers Claude.
What did not change
The endpoint address, the per-role and per-connection allowlists and every ability switch are as they were. The plugin still contacts no AI provider and sends no telemetry.
The full history is on the changelog. The release before this one was 1.10.0, and the safety controls docs cover the switches that sit beside these.