What it does

The connector works over the Model Context Protocol (MCP). When your assistant needs data to answer a question, it calls the matching AgentFabrica tool and sees only what that tool returns. Every call is limited to the account that allowed the connection: the assistant sees that account's brands and nothing else.

Example questions

Which posts go out this week?
What is waiting for my approval? Give me the review link for each item.
How many leads did each brand get in the last 30 days, and how many came from ads?
How many WhatsApp conversations are waiting for a human reply?
Send this caption to the approval queue as a draft for Instagram and LinkedIn.
Which objections did the panel write in the latest ad pre-tests?

Requirements

An AgentFabrica plan with API access: Enterprise, E-commerce Premium, E-commerce Enterprise, Plus Starter, Plus Pro, Plus Enterprise, Network / MCN, Live Chat Pro, and Live Chat Enterprise. In the comparison on the pricing page this is the “Custom integration API” row.
If your plan does not include API access, the consent screen says so, links to the plan page and no connection is made.
If your plan loses API access after you connect, the assistant loses access too: AgentFabrica checks the plan on every call.
Claude: a custom connector added by URL works on the Free, Pro, Max, Team and Enterprise plans. The Free plan allows one custom connector. On Team and Enterprise plans an Owner adds the connector.
ChatGPT: custom MCP servers are added in ChatGPT on the web. Workspace permissions and security restrictions apply.

Connect in Claude

On a Free, Pro or Max plan

Open Customize > Connectors.
Click Add custom connector.
Enter AgentFabrica as the name and https://agentfabrica.com/api/mcp as the MCP server URL.
If the dialog asks for authentication settings, choose Sign in now for Authentication and Register automatically for OAuth client. If the dialog shows the name, the URL and Advanced settings on one screen, leave Advanced settings empty. Do not enter a client ID or secret.
Click Add. When Claude asks you to sign in, the AgentFabrica sign-in screen opens.
After you sign in, the consent screen shows the name of the app asking for access, its return address and the permissions you are about to grant. If they look right, click Allow.

On a Team or Enterprise plan

An Owner opens Organization settings > Connectors, selects Add, then Custom. If Claude asks for the connector type, they choose Web.
The Owner enters https://agentfabrica.com/api/mcp as the server URL, sets authentication as described above and clicks Add.
Each member opens Customize > Connectors, finds the AgentFabrica connector with the Custom label, clicks Connect and allows access with their own AgentFabrica account.

Use it in a chat

Click the + button in the chat box, open Connectors and turn AgentFabrica on for that conversation.

Do not choose Use Claude's published identity. That option needs the server to support Client ID Metadata Documents, and AgentFabrica does not. Register automatically uses dynamic client registration.

Source: Claude documentation, “Add a connector that isn't in the directory”. Checked on 14.10.2026.

Connect in ChatGPT

Follow these steps in ChatGPT on the web.

Open ChatGPT Plugins.
Select the plus button, then Add custom MCP server.
Enter AgentFabrica as the name. Under Connection, enter https://agentfabrica.com/api/mcp as the Server URL.
Choose OAuth for authentication. AgentFabrica does not issue static client credentials, so leave the client credential fields empty. If ChatGPT asks how to register the client, choose dynamic client registration (DCR); AgentFabrica does not support Client ID Metadata Documents (CIMD).
Read the risk warning and select I understand and want to continue.
Select Create as a plugin. The plugin appears in your personal plugins or in the workspace where you created it; install it before you use it in a conversation.
When ChatGPT asks you to sign in, the AgentFabrica sign-in and consent screen opens. If the details look right, click Allow.

Use it in a chat

Type @ in the prompt box and select the AgentFabrica plugin. ChatGPT asks for your confirmation before it runs a tool that is not marked read-only. In AgentFabrica that is only submit_post_for_approval.

Source: ChatGPT developer documentation, “Add custom MCP server”. Checked on 14.10.2026.

Tools

The connector offers 11 tools. The names below are the technical names the assistant sees; the assistant picks the right tool from your question. It calls list_brands first, and most other tools take a brand id from there.

Brands and plan

list_brandsRead-only: Lists the brands on the account: name, industry, website, tone of voice and brand role (your own brand or a client brand).
quota_statusRead-only: Shows your plan's generation quotas: used, limit, remaining and whether it is unlimited, for each generation type. Types that cannot be measured are left out.

Content calendar and approvals

scheduled_postsRead-only: Posts on the content calendar: title, platforms, time, status (planned, needs approval, publishing, published, failed, rejected) and the start of the caption. Upcoming or past posts, up to 50 at a time.
pending_approvalsRead-only: Everything waiting for a human decision: calendar posts that need approval and queued approvals (ads, blog posts, product imports). Each item carries a link to the panel page where you approve or reject it. This tool cannot approve anything.
generated_assetsRead-only: The record of generated content, newest first: type, category and date. It returns the record, not the content itself.
submit_post_for_approvalAdds a draft: Adds a draft post to the brand's approval queue. The post is not published until a person approves it in the panel. Platforms are Facebook, Instagram and LinkedIn; Instagram needs at least one image URL (up to 4, https). Captions can be up to 2,200 characters. Without a time, the brand's measured best posting time is used; a given time must be at least 5 minutes and at most 365 days ahead. Sending the same caption again within 10 minutes returns the existing draft instead of opening a new one.

Ad pre-tests and learning

ad_pretestsRead-only: Ad pre-tests run before launch: panel score, click-intent rate and the objections the panel wrote. If the ad later ran, the real result (impressions, clicks, click-through rate) is included. The score is the panel's reaction, not a reach forecast.
calibration_reportRead-only: How well the panel holds up for this brand: in how many comparisons it ranked the ads correctly. With too little data it reports the situation instead of producing a number.
learning_poolRead-only: Quality evidence drawn from the brand's own ad data: which hook, angle, visual style and format works, each with its evidence tier and reasoning. Attributes that always appeared in the same ads name each other: this data cannot tell which one made the difference, so they count as one finding. Returns nothing when no pool exists.

Lead and WhatsApp counts

lead_summaryRead-only: Lead counts per brand: total, by status (new, contacted, appointment, won, lost) and by source (Meta lead form, comment and DM, WhatsApp, website form, manual), all time. For the chosen period (1-365 days, 30 by default) it adds the new leads and how many of those came from ads. Counts only.
whatsapp_summaryRead-only: WhatsApp line status per brand (connected, disconnected, pairing, connecting, paused, not set up; QR or official line) and conversation counts: active, unread, needing a human, on automatic replies, hot leads, new in the last 7 days. Counts only.

What it never does

It does not publish. The only write tool adds a draft to the approval queue; publishing needs a person's approval in the panel.
It does not approve or reject. For items waiting for approval it returns the panel link only.
It does not delete or change anything. No existing record is touched.
It does not generate content or spend quota.
It does not send personal data. Names, phone numbers, form fields and message text of your leads and WhatsApp contacts never reach the assistant; only counts do.
It does not cross accounts. Every call is limited to the brands of the account that allowed the connection.

Disconnect

From the AgentFabrica panel

Open the API keys page in the panel.
Find the connection on the Connected assistants card. The card shows when each connection was made and when it was last used.
Click Disconnect. The assistant's access closes immediately. To connect again, start the connection from the assistant.

From the assistant app

In Claude, click Remove on Customize > Connectors; on Team and Enterprise plans an Owner removes the connector from Organization settings > Connectors. Removing it on the assistant side deletes the connector from that app. To make sure access is closed on the AgentFabrica side, use the panel card.

Automatic closing

A connection unused for 30 days closes on its own.
A connection in use asks for your permission again 90 days after you allowed it, at the latest.
Setting or changing your password, signing out of all devices or requesting account deletion closes every assistant connection.
If your plan loses API access or your account is suspended, the assistant cannot get in.

Data and retention

What the assistant receives

Brand details: name, industry, website, tone of voice and brand role.
Calendar posts and approval queue items: title, platform, time, status and the start of the text.
The record of generated content: type, category and date.
Ad pre-test results, panel calibration and the learning pool.
Quota status, and lead and WhatsApp counts.

The assistant processes the tool results under its own provider's privacy policy: Claude's policy for Claude, ChatGPT's policy for ChatGPT.

Connection records kept by AgentFabrica

The assistant app's registration: the name the app declares for itself and its return addresses.
The connection record: your account, the connected app, its return address, when it was connected, when it was last used and, if so, when it was removed.
Tokens and authorization codes are never stored in plain form; only their SHA-256 hashes are kept.

Lifetimes

Record	Validity and deletion
Access token	Valid for 1 hour. The record is deleted after 2 days.
Refresh token	Valid for 30 days, restarting on each use. The record is deleted after 31 days.
Authorization code	Valid for 10 minutes. The record is deleted after 2 days.
Consent screen request	Valid for 10 minutes. The record is deleted after 1 day.
Connection	At most 90 days from when you allowed it. A removed or expired connection's record is deleted 7 days later; a connection unused for 90 days is deleted too.
Assistant app registration	Deleted when no connection uses it: 7 days after registration if it never connected, 90 days after its last use otherwise.

The deletion job runs every day. For your personal data rights, see the privacy policy.

Support

Questions about the connection, the tools or data handling: write to destek@agentfabrica.com.