> ## Documentation Index
> Fetch the complete documentation index at: https://docs.armin.cx/llms.txt
> Use this file to discover all available pages before exploring further.

# Connect AI tools (MCP)

> Use Claude, ChatGPT, Cursor and other AI tools with your Chatarmin CX data.

Connect Claude, ChatGPT, Cursor, VS Code or Codex to your Chatarmin CX workspaces. Your AI tool can then answer questions about your support numbers, read tickets, review how the AI performed, and draft Torben AI Playbooks for you to review. If you allow it, it can also reply to customers as you.

It uses the Model Context Protocol (MCP). The server address is:

```
https://api.armin.cx/mcp
```

## Who can connect

Workspace admins, in workspaces where AI tool connections are enabled. On the consent screen you pick the workspaces the connection covers and the access it gets in each. It acts with your admin access, only in those workspaces.

When a connection covers several workspaces, your AI tool first lists them and then works in one at a time. Say which workspace you mean, or it asks.

## What your AI tool can do

| Access | What it covers |
| - | - |
| Read analytics | Ticket volume, backlog, response and resolution times, CSAT and automation rate, per inbox, channel, agent or team. |
| Read tickets | Tickets, their messages and contacts, ticket fields, views, macros and the orders of a ticket. This includes customer messages. |
| Read AI setup and performance | AI Agents and Torben AI Playbooks with their triggers, drafts, skills, tests and runs, the actions agents can use, and your custom HTTP actions (names, descriptions, parameters and host, never URLs, headers or secrets). Run details that quote customer messages also need Read tickets. |
| Read rules, workflows and knowledge | Ticket rules, classic workflows, help center articles, knowledge sources with the knowledge agents that own them and their training status, Torben AI knowledge bases with their sources, and learned knowledge-base answers. Learned answers quote staff replies, so they also need Read tickets. |
| Read AI usage | AI credits used and what credits beyond your plan cost, per billing period, agent and run. No invoices or payment links. |
| Change tickets | Status, priority, tags, assignee, ticket fields, internal notes, reply drafts for your team to send, and merging duplicate tickets or contacts. Needs Read tickets too. |
| Edit views, tags and macros | Create, change and delete views, tags and macros for the whole team. Needs Read tickets too. |
| Edit AI agents and Playbooks | Create, edit, test, turn on or off and archive AI Agents and their skills. With Torben AI: create and edit draft Playbooks, check routing and start test runs. An active agent answers customers; you activate Playbooks in Chatarmin CX. Needs Read AI setup and performance too. |
| Edit knowledge | Add web pages, text and Q\&As to the knowledge AI answers from: an AI agent's knowledge, or a Torben AI knowledge base. Every addition needs a confirmation in the call. Needs Read rules, workflows and knowledge too. |
| Send replies and approve AI actions | Send replies to customers as you, and approve or reject what AI Agents proposed (replies, refunds and other actions). Never ticked by default. Needs Read tickets too; approvals also need Read AI setup and performance. |
| Start new email tickets | Create email tickets for existing contacts and send their first email as you. Every new ticket needs a confirmation in the call. Never ticked by default. Needs Read tickets too. |

Every read access except Read AI usage also includes the lists of inboxes, tags, teams and views: the other tools take their ids. The list of members (names and emails) comes with Read analytics and Read tickets only. A write access works only together with the read access it names, which lists the tickets, agents or Playbooks it changes.

**Read tickets** includes contact data: your customers' names, email addresses and phone numbers, also through the contact tools. There is no ticket access without contact data. Your AI tool only sees the tools the connection allows, and Chatarmin CX checks every call again. Without write access there are no write tools, including ones added later.

Your AI tool sends a message to a customer only with **Send replies and approve AI actions** or **Start new email tickets**. It sends as you, on the ticket's channel with its signature, and every send and approval needs a confirmation in the call. Rejecting a proposal sends nothing. With **Start new email tickets**, a new ticket it creates is an email ticket from one of your email inboxes to a contact that already has that email address in Chatarmin CX; it creates no contacts, and an internal note ticket sends nothing. Send replies and approve AI actions alone never starts a new ticket. On WhatsApp it can't reply once the 24-hour window is closed: send an approved template from Chatarmin CX. Without this access, **Change tickets** lets it save a reply as a draft for your team: it appears as the suggestion in the ticket's reply editor, marked **Draft by** with your name and the app's name, and a teammate edits and sends it. Drafts don't use AI credits. In the inbox, replies it sent show **Sent via** and the app's name.

Write access starts unticked on the consent screen; tick only what the tool needs. Ticking a write access also ticks the read access it needs. Changes to what customers get (an agent's instructions or status, a skill, a macro) need an explicit confirmation in the call, and Chatarmin CX asks your AI tool to show the change and wait for your go. Chatarmin CX cannot check that a person approved: that depends on your AI tool. Keep its approval prompt on for write tools, and don't let it auto-approve them. Torben AI Playbooks it creates are drafts; you activate them in Chatarmin CX.

With **Edit knowledge**, your AI tool adds knowledge to the target you name. A workspace can have two kinds: the knowledge of an AI Agent's knowledge agent, which AI Agents search, and, with Torben AI, knowledge bases, which Torben AI Playbooks search. When both could fit, your AI tool asks which one you mean. New knowledge trains in the background: AI Agents and Torben use it only once training finished, and your AI tool can check its status. Files and whole websites for an AI Agent are added in Chatarmin CX.

## Connect

> [tab] Claude
>
> In Claude, open **Settings → Connectors → Add custom connector**, paste `https://api.armin.cx/mcp` and connect. Sign in to Chatarmin CX, pick the workspaces and allow access.

> [tab] ChatGPT
>
> In ChatGPT on the web:
>
> 1. Open **Plugins → Add → Add custom MCP server**.
> 2. Enter a name and the server URL `https://api.armin.cx/mcp`, and set **Authentication** to **OAuth**.
> 3. Tick **I understand and want to continue**, then click **Create as a plugin** once. A second click reports that the app already exists.
> 4. Open the plugin, click **Connect** and confirm. The sign-in page opens in the same tab.
>
> Some ChatGPT accounts still show the older path (**Settings → Apps & Connectors → Advanced → Developer mode**). If yours does, use that path with the same URL and OAuth.

> [tab] Claude Code
>
> ```bash theme={null}
> claude mcp add --scope user --transport http chatarmin-cx https://api.armin.cx/mcp
> ```
>
> Then start a new Claude Code session, run `/mcp` and sign in.

> [tab] Codex
>
> ```bash theme={null}
> codex mcp add chatarmin-cx --url https://api.armin.cx/mcp
> codex mcp login chatarmin-cx
> ```

> [tab] Cursor and VS Code
>
> Open **Settings → MCP** in Chatarmin CX and use the **Cursor** or **VS Code** button.

Tools that can't sign in with OAuth can use a personal access token. Create one in **Settings → MCP** and send it as `Authorization: Bearer <token>`. The token is shown once.

## Try it

Ask your AI tool:

* "How did our support do last week compared to the week before?"
* "Why did the AI hand off ticket #1234?"
* "Which Torben Playbooks fail most often this month, and why?"
* "Draft a Playbook for 'where is my order' requests and test it on a recent ticket."

Most AI tools also show ready-made prompts from Chatarmin CX, such as **Support report** and **AI performance review**.

Setting up your own agent or system prompt? Chatarmin CX serves guides through the `cx_guide` tool: `start` (read first), `entities`, `tickets`, `agents`, `playbooks`, `reporting` and `glossary`. The server asks your AI tool to read `start` first; if it doesn't, tell it to.

For Claude and Claude Code, **Settings → MCP** also offers the same guides and rules as a skill package: **Download skill package** gives a zip for Claude's skill upload, and the Claude Code line unpacks it into `~/.claude/skills`. It holds no workspace data and no token.

## Disconnect

Open **Settings → MCP** in the workspace and choose **Disconnect** next to the app or token. It loses access to that workspace right away; other workspaces it covers keep theirs. A connection also stops working in a workspace where the person who connected it is no longer an admin.

## Troubleshooting

* **"Only workspace admins can connect"**: ask an admin to connect, or to change your role.
* **"Not enabled for this workspace yet"**: AI tool connections are being rolled out. Contact your Chatarmin contact.
* **Playbook tools are missing**: Torben AI isn't enabled for the workspace yet, or the connection doesn't have **Edit AI agents and Playbooks** and **Read AI setup and performance** there. Reconnect and allow both.
* **Write tools are missing**: a write access needs its read access in the same workspace (see the table). Reconnect and allow both.
* **Knowledge base tools are missing**: knowledge bases need Torben AI in the workspace. Without it, your AI tool adds knowledge to AI Agents only.
* **Your AI tool can't send replies or approve**: the connection needs **Send replies and approve AI actions** in that workspace, and approvals also need **Read AI setup and performance**. Reconnect and tick them.
* **A workspace is missing**: reconnect the app and tick it on the consent screen. Reconnecting replaces the earlier connection, so tick every workspace you want.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.