# MCP and AI Agentic integration

**URL:** <https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660>\
**Category:** Feature requests\
**Tags:** zammad-73\
**Created:** [February 23, 2026, 3:10pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660 "2026-02-23T15:10:02Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![MarcosMG86](https://community.zammad.org/user_avatar/community.zammad.org/marcosmg86/32/11629_2.png) [@MarcosMG86](https://community.zammad.org/u/MarcosMG86)\
**Post date:** [February 23, 2026, 3:10pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660/1 "2026-02-23T15:10:02Z")

</div>

Title: MCP and AI Agentic integration

1. What is your original issue/pain point you want to solve? Any plans for AI Agentic integration with Zammad? Or maybe MCP server similar to this one: [GitHub - johncallhub/civicrm-mcp-server: Model Context Protocol server for CiviCRM · GitHub](https://github.com/johncallhub/civicrm-mcp-server)
2. Which are one or two concrete situations where this problem hurts the most? For Tier 0 troubleshooting/help for final users, we would like to use a chatbot/AI agent.
3. Why is it not solvable with the Zammad standard? I could not find any reference to MCP on the forum.
4. What is your expectation/what do you want to achieve? Seeking open source based Agentic service desk tools

Thank you!

---

<div class="post-metadata">

**Author:** ![MrGeneration](https://community.zammad.org/user_avatar/community.zammad.org/mrgeneration/32/9260_2.png) [@MrGeneration](https://community.zammad.org/u/MrGeneration)\
**Post date:** [February 23, 2026, 3:10pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660/2 "2026-02-23T15:10:54Z")

</div>

MCPs not really, but the upcoming 7.0 is all AI.

> **[Zammad AI. Free choice of LLM. For secure customer support.](https://zammad.com/en/product/artificial-intelligence)**
>
> AI customer support software built for your data protection requirements. AI summary, sentiment analysis, auto-routing and writing assistance — with free LLM choice or self-hosting. Fully GDPR compliant.

---

<div class="post-metadata">

**Author:** ![YetAnotherGerrit](https://community.zammad.org/user_avatar/community.zammad.org/yetanothergerrit/32/8413_2.png) [@YetAnotherGerrit](https://community.zammad.org/u/YetAnotherGerrit)\
**Post date:** [February 24, 2026, 4:47pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660/3 "2026-02-24T16:47:30Z")

</div>

Thank you for your suggestion. We will consider it. By the way, I am aware of at least two open-source projects that already provide an MCP server for Zammad. Bear in mind that they are not affiliated with us, so you should assess the risks yourself. You can find them with a quick web search.

---

<div class="post-metadata">

**Author:** ![awedor](https://community.zammad.org/letter_avatar_proxy/v4/letter/a/7c8e57/32.png) [@awedor](https://community.zammad.org/u/awedor)\
**Post date:** [November 20, 2026, 12:48pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660/4 "2026-11-20T12:48:20Z")

</div>

Hi.

I was on the hunt for a Zammad MCP Server and found this page: [MCP Server — documentação Zammad User Documentation](https://user-docs.zammad.org/pt-br/latest/extras/mcp-server.html)  
Since this page is not available in other languages it was probably not planned to publish this, but it looks like you are developing an integrated MCP Server yourself. Can you give us a timeline?

---

<div class="post-metadata">

**Author:** ![dominikklein](https://community.zammad.org/user_avatar/community.zammad.org/dominikklein/32/8144_2.png) [@dominikklein](https://community.zammad.org/u/dominikklein)\
**Post date:** [November 20, 2026, 1:30pm UTC](https://community.zammad.org/t/mcp-and-ai-agentic-integration/19660/5 "2026-11-20T13:30:18Z")

</div>

It’s the pre-release version (maybe a mistake that it’s already available for this language but for english it’s normal: [MCP Server — Zammad User Documentation documentation](https://user-docs.zammad.org/en/pre-release/extras/mcp-server.html?highlight=mcp) ) Here also another part: [MCP Server — Zammad System Documentation documentation](https://docs.zammad.org/en/pre-release/api/mcp.html).

> <https://github.com/zammad/zammad/issues/6519>
>
> As an agent, I want to work on my Zammad tickets from the AI client I already us…e — reading, finding and updating them there — so that I get help without copying ticket content between windows and without giving anyone more access than my own.
> 
> \## Problem and business value
> 
> AI clients are becoming part of daily support work. An AI client is an application that the agent chooses themselves, such as a chat application or a coding tool, and that connects to other systems through the Model Context Protocol (MCP). Agents use them to summarize a long ticket, find out what is open for a customer, draft an answer or record a finding. The client cannot see Zammad, so every one of these tasks starts with copying text out of Zammad and ends with copying a result back. This costs time on every ticket. It also loses information: internal notes are left out, the client cannot tell the customer's words from the team's, and the copied text is a snapshot that is outdated as soon as the ticket changes. Customer data ends up wherever the agent pasted it, and nobody can tell afterwards what was shared.
> 
> Teams that try to solve this today install a community-built Zammad MCP server (a program that offers Zammad's data to AI clients), run it themselves and hand it an API token. The result is only as safe as that program. The token usually allows everything its owner may do, so permissions and safeguards depend on a third party and differ from one tool to the next. Many organizations that run Zammad on their own infrastructure, for reasons of data protection or sovereignty, cannot accept this, and so they get no AI help in their ticket work at all.
> 
> An MCP server that is part of Zammad closes this gap. It gives every installation, hosted or self-hosted and with or without internet access, a supported way for AI clients to work with tickets and the knowledge base. Zammad stays the one place that decides who may see and do what, because the server uses the same permission checks as the interface and the API. Every change is made as the person who connected, and nothing reaches a customer without that person's approval. For agents, this means less copying and better answers. For admins and organizations, it means that AI use is possible without losing control over customer data. For Zammad, it makes the open-source helpdesk a system that AI works well with, and gives the in-product chat assistant and external clients one common set of tools.
> 
> \## Use case scenarios
> 
> \### Nina pastes a ticket link
> Nina is a second-level agent on a 25-person IT service desk at a managed service provider. A colleague sends her the link to a ticket, and she wants her AI client to tell her what the customer wants and what has been tried so far. Today she copies the messages over by hand. She misses the internal notes, and the client cannot tell which lines the customer wrote and which ones the team wrote. She expects the client to read the ticket itself from the link, the ticket number or the ID, and to see exactly what she is allowed to see in Zammad.
> 
> \### Nina asks what is open for a customer
> Before a call with Example Corp, Nina wants to know what is still open for that organization and which of those tickets are hers. Today she filters in Zammad, copies the ticket numbers into her AI client and asks her question there. She expects the client to find the tickets itself, by words or by names such as the organization, the state or "mine", to answer only from tickets she may see, and to link back to every ticket it mentions. The same goes for a how-to question: she expects the answer to come from the team's knowledge base instead of from her pasting an article.
> 
> \### Nina records what she found
> Nina and her client work out why Outlook stopped syncing for a customer. She wants to leave the finding on the ticket, and to prepare a reply for a colleague to check, without copying the text back into Zammad. She also does not want the client to send anything to the customer on its own. She expects the note to appear on the ticket as hers, and anything that reaches the customer to happen only after she has explicitly approved the actual message and its recipients.
> 
> \## Acceptance criteria
> 
> \- An agent can connect an AI client that supports MCP to Zammad with a personal access token they create themselves in their profile.
> - No software has to be installed on the Zammad side and no separate service has to run.
> - Signing in without a token (OAuth) is not part of this story.
> \- Everything a client does is done as the token's owner, within the permissions of both the token and its owner.
> - Tickets, messages, internal notes and knowledge base answers the user may not see are not returned and cannot be told apart from ones that do not exist.
> - A user without agent permissions, such as a customer, is offered no tools.
> \- A client can read a ticket named by its number, its ID or a pasted link, including its messages and the internal notes the user may see, and can look up the ticket's customer and organization.
> \- A client can find tickets by words and by names (customer, organization, state, owner, group, priority), can list the user's own overviews and the tickets in them, and can find knowledge base answers by words or, where semantic search is configured, by meaning.
> - Results are paged, a single answer never floods the client with a whole long ticket, and the client can ask for more.
> - Every object in a result links back to Zammad.
> \- A client can change what the user could change in the ticket view: fields, tags, internal notes, shared drafts and links to other tickets or knowledge base answers. It can also create a ticket.
> - Values are given in words (for example the name of a state), and the same rules apply as in the interface, including Core Workflow and read-only tickets.
> - Every change is recorded as made by the token's owner, in the ticket and in its history.
> \- Nothing reaches a customer unless the user has approved it.
> - A customer-visible message is created only through the tools that send one and never through the tools for internal notes or for saving drafts; sending a saved draft asks the same approval as a reply.
> - Where the client supports it, Zammad asks the user to approve the prepared message, showing its recipients, before it is sent. Otherwise the client's own approval step applies, and the documentation says so openly.
> - A declined, cancelled, outdated or repeated approval never sends anything.
> \- Text written by customers is marked as such in what a client receives, so the client does not treat it as an instruction.
> \- An admin can switch the MCP server on and off for the whole installation, in the API settings.
> - It is off by default. While it is off, every request is refused whatever the token.
> \- The MCP server is part of every Zammad installation (package, Docker, Kubernetes, hosted), needs no connection to a Zammad-operated service and works without internet access.
> \- The server passes the official MCP conformance tests for the features it offers.
> \- The documentation tells an admin how to enable the server and tells a user how to connect a client.
> 
> \## Background information
> 
> The server answers on a single endpoint, keeps no state between requests and supports the two current revisions of the MCP standard, so it runs the same way in a package install, in Docker and in Kubernetes. The in-product chat assistant for agents (#6235) works with the same tools, so what an external client can do and what the assistant can do do not drift apart.
> 
> Signing in with OAuth, so that hosted clients can connect without a token, and finer permission scopes for what a connection may do are out of the first iteration and follow in the second one. Running macros and time accounting are planned as follow-up stories as well. Earlier plans for this server also asked for a permission of its own for MCP, rate limits, an audit log of every call and a separate switch for write tools. These were dropped on purpose: a token reaches the MCP server exactly as it reaches the REST API, and what it may do follows the permissions the token and its owner hold. Customer and employee self-service through MCP is out of scope for this story.
