Zendesk MCP: The Server, the Client, and What Works Today

Zendesk, the customer service platform, is one of the most searched names in MCP right now. MCP is the open standard that lets an AI assistant like Claude or ChatGPT call another product’s tools directly. It can read a ticket or draft a reply without you pasting anything.

Search “zendesk mcp” and one of the first results is a Zendesk Marketplace app called MCP Server. Zendesk did not build it. The developer is Swifteq, a third party, and since October 1, 2026 its paid plan costs €49 a month.

That is the whole problem with this term. Three different things answer to the same two words, and they point in opposite directions. One of them is a press release. And the setup step most guides still tell you to follow, creating an API token, stops working for new tokens on October 27, 2026.

This guide separates them, says what you can actually connect today, and shows the setup that will still work next year.

Two directions, one name

Every MCP connection has a server and a client. The server owns the data and offers tools. The client is the AI that calls those tools. If the protocol itself is new to you, our plain explanation of MCP covers it without the acronyms.

So “Zendesk MCP” can mean either of two things.

• Zendesk as the server: an outside AI, say Claude on your laptop, reads ticket #4821, summarizes the thread and posts an internal note. The data flows out of Zendesk.
• Zendesk as the client: the Zendesk AI agent answering your customers calls a tool in another system, say your order database, to check where a parcel is. The data flows into Zendesk.

A support lead searching this term almost always wants the first one. Zendesk’s documentation is almost entirely about the second. That mismatch is why the results page reads like five people answering different questions.

n8n has the same split, and our n8n MCP guide shows how a product can do both at once. Zendesk is not there yet.

What Zendesk actually ships today

The thing Zendesk documents in full is the client. Its support article on connecting to MCP servers in action flows covers it, and the details matter.

• Plans: Suite Growth, Professional, Enterprise or Enterprise Plus, or Support Team, Professional or Enterprise.
• Where: Admin Center, then Apps and integrations, then Connections, then Add connection.
• Auth: OAuth only. A server that wants an API key is out.
• Transport: HTTP servers only. A stdio server, the way most open-source servers run on your own machine, will not connect.
• Size: up to 250 tools per server, with no cap on the number of servers.
• Cost: every tool call made inside an action flow uses action credits.

The article carries no early access label now, which suggests the client has moved past its waitlist. We found no general availability announcement, so check your own Admin Center before you plan around it.

This is useful if you run a Zendesk AI agent and want it to act on other systems. It does nothing for the support lead who wants Claude to read the queue.

The server is still a press release

Zendesk did announce the server. Its Relate 2026 release on May 19, 2026 promised a Zendesk MCP Server to connect tickets and knowledge to outside AI systems, marked “Early access this summer”.

Summer has ended. Checked on October 8, 2026, Zendesk’s list of early access programs has no general MCP server on it. The client article mentions in one line that Zendesk can also act as an MCP server, and stops there. There is no product name, no setup page and no tool list.

The one MCP server Zendesk does run in early access faces the other way again. It puts your Help Center inside ChatGPT for your customers, so they get answers there and a ticket is raised when a human is needed. That is a support channel, not a way for your team to work tickets.

What to assume

Plan as if there is no official Zendesk MCP server, and treat its arrival as a bonus. Anything you set up today on OAuth will keep working when it lands.

Three ways to put tickets in front of an AI today

None of these is made by Zendesk. All three are real, and the right one depends on which assistant your team already pays for.

In ChatGPT: OpenAI’s own plugin

OpenAI builds a Zendesk plugin for ChatGPT and Codex, documented in its help center. You open Plugins, pick the Zendesk plugin developed by OpenAI, type your subdomain (acme, not the full address) and sign in with your own Zendesk account.

It is built to review tickets and customer history, find knowledge and prepare replies. It never grants more than your Zendesk role already allows, and each person connects their own account. In a managed workspace an admin may need to switch it on first.

If your team lives in ChatGPT, stop here. It is the least work of anything on this page.

In Claude: the open-source server

Claude’s connector directory had no Zendesk entry when we searched it on October 8, 2026. The most-starred open-source server, reminia/zendesk-mcp-server, fills the gap. It is Python, Apache-licensed, and was last updated on August 27, 2026, when it moved to OAuth.

Read from its code, it exposes seven tools. Get one ticket, list the latest tickets, read a ticket’s comments, add a comment, create a ticket, update a ticket and fetch an attachment. It also ships two prompts, one to analyze a ticket and one to draft a response, and reads your Help Center as a knowledge base.

One gap to know before you start. There is no search tool. The list tool pages through your most recent tickets, up to 100 at a time, sorted by date, status or priority. “Find every ticket that mentions refunds” means the agent reads pages until it finds them, so keep requests scoped to recent work.

Setup takes about ten minutes.

1. In Admin Center, go to Apps and integrations, then APIs, then OAuth clients, and add a public client with the redirect URL http://localhost:4567/callback.
2. Clone the repository, run uv venv && uv pip install -e ., and put ZENDESK_SUBDOMAIN and ZENDESK_CLIENT_ID in a .env file.
3. Run uv run zendesk-auth once. A browser opens, you approve, and the server renews its own tokens after that.
4. Add it to Claude Desktop’s config file, as below.

{
  "mcpServers": {
    "zendesk": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/zendesk-mcp-server",
        "run",
        "zendesk"
      ]
    }
  }
}

Because each person authorizes with their own login, comments the agent posts are authored by them and Zendesk applies their normal permissions. That is the right default for a support team.

Hosted: Swifteq’s MCP Server app

The Marketplace app from the opening of this guide is Swifteq’s MCP Server. It connects over OAuth, names Claude, ChatGPT and Cursor as supported clients, and is built around ticket analysis, with summaries, tagging, trends and search.

The free tier is 500 MCP calls a month. Paid plans began on October 1, 2026 at €49 per account a month plus €9 per extra seat. That buys you no install and no Python, which is a fair trade for a team without anyone who wants to run a server.

General integration hubs such as Zapier and Composio also expose Zendesk actions over MCP, billed against their own plans. They are worth a look if your team already pays for one.

The API token deadline most guides skip

Most Zendesk MCP guides still open with “create an API token”. That advice has a date on it.

Zendesk announced on June 1, 2026 that it is retiring API tokens in three steps.

• July 28, 2026: tokens unused for 30 days were switched off, and new accounts could no longer create them.
• October 27, 2026: no account can create a new API token, in the interface or through the API.
• April 30, 2027: every remaining token stops working. Zendesk says there is no opt-out and no extension.

So a guide that tells you to paste an email and a token into an environment variable is teaching a setup that cannot be repeated after October 27 and dies outright next April. The reminia server still accepts the old variables, logs a warning, and prefers OAuth the moment you set a client ID.

There is a second reason to move, and it is the better one. A Zendesk API token cannot be scoped. Zendesk’s own announcement says it can be used to impersonate any user in your account, including admins. OAuth gives each person their own permissions, and that is a far safer thing to hand an agent. Our guide to vetting an MCP server covers the rest of that check.

Limits that bite when an agent works the queue

Zendesk’s API rate limits are generous for reading. The Support API allows 200 requests a minute on Team, 400 on Professional and 700 on Enterprise.

Writing is where an agent hits a wall. Zendesk allows 30 updates per 10 minutes per user on any one ticket, and caps ticket updates at 100 requests a minute across the whole account. An agent told to “retag and reprioritize everything from this week” can reach that in one go, and so can everyone else’s integrations sharing the account.

The fix is boring. Ask for a plan first, approve it, then let the agent apply it in batches.

For the wider support process, our guide to AI customer support for small SaaS teams covers choosing the first questions to automate, preparing the documentation, and handing unresolved conversations to a person.

Which one you want

For most support leads using Claude, it is the open-source server. It is free, it acts as you, and it already speaks OAuth.

Copy this prompt

Use Zendesk, read only. Do not change anything.
1. List the 20 most recently updated tickets with ID,
   subject, status and priority.
2. Group them by the problem the customer describes.
3. For the biggest group, quote the one customer line
   that shows the problem best.
4. Suggest a macro reply for that group. Do not post it.

Run that first. It only reads tickets and their comments, changes nothing, and shows you in two minutes whether the connection is worth keeping.

What none of them reach

Every option on this page stops at the edge of Zendesk. That is the correct scope, and it leaves a real gap.

Zendesk holds the complaints customers filed. A lot of complaints are never filed. They are a Reddit thread titled “is anyone else’s sync broken”, an X post tagging your company, a LinkedIn post from a customer’s head of ops. Your support lead saw them, saved a few, and meant to come back.

When Claude summarizes the week’s tickets, it reasons from the half of the story that arrived through the front door.

Where ContextBolt fits

ContextBolt is ours, so weigh the bias. It is a Chrome extension that keeps the other half.

It saves any web page with its text in one click, and it captures your social saves into the same library. X bookmarks are captured as you save them. Reddit saves sync from your saved list within minutes. On LinkedIn it adds its own save button to every post, and it imports LinkedIn’s Saved Items export for everything you saved before. Every page and post you capture is tagged and clustered as it arrives, and search in the extension is instant and on your device.

Pro adds an MCP endpoint, signed in with OAuth, so Claude can search that library by meaning in the same session it reads your Zendesk tickets. Ask for the week’s refund tickets and the Reddit threads you saved about refunds, and you get both in one answer. It only knows what you chose to save. It does not monitor mentions for you.

Try this on Pro

Read only. Pull this week's Zendesk tickets about sync.
Then search my ContextBolt bookmarks for saved Reddit
and X posts about sync. Tell me which complaints show
up only in the saved posts, with a link to each one.

Basic is free up to 150 bookmarks. Pro is $6 a month and removes the cap, adds cloud sync and the MCP endpoint.

Worth doing?

Yes, and sooner than you might think, because of the token date.

If you set Zendesk up with an AI this month, set it up on OAuth. Pick the route that matches the assistant your team already uses, start with a read-only prompt, and keep bulk changes behind an approval step.

Do not wait for Zendesk’s own server. It was due this summer and there is still no page explaining how to use it. When it does arrive, an OAuth setup moves across with one config change.

Then give the agent the other half of the story. Install ContextBolt, free up to 150 bookmarks, and on Pro the same Claude session that reads your queue can find the Reddit thread where the customer complained first.