Zendesk MCP server for Claude, what works today

Zendesk doesn’t have an MCP server you can use yet. It announced one on 19 May 2026, for early access “this summer”, and on 10 November 2026 its developer docs, its help centre and its list of early access programmes still had nothing on it. So every way to connect Zendesk to Claude, ChatGPT or Cursor today is someone else’s. Swifteq hosts one you sign in to, two community servers on GitHub run on your own computer, and Zapier and Pipedream offer Zendesk on their hosted servers.

Older guides miss that since 27 October 2026, no Zendesk account can create a new API token, and every token stops working on 30 April 2027, so a server that only signs in with a token can’t be set up from scratch any more.

This page is by the team behind Deacon, an AI support agent built for founders who’d rather not run a helpdesk at all. It answers the questions people ask in the chat on your site from your own docs, and it doesn’t connect to Zendesk. Unlike Zendesk, it already has an MCP server you can use today, made for a coding agent such as Claude Code rather than for a support team, and it comes near the end. You can try Deacon on the free plan. Everything here about Zendesk and the five routes was read on their makers’ own pages on 10 November 2026, and each fact links to where we read it.

Five ways to connect Zendesk to an AI tool today

An MCP server is what lets an AI tool such as Claude call Zendesk while you talk to it. Before you pick a route, check in the fourth column whether it can reply to your customers. On a Zendesk ticket a public comment is the reply your customer gets, and an internal note is the only kind they don’t see.

If you want Claude or ChatGPT reading your tickets today, Swifteq’s server is the shortest way there. It signs in with your Zendesk login, can’t delete anything and keeps public replies off until an owner allows them. If you’d rather your tickets didn’t pass through another company, reminia’s server runs on your own computer and signs in with OAuth, once an admin has made an OAuth client for it. Zapier suits a team that already pays for it and wants Zendesk in the same server as its other apps, and Pipedream does that for free for your own use. Skip mattcoatsworth’s unless you already hold a working API token, since it stops on 30 April 2027 either way.

Zendesk’s own MCP server, as of 10 November 2026

Zendesk’s Relate press release said its MCP Server would let businesses “connect Zendesk tickets, knowledge, and other data to external AI systems in a trusted and governed way”. Its footnote gave “Early access this summer”, and the page still says so. The index of Zendesk’s developer docs lists no page on it, and the early access list has no programme for it. Zendesk’s help article on its MCP client says Zendesk “can also act as an MCP server in other scenarios”, but links to nothing that shows how.

The one Zendesk programme that runs an MCP server today is ChatGPT as a customer support channel. It answers your customers inside ChatGPT from your help centre articles, runs action flows for customers who sign in and opens a ticket when they need a person. You apply with a form, it’s free during early access, and Zendesk says there will be “pricing and packaging” once it’s generally available. It serves your customers, so it doesn’t let your own Claude or Cursor into your tickets.

Zendesk API tokens stop working on 30 April 2027

Older Zendesk MCP guides start by asking you to make an API token in Admin Center, and mattcoatsworth’s README still asks for one. Zendesk is removing API tokens in three steps.

• From 28 July 2026, a token that goes 30 days unused is switched off, and it’s deleted 60 days after that. Accounts created on or after 28 July 2026 can’t create or use tokens at all.
• From 27 October 2026, no account can create a new token. A token that already works keeps working while it’s used at least once every 30 days.
• On 30 April 2027, every token stops for good, and Zendesk says there will be no extension.

Zendesk is removing tokens because of what one can reach. Its announcement says a token “can be used to impersonate any user in your account, including admins” and gives “all-or-nothing account access”. OAuth replaces it, with tokens that expire and can be limited to the scopes you choose. Four of the five routes here sign in with OAuth, and mattcoatsworth’s server takes only a token.

Swifteq’s MCP Server, hosted for you

The MCP Server app in Zendesk’s Marketplace is made by Swifteq, not Zendesk. Swifteq is an independent company in Dublin that makes Zendesk apps, and it runs the server itself, so there’s nothing for you to host. You sign in to Zendesk through Swifteq’s site, then add one address to your AI tool.

Its tool list runs to more than 40 tools. It reads tickets with their full conversations and attachments, users and organisations, macros, Help Center articles, triggers and, for admins, CSAT ratings. It also reports on your queue. count_tickets splits a count by group, channel or week, get_ticket_metrics_summary gives first reply and resolution times, and get_sla_queue checks open tickets against your SLA policies. So once it’s connected you can ask “Which tickets will breach their SLA next?”, “What was our median first reply time last month, by group?” or “How many messaging tickets did our Zendesk AI agent pass to a person this week?”

With write access it can add internal notes, change ticket fields and tags, create tickets, draft Help Center articles, apply and edit macros and merge duplicate tickets. It has no tool that deletes anything. Its write permissions keep public replies, publishing articles, changing users and organisations, changing macros and merging off until an account owner turns each one on, and an owner can hold any of them back from one person. A refused action isn’t done another way, so with public replies off it won’t quietly post an internal note instead.

It’s free for 500 calls a month for each Zendesk account. A call is one tool use, so a search that then opens three tickets is four calls, and a call that fails isn’t counted. When the 500 run out, calls are refused until the 1st of the month and nothing is charged. Asking your AI tool “What’s my Swifteq MCP usage?” tells you where you stand and uses no calls.

The Team plan has unlimited calls for €49 a month, which covers your account and its first user, plus €9 a month for each extra user, so a team of five pays €85 a month. Paid yearly it’s €490, plus €84 a year for each extra user. Prices are in euros, before VAT. Only the Team plan can change up to 100 tickets in one request or read 10 conversations at once, and on the free allowance your AI tool works through them one at a time.

Have one person sign up and invite the rest. The free 500 calls are shared by every Swifteq account on your Zendesk account, but the Team plan is charged for each Swifteq account, so three colleagues who each sign up and subscribe pay three times. Swifteq’s pricing article says one account with colleagues invited to it, each on their own Zendesk login, is usually much cheaper.

It works with Zendesk Support and Suite. Swifteq says some of the searches it uses need Suite Growth or higher, and macro usage counts need Professional or Suite Growth and above. Zendesk pricing has what those plans cost.

A few things only show up when you set it up. Swifteq’s guide says Claude needs a paid plan to add a custom connector, but Anthropic’s own guide says the free plan can add one. In Claude Code, sign-in fails with “Callback URL mismatch” unless you add --callback-port 8080. Swifteq’s ChatGPT guide lists Plus too, but OpenAI’s page gives full MCP support only to Business and Enterprise, and on Pro a custom server can only read, so it can’t change anything in Zendesk. Cursor works only once you’ve sent Swifteq the callback address it uses. There are no API keys for scripts that run on their own, and sessions expire, so you sign in again now and then.

Your tickets also pass through Swifteq on their way to your AI tool. Its security page says the server runs on AWS in Ireland and stores no ticket conversations. It also says read access includes internal notes and attachments, since Zendesk offers it nothing finer than read, or read and write.

Setting up Swifteq’s server in Claude

These are the steps from Swifteq’s guides to authorising Zendesk and connecting Claude, with Claude’s menus as Anthropic’s guide names them today.

1. Sign up at agenthelper.swifteq.com/auth/mcpserver with your email address.
2. When it asks you to connect Zendesk, log in to Zendesk as the admin or agent the server should act as. Read Helpdesk Data is required. Leave Write Helpdesk Data off if you only want it to read.
3. Enter your subdomain, which is yourcompany for yourcompany.zendesk.com. Click Authorize in Zendesk Support, then Allow.
4. In Claude, open Customize, then Connectors, press + Add and choose Add custom connector. Give it a name, paste https://mcp.swifteq.com/api/mcp/mcp, keep the sign-in settings Claude finds and press Add. Sign in with your Swifteq account when it asks.
5. In each new chat, press +, choose Connectors and switch it on. It isn’t on in every chat by itself.

In Claude Code, run this line, then run /mcp, pick zendesk-mcp and sign in when your browser opens.

claude mcp add zendesk-mcp https://mcp.swifteq.com/api/mcp/mcp --transport http --callback-port 8080

ChatGPT takes the same address. Open Plugins in ChatGPT, press +, choose Add custom MCP server, paste the address, choose OAuth and tick the risk warning. OpenAI’s guide says you don’t need to turn on Developer mode first, and on a Business workspace only an admin or owner can add one. If you started read only and want write access later, you have to authorise Zendesk again with Write Helpdesk Data on, because changing Swifteq’s settings alone won’t do it.

Claude asks before every tool call unless you change it. Swifteq suggests setting the read-only tools to Always allow and leaving Needs approval on anything that changes Zendesk, so a person still says yes to every change.

The two community servers on GitHub

Both sit on one person’s GitHub account and run on your own computer, so your tickets go between Zendesk and your AI tool with no other company’s server in between.

reminia/zendesk-mcp-server, which signs in with OAuth

reminia/zendesk-mcp-server is kept by the GitHub user reminia, under the Apache 2.0 licence. Its last commit, on 27 August 2026, added OAuth sign-in with PKCE, and its README now calls API tokens deprecated. It’s written in Python 3.12 or later and runs with uv.

It has seven tools, get_tickets, get_ticket, get_ticket_comments, get_ticket_attachment, create_ticket, update_ticket and create_ticket_comment. It also gives your AI tool your whole Help Center as a resource, and two prompts, analyze-ticket and draft-ticket-response. Watch create_ticket_comment, because its README says a comment is public unless the AI tool sends public: false.

Its README sets it up in four steps.

1. In Zendesk’s Admin Center, open Apps and integrations, then APIs, then OAuth clients, and press Add OAuth client. Set Client kind to Public, add the redirect URL http://localhost:4567/callback, and set Allowed scopes to tickets:read tickets:write ticket_attachments:read users:read hc:read. Copy the client’s Identifier. Zendesk only lets an admin make one.
2. Clone the repo, run uv venv && uv pip install -e ., copy .env.example to .env, and set ZENDESK_SUBDOMAIN and ZENDESK_CLIENT_ID to your subdomain and that Identifier.
3. Run uv run zendesk-auth once. Your browser opens on Zendesk, you approve, and the tokens are saved in ~/.config/zendesk-mcp/tokens.json. It renews them by itself unless they go 90 days unused.
4. Add it to Claude Desktop as a command, uv with the arguments --directory /path/to/zendesk-mcp-server run zendesk.

Claude Code takes the same command after --, which is how its MCP docs add any server that runs on your computer.

claude mcp add zendesk -- uv --directory /path/to/zendesk-mcp-server run zendesk

If Zendesk won’t accept http://localhost:4567/callback as a redirect URL, the README says to register https://localhost instead and run uv run zendesk-auth --manual, which has you paste the result back by hand. For a server that only reads, set ZENDESK_OAUTH_SCOPES to tickets:read users:read hc:read. Copy scope names exactly. Zendesk’s scope docs say a misspelt scope still gets a token, and then every request made with it fails with 403 Forbidden. zendesk-auth prints the scopes Zendesk actually granted, so you can compare.

mattcoatsworth/zendesk-mcp-server, which needs an API token

mattcoatsworth/zendesk-mcp-server is a Node.js server, and all 21 of its commits were made on 4 April 2025, with nothing since. Its README says MIT, but the repo has no licence file, and GitHub shows no licence for it.

It has 49 tools. Tickets, users, organisations, groups, macros, views, triggers, automations and Help Center articles each get list, get, create, update and delete, so nine of them delete things. The support_info tool only returns a placeholder message. Its update_ticket adds a comment without saying whether it’s public, and Zendesk’s comment rules then follow the ticket’s first comment, which on a ticket a customer sent is public.

It signs in only with your Zendesk email and an API token in a .env file. Unless you already have a token that still works, you can’t set it up, and it stops on 30 April 2027 either way.

Zapier’s and Pipedream’s hosted servers

These two put Zendesk in one MCP server with the other apps you’ve connected to them. You sign in to Zendesk through them once and their servers make the calls, so your tickets pass through them the way they pass through Swifteq.

Zapier MCP

Zapier’s Zendesk page lists actions rather than a fixed set of tools. You make a server at mcp.zapier.com and connect Claude, ChatGPT, Cursor or VS Code to it. Its seven Zendesk searches include Find Tickets, Find Ticket Comments and Find a User. Its actions include Create Ticket, Update Ticket and Add Comment to Ticket, which asks whether the comment should be public, along with tags, file attachments, creating and updating organisations, creating, updating and deleting users, and a raw API Request in beta.

Most Zapier servers let your AI tool switch on the tools it needs as you talk, from the apps you’ve connected. You can choose a fixed set instead, which is how you keep Delete User out of its reach.

Each successful tool call uses two tasks, and a call that fails uses none. Zapier’s Free plan has 100 tasks a month, which is 50 calls. Zapier lists Zendesk as a premium app, though, and its pricing page keeps premium apps for paid plans, starting from $19.99 a month billed yearly.

Pipedream MCP

Pipedream’s Zendesk server shares one address, https://mcp.pipedream.net/v2, with every other app it covers. Its setup guide has you add that address to Claude, ChatGPT, Cursor or VS Code, sign in to a Pipedream MCP account, which is separate from a pipedream.com account, and pick the apps to turn on. Pipedream says anyone can use it for their own use for free.

Its Zendesk tools are built from the 33 Zendesk actions in Pipedream’s open-source registry, whose Zendesk code was last changed on 6 October 2026. They search and list tickets, comments, articles and macros, create and update tickets, set tags and custom fields, handle side conversations, and include Delete Ticket. A comment added with Update Ticket is public unless the AI tool says otherwise, because the shared setting defaults to true.

Zendesk’s AI calling other MCP servers

Zendesk does use MCP today in its own AI, which can call out to other companies’ servers. Its MCP client became generally available on 10 August 2026. An admin connects an outside MCP server in Admin Center, under Apps and integrations, then Connections, Add connection and Connect to an MCP server, and ticks which of its tools to allow. Zendesk’s own examples are Asana, Stripe and GitHub.

Those tools become steps in action flows, which auto assist, AI agents and custom agents can run. Zendesk’s setup article sets the limits. The server has to use OAuth over HTTP, and one that needs no sign-in isn’t supported. Zendesk reads up to 250 of its tools, each call has 30 seconds, and every call in an action flow uses action credits.

Zendesk gives two answers on which plans have it. The setup article lists every Suite plan from Team up, and Support Team, Professional and Enterprise, while the announcement says Suite Growth and above, and Support Team and above. Whichever is right, this is Zendesk’s AI reaching out to your other tools, so it’s no help if what you want is Claude or Cursor reading your tickets.

What to check before an AI tool reads your Zendesk

These apply whichever route you pick.

• Know what the sign-in reaches. Zendesk says an API token can be used by any verified user, or anyone with their email address, so a token in a shared config opens the whole account. Signing in through Swifteq or reminia’s server acts as you, with your Zendesk role, and scopes such as tickets:read users:read hc:read narrow it to reading.
• Remember who writes your tickets. Anyone who emails your support address writes text that your AI tool will read, and a message can carry instructions aimed at an AI. Don’t let your AI tool act on instructions it finds in a ticket, and keep its approval prompt on for anything that changes Zendesk, so a human in the loop says yes to each write.
• Check what a comment does by default. In Zendesk a comment is either public, which the customer gets, or an internal note. reminia’s create_ticket_comment, Pipedream’s Update Ticket and mattcoatsworth’s update_ticket post public by default, while Swifteq writes internal notes until an owner allows public replies.
• Decide whether a company should sit in the middle. A hosted server sees your tickets on the way through. A server on your own computer, like reminia’s, passes them only between Zendesk and your AI tool.

Where Deacon fits beside Zendesk

If you’re the one answering in Zendesk, a ticket queue may be more than your support needs. Deacon handles support in the chat on your website instead. It answers in seconds, at any hour, from your docs, and when they don’t cover a question it says so rather than guessing, then asks for the visitor’s email. You reply to the rest from Deacon’s own dashboard, so there’s no helpdesk to run. And where Swifteq’s tools work out first reply times across a queue, Analytics › Widget tells you how many conversations Deacon handled with nobody from your team, and lists the features visitors asked you to build, in their own words.

Deacon’s MCP server is its own, and you can connect it today, so there’s no other company between your chats and your agent and no token to paste in. Claude Code, Codex, Cursor or VS Code signs in through your browser, an owner on your team presses Allow, and every owner gets an email whenever an agent connects or disconnects. It’s also built for your coding agent, where the routes above are for working tickets. Your coding agent reads what customers said was broken and on which page, beside the JavaScript errors Deacon’s script caught on your site, and goes after the cause in your code.

The checks in the last section apply here too, and Deacon’s server is built around them. Everything a visitor typed reaches your agent under one untrusted field, so it reads as data rather than as instructions. There’s no tool that replies to anyone, so there’s no public or private setting to get wrong. What it can change stays inside Deacon, such as a suggested answer that waits for you to save it before any customer sees it.

Zendesk is built for a support team, and it has plenty Deacon leaves out. It has the ticket queue, views, macros, triggers and SLA policies the tools above read and change, and AI agents that work across messaging, email and voice. Deacon has no ticketing, email inbox, phone channel or saved replies. If a team works your tickets every day, keep Zendesk and connect it through Swifteq’s server. If support is you, Deacon answers the questions your docs already cover, and your coding agent reads the bug reports straight from it.

Support tools with an MCP server sets Zendesk beside 31 other support tools, and if you’re weighing up Zendesk itself, Zendesk alternatives looks at what else a small team can use.

Questions about Zendesk’s MCP server

Does Zendesk have an official MCP server?

Not one you can use yet. Zendesk announced it on 19 May 2026 for early access that summer. On 10 November 2026 it had published no docs, no early access sign-up and no price for it. The MCP Server app in its Marketplace is Swifteq’s.

Can I still create a Zendesk API token?

No. Since 27 October 2026 no account can create one, and accounts created on or after 28 July 2026 can’t use them at all. A token you already have keeps working until 30 April 2027, as long as it’s used at least once every 30 days. Anything you set up now should sign in with OAuth.

How do I connect Zendesk to Claude?

The quickest way is Swifteq’s server. Sign up at Swifteq, authorise your Zendesk, then add https://mcp.swifteq.com/api/mcp/mcp as a custom connector in Claude. Even Claude’s free plan allows one. Zapier and Pipedream work in Claude too, and reminia’s server runs in Claude Desktop or Claude Code if you’d rather keep it on your own computer.

Can Claude or ChatGPT reply to Zendesk tickets?

Yes, through every route here (in ChatGPT, only on Business or Enterprise), so decide whether you want it to. Swifteq only writes internal notes until an account owner allows public replies, and Zapier’s Add Comment to Ticket asks whether each comment should be public. reminia’s and Pipedream’s servers post public comments unless told otherwise.

Which Zendesk plan do I need?

Zendesk’s API, and the OAuth these servers sign in with, come on every Suite plan from Team up and on Support Team, Professional and Enterprise. Swifteq says some of the searches it relies on need Suite Growth or higher. Zendesk pricing has what each plan costs.

Do I need Zendesk if I answer support myself?

Not for the people who ask on your site. Deacon replies to them from your docs, keeps a list of what it couldn’t answer, and you write back to those from its dashboard. Zendesk earns its place once a team works email, phone and messaging queues, which Deacon doesn’t cover, since it has no email inbox, phone channel or ticketing. If you keep Zendesk for those, the two run side by side with no built-in connection.