zendesk-mcp-server

# Zendesk MCP Server

[![Glama score](https://glama.ai/mcp/servers/fruggr/zendesk-mcp-server/badges/score.svg)](https://glama.ai/mcp/servers/fruggr/zendesk-mcp-server)
[![MCP Registry](https://img.shields.io/badge/MCP_Registry-io.github.fruggr%2Fzendesk--mcp--server-0a7ea4)](https://registry.modelcontextprotocol.io/?search=io.github.fruggr/zendesk-mcp-server)
[![npm version](https://img.shields.io/npm/v/@fruggr/zendesk-mcp-server?logo=npm&color=cb3837)](https://www.npmjs.com/package/@fruggr/zendesk-mcp-server)
[![License: MIT](https://img.shields.io/npm/l/@fruggr/zendesk-mcp-server?color=blue)](LICENSE)
[![Node.js](https://img.shields.io/node/v/@fruggr/zendesk-mcp-server?logo=nodedotjs&logoColor=white&color=339933)](https://nodejs.org)

A [Model Context Protocol](https://modelcontextprotocol.io) (MCP) server that
puts Zendesk inside your AI assistant. It finds answers in the Help Center;
drafts, updates and translates articles while keeping the languages in sync; and
handles Support tickets end to end, comments, triage and image attachments
included. It all happens in plain language, without switching apps.

It does roughly what the
[Zendesk agent for Microsoft 365 Copilot](https://support.zendesk.com/hc/en-us/articles/9958331458458-Using-the-Zendesk-agent-in-Microsoft-365-Copilot)
does, minus the tie to one vendor: it drops into any MCP client (Claude Desktop,
Claude Code, Cursor, VS Code, and the rest). And it always acts with each user's
own Zendesk permissions, never a shared admin key.

## What your assistant can do

Ask in natural language; the assistant works out the context and the intent,
then calls the right tools on your behalf.

- Find answers in the Help Center. "How do I request a software license?" or
  "what's the time-off policy?" surfaces the right article, matched by meaning
  rather than by keyword.
- Create, view and update tickets without leaving the conversation: open a
  ticket, check its status, add a public reply or an internal note, change the
  priority or the assignee, mark it solved.
- Summarize a ticket for a report or a quick decision. The assistant pulls the
  details and the full comment thread and gives you the gist in a sentence.
- Read the screenshots and photos attached to a ticket. Error dialogs, UI
  captures and product photos are handed to your assistant's own model as
  images, so it can describe them or act on what they show.
- Search and triage your queue in plain language: "show me my open tickets about
  billing from this week."
- Draft and maintain knowledge-base articles. You can write a new one, or revise
  a large one a single section at a time, so the whole HTML body never has to
  round-trip through the model.

## Why this server

Most Zendesk integrations run on a shared admin API key, which hands every user
full access to every ticket, and bolt on a fixed set of tools. This one is built
differently.

- Per-user authentication, OAuth only. Both transports use OAuth 2.1 PKCE: each
  user signs in with their own Zendesk credentials, so the assistant sees and
  touches exactly what that person is allowed to, the same scoping you get by
  signing into Zendesk directly. Static API tokens are deliberately not
  supported ([why](#what-this-server-does-not-do)).
- Section-based article editing. For large Help Center articles, read and
  rewrite one section at a time (parsed by `h1`/`h2`/`h3` headings) instead of
  shuffling the full HTML body through the assistant. On a targeted edit that
  cuts tokens by a factor of 10 to 100.
- Native multimodal attachments. Ticket images come back as native MCP image
  content, so the client's own model (Claude, GPT, Gemini, whichever) sees the
  pixels directly. No server-side vision model, no extra API key, and nothing
  tying you to one provider. Non-image attachments come back as text references,
  and both image caps are configurable.
- A tool surface you can cap. Expose every operation as its own tool, group them
  into namespace proxies, or collapse everything into a single unified tool. You
  can also filter by namespace or down to read-only operations, so each context
  loads only the surface it needs (see [Tool surface](#tool-surface)).
- Two deployment shapes, same auth story. Run it on your laptop as a stdio MCP
  server, or deploy it as a private remote MCP server reached over HTTP, with one
  Zendesk session per request and each client carrying its own user's token.
- A lean stack: the official `@modelcontextprotocol/sdk` plus `zod`, speaking to
  the Zendesk Support and Help Center (Guide) APIs.

Look elsewhere when:

- You need Zendesk products outside Support and Guide (Talk, Explore analytics,
  Sell). Those endpoints aren't covered.
- You need a single shared service account, or static API-token auth. This
  server supports neither, by design (see below).

## What this server does *not* do

There is no API-token authentication. The server speaks OAuth 2.1 PKCE and
nothing else: no `ZENDESK_EMAIL` + `ZENDESK_API_TOKEN` (Basic auth) mode, in any
transport. That is deliberate, for two reasons.

1. **API tokens are insufficiently secure.** A Zendesk API token is a
   long-lived, static, shared secret that carries the full rights of the user
   who issued it. There is no per-user scoping, no short expiry, and no per-user
   consent or revocation. OAuth 2.1 PKCE issues per-user, revocable tokens
   instead.
2. **API tokens don't scale.** A single static credential can't attribute
   actions to individual users, and it can't be revoked granularly. It also
   makes a multi-user remote deployment unsafe: over HTTP it would expose the
   issuing user's rights to every caller.

If you specifically need an API-token or service-account mode (headless CI with
a shared account, say), use one of the other Zendesk MCP servers that support
it. A few are listed under
[Inspiration & related projects](#inspiration--related-projects).

## Quick start: local (stdio)

The default shape: one developer, one Zendesk account, OAuth 2.1 PKCE in the
browser. You need **Node.js >= 20** and a **Zendesk** instance (Support or
Suite).

### Install

```bash
# Run without installing
npx -y @fruggr/zendesk-mcp-server <your-subdomain>

# Or install globally
npm install -g @fruggr/zendesk-mcp-server
zendesk-mcp-server <your-subdomain>
```

Signing in needs a Zendesk OAuth client, so register one first (next section).

### Zendesk OAuth setup

1. Go to **Admin Center → Apps and integrations → APIs → OAuth Clients**
2. Create a **public** client:
   - **Identifier**: `<your-subdomain>_zendesk` (or set `ZENDESK_OAUTH_CLIENT_ID`)
   - **Redirect URL**: `http://localhost:27439/callback` (change the port to match
     `ZENDESK_OAUTH_CALLBACK_PORT` / `--callback-port` if you override it; Zendesk
     accepts several redirect URLs, one per line)

On the first tool call the server starts the sign-in flow: it opens a browser
window and returns the authorize URL in a tool message. The call does not block
waiting for sign-in, so authenticate in the browser and then retry the request.
The token is persisted to an owner-only file and reused across restarts, so you
don't authenticate again every time your MCP client respawns the server (path
and overrides: [`ZENDESK_TOKEN_FILE`](docs/configuration.md#zendesk_token_file)).

### MCP client wiring

<details>
<summary><strong>Claude Desktop</strong></summary>

Add to your `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "zendesk": {
      "command": "npx",
      "args": ["-y", "@fruggr/zendesk-mcp-server", "<your-subdomain>", "--mode", "single"]
    }
  }
}
```

</details>

<details>
<summary><strong>Claude Code</strong></summary>

```bash
claude mcp add zendesk -- npx -y @fruggr/zendesk-mcp-server <your-subdomain> --mode single
```

</details>

<details>
<summary><strong>VS Code (Copilot / Continue / Cline)</strong></summary>

Add to your `.vscode/mcp.json`:

```json
{
  "servers": {
    "zendesk": {
      "command": "npx",
      "args": ["-y", "@fruggr/zendesk-mcp-server", "<your-subdomain>", "--mode", "single"]
    }
  }
}
```

</details>

Something not working? See [Troubleshooting](docs/troubleshooting.md).

## Quick start: remote (HTTP)

> **Experimental.** The HTTP transport ships, but it has not yet been exercised
> end-to-end against a real Zendesk tenant from every MCP client. Local stdio is
> the supported path.

You can also deploy a private remote MCP server for **one** Zendesk account,
where every MCP client presents its **own** user's OAuth bearer in
`Authorization:` and the server never sees a shared admin key. The full guide
covers OAuth setup, `--public-url` behind a reverse proxy, per-platform config,
the discovery endpoints, MCP client wiring, CORS and what stays the operator's
job: **[docs/http-deployment.md](docs/http-deployment.md)**.

## Tool surface

Tools are grouped into four namespaces: **Tickets**, **Help Center**, **Users &
Organizations** and **Search**. The server registers them in one of three modes,
so you can trade granularity against context budget:

- **`all`**: every operation as its own tool, for clients with good tool selection;
- **`namespace`** (default): one proxy tool per namespace, a balanced middle ground;
- **`single`**: a single `zendesk` tool, for models with limited tool slots.

Proxies take `{ "operation": "<tool_name>", "params": { … } }` and validate
`params` through the original schema. `--namespace`, `--tool` and `--read-only`
filter tools *before* the proxies are built, so each proxy describes only the
operations that survive.

Every tool with its description and its `read`/`write` mode:
**[docs/mcp-tools-reference.md](docs/mcp-tools-reference.md)**. The flags and
worked examples: **[docs/configuration.md](docs/configuration.md)**.

## Help Center context

Beyond tools, the server hands the LLM the structure of *your* Help Center: the
active locales, the category → section tree with IDs, the visibility segments
and the permission groups. With thos