zendesk-mcp

Read-only MCP server for auditing Zendesk Support tickets and Chat (Zopim) transcripts from Claude.

Why

Fully open-source, GET-only Zendesk MCP — no writes, no side effects, no hidden telemetry. Built for searching, fetching and reading tickets and chat transcripts so an LLM can audit support communication quality. Every tool issues only GET requests; there are no POST/PUT/PATCH/DELETE paths in the codebase.

Setup

1. Build

git clone https://github.com/mikhailrojo/zendesk-mcp.git
cd zendesk-mcp
yarn install --immutable
yarn build

Requires Node.js 20+.

2. Get a Zendesk API token

Admin Center → Apps and integrations → APIs → Zendesk API → enable Token access → Add API token. Copy it (shown once). Note the agent email and subdomain (<sub> in https://<sub>.zendesk.com).

3. Add to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) and add the zendesk entry under mcpServers:

{
  "mcpServers": {
    "zendesk": {
      "command": "/absolute/path/to/node",
      "args": [
        "/absolute/path/to/zendesk-mcp/dist/index.js"
      ],
      "env": {
        "ZENDESK_SUBDOMAIN": "yourcompany",
        "ZENDESK_EMAIL": "agent@yourcompany.com",
        "ZENDESK_API_TOKEN": "..."
      }
    }
  }
}

Restart Claude Desktop. For Claude Code, use the same structure in its MCP config.

Tools

Support API

| Name | What it does | |---|---| | search_tickets | Search tickets via Zendesk Search (ZQL). Auto-prefixes type:ticket. | | get_ticket | Ticket metadata (no comments). | | get_ticket_comments | Full comment history with authors, timestamps, attachments. Auto-paginated. | | list_views | Active Zendesk views. | | get_tickets_from_view | Tickets inside a view. | | list_groups | Support groups. |

Chat API (Zopim)

| Name | What it does | |---|---| | get_chat | Single chat session with normalized transcript. | | search_chats | Search chats by text, timestamp:[X TO Y], or tag:foo. | | list_chats | Incremental export of chat ids since start_time. | | get_chat_for_ticket | Resolves the chat that produced a Support ticket — looks up requester, matches by name/email within +/- window_minutes of created_at, returns transcript + candidates. |

All tools return JSON in content[0].text. Chat transcripts are arrays of { type: "message" | "event", author_type, author_name, timestamp, body, ... } — filter type === "message" for conversation turns.

Example prompts

• "For ticket 869955 use get_chat_for_ticket and score the agent on tone, empathy and resolution."
• "Search tickets status:solved created>2026-04-01 group_id:12345, then get_ticket_comments for each and flag responses lacking empathy."
• "Use list_chats for yesterday, then get_chat on each to find sessions where the agent took >60s to first respond."

Security

• Read-only — only GET requests.
• Token via env only — never commit; .env is gitignored.
• stdout reserved for MCP JSON-RPC; logs go to stderr.
• PII (customer emails, message text) is intentionally returned to the LLM. Only run against authorized tenants.

yarn audit            # yarn audit + trivy

Troubleshooting

• 401 — wrong token/email, or agent lacks API access. For Chat endpoints, confirm Chat product is enabled.
• 429 — client retries 3x honoring Retry-After. Lower per_page if persistent.
• Empty search_tickets — Zendesk Search has eventual consistency (minutes).
• Server missing in /mcp — confirm absolute paths, run node dist/index.js manually with env vars; should print Server ready on stdio transport.