Zendesk CLI, MCP Server & Claude Code Skill

A Claude Code skill for Zendesk Support integration. The primary interface is a skill file (SKILL.md) that gives Claude full command knowledge — backed by a CLI (zd-cli) that Claude runs to interact with Zendesk. An MCP server is also included for other AI assistants.

This project was built almost entirely by Claude Code through iterative conversation — designing, implementing, testing against a live Zendesk instance, and fixing issues together over multiple sessions. It's a tool I use often and am sharing in case others find it useful.

Features

• Claude Code skill — SKILL.md gives Claude full command documentation, workflows, and the save-first/query-later pattern
• CLI (zd-cli) — 28 commands covering tickets, users, organizations, views, metrics, and more
• MCP server — alternative interface for other AI assistants
• Slack integration — send formatted support reports to Slack channels
• Support metrics analysis — response times, resolution rates, CSAT, after-hours and on-call tracking
• Local response storage — API responses saved to temp directory, queryable with jq without re-fetching
• Markdown formatting — write operations convert Markdown to HTML for proper rendering in Agent Workspace
• Encrypted credentials — API tokens and OAuth tokens encrypted at rest

Quick Start

# Run directly (no install needed)
uvx zd-cli --help

# Set up credentials (see Authentication below)
uvx zd-cli auth login

# Verify it works
uvx zd-cli me

Authentication

Three methods are supported. API token is the simplest; OAuth adds auto-refresh. A relay server mode exists for shared team deployments but is not covered here.

Method 1: API Token

1. In Zendesk: Admin Center → Apps and integrations → APIs → Zendesk API → Add API token
2. Copy the token (shown only once)

# Interactive setup (prompts for email, token, subdomain)
zd-cli auth login

# Or non-interactive
zd-cli auth login --email you@company.com --token YOUR_TOKEN --subdomain yourcompany

# Or via environment variables
export ZENDESK_EMAIL="you@company.com"
export ZENDESK_TOKEN="your-api-token"
export ZENDESK_SUBDOMAIN="yourcompany"

Method 2: OAuth 2.0 (with auto-refresh)

1. In Zendesk: Admin Center → Apps and integrations → APIs → OAuth Clients → Add OAuth client
2. Add redirect URIs for ports 8080–8089:

   http://127.0.0.1:8080/callback
   http://127.0.0.1:8081/callback
   ...
   http://127.0.0.1:8089/callback
   

3. Note the Client ID and Client Secret

# Opens a browser for authorization
zd-cli auth login-oauth --subdomain yourcompany --client-id YOUR_ID --client-secret YOUR_SECRET

# Headless/SSH: browser redirects to localhost (shows connection refused — that's expected),
# paste the full redirect URL back into the prompt
zd-cli auth login-oauth --subdomain yourcompany --client-id YOUR_ID --client-secret YOUR_SECRET --manual

# Check auth status
zd-cli auth status

# Remove credentials
zd-cli auth logout          # API token
zd-cli auth logout-oauth    # OAuth token

Installation

Prerequisites: Python 3.12+, uv, and jq (for the query command). On macOS, brew install uv jq installs both.

Via uvx (no install, recommended)

uvx zd-cli --help

uvx runs the tool in an isolated environment on demand — nothing is permanently installed.

Via uv tool install

uv tool install zd-cli
zd-cli --help
uv tool upgrade zd-cli    # later, to update

Development / Claude Code skill

git clone https://github.com/andmarios/zendesk-skill zendesk-skill
cd zendesk-skill
uv sync
uv run zd-cli --help

Point Claude Code at SKILL.md to give Claude full command documentation and workflows.

Commands

Tickets

Command	Description
search "status:open priority:urgent"	Search tickets
ticket 12345	Get ticket by ID
ticket-details 12345	Ticket + all comments
linked-incidents 12345	Incidents linked to a problem ticket
attachment --ticket 12345 <url>	Download an attachment

Write Operations

All write commands convert Markdown to HTML by default (use --plain-text to skip). Bare URLs are linked and lists get a blank line after them. Raw HTML inside Markdown is escaped; content that starts with an HTML tag is sent unchanged.

Command	Description
update-ticket 12345 --status pending --tags "waiting-customer"	Update ticket
create-ticket "Subject" "**Bold** description"	Create ticket
add-note 12345 "Internal **note**"	Add internal note
add-comment 12345 "Public reply"	Add public comment

Metrics & Analytics

Command	Description
ticket-metrics 12345	Reply/resolution times for a ticket
list-metrics	Metrics across tickets
satisfaction-ratings --score bad	CSAT ratings

Views

Command	Description
views	List available views
view-count 123	Ticket count in a view
view-tickets 123	Tickets from a view

Users & Organizations

Command	Description
user 12345	Get user by ID
search-users "john@example.com"	Search users
org 67890	Get organization by ID
search-orgs "Acme"	Search organizations

Configuration & Info

Command	Description
me	Current user (tests auth)
groups	List support groups
tags	Popular tags
sla-policies	SLA policies

Query

# Query a saved response file with jq
zd-cli query <file> -q comments_slim       # named query
zd-cli query <file> --jq '.data.ticket'    # custom jq
zd-cli query <file> --list                 # show available named queries

All commands save their full API response to /tmp/zd-cli-$UID/ and print the path. Use zd-cli query to extract data from saved files without re-fetching.

Search Query Syntax

# Status & priority
zd-cli search "status:open priority:urgent"
zd-cli search "status:pending assignee:me"

# Time filters
zd-cli search "created>2024-01-01 status:open"
zd-cli search "updated<1week status:open"

# By person/org
zd-cli search "requester:user@example.com status:open"
zd-cli search "organization:acme type:incident"

# Tags
zd-cli search "tags:billing tags:escalated"

Slack Integration

# Configure
zd-cli auth login-slack --webhook "https://hooks.slack.com/services/..." --channel "#support"

# Send a report
zd-cli slack-report

# Check / remove
zd-cli auth status-slack
zd-cli auth logout-slack

Support Metrics Analysis

# Default: last 2 weeks
uv run python src/zendesk_skill/scripts/analyze_support_metrics.py

# Custom range with accurate reply counts from the Ticket Metrics API
uv run python src/zendesk_skill/scripts/analyze_support_metrics.py \
  --start 2026-01-01 --end 2026-01-15 --fetch-metrics

# Pipe to Slack
uv run python src/zendesk_skill/scripts/analyze_support_metrics.py | zd-cli slack-report

# Generate markdown report
zd-cli markdown-report -o report.md

Business Hours & On-Call Tracking

Configure in ~/.config/zd-cli/config.json to add after-hours and on-call sections to your reports:

{
  "business_hours": {
    "timezone": "Europe/Berlin",
    "start_hour": 9,
    "end_hour": 18,
    "workdays": [0, 1, 2, 3, 4]
  },
  "oncall": {
    "enabled": true,
    "start_hour": 19,
    "end_hour": 9,
    "customers": ["bigclient.com"],
    "priorities": ["urgent"]
  }
}

When configured, reports include after-hours ticket counts, messages, replies, and on-call engagements (tickets matching the configured priority + customer domain during on-call hours). FRT for on-call tickets is calculated in calendar time; all others use business hours.

Security: Prompt Injection Protection

All content from Zendesk is treated as untrusted input and screened through a multi-layer security pipeline before reaching the LLM:

1. Regex detection — known prompt injection patterns (instruction overrides, prompt extraction, leetspeak evasion)
2. Semantic matching — fuzzy similarity against known injection templates
3. Haiku/LLM screening — when configured, an LLM-based classifier for sophisticated attacks

Every LLM-facing field (ticket subjects, comment bodies, user names, emails, org names, view titles) is wrapped with session-scoped security markers. Stored response files are also scanned at save time using both regex and semantic screening tiers, with detection results preserved in file metadata.

Attachment scanning is size-aware:

• Text files up to 1 MB are scanned inline through the full screening pipeline
• Text files over 1 MB and binary files receive a security warning with a hint to scan manually using the CLI tool:

  uvx prompt-security-utils <file>
  

Security is enabled by default. Configure in ~/.config/zd-cli/config.json:

{
  "security_enabled": true,
  "allowlisted_tickets": ["12345", "67890"]
}

• security_enabled — set to false to disable all screening and wrapping (default: true)
• allowlisted_tickets — ticket IDs to skip wrapping (for trusted/internal tickets)

# View current security status and session markers
zd-cli security-info

# Include full MCP security instructions
zd-cli security-info --instructions

Requires prompt-security-utils (included as a dependency).

Local file permissions

Saved responses, exports and downloaded attachments contain customer data, so they are written owner-only: directories 0700, files 0600. The tool refuses to use /tmp/zd-cli-$UID/ if it is a symlink or owned by another user, and it never changes permissions on anything outside that directory. If a directory cannot be secured the command fails with a File system error message; if only a file's mode cannot be tightened, a warning is printed on stderr.

MCP Server

The MCP server exposes the same functionality for AI assistants that support the Model Context Protocol.

uv run zendesk-mcp

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "zendesk": {
      "command": "uvx",
      "args": ["--from", "zendesk-skill", "zendesk-mcp"]
    }
  }
}

Development

uv run pytest -v
uv run zd-cli --help

License

MIT