zscaler-mcp-server

zscaler-mcp-server is a Model Context Protocol (MCP) server that connects AI agents with the Zscaler Zero Trust Exchange platform. By default, the server operates in read-only mode for security, requiring explicit opt-in to enable write operations.

Support Disclaimer

-> Disclaimer: Please refer to our General Support Statement before proceeding with the use of this provider. You can also refer to our troubleshooting guide for guidance on typical problems.

[!IMPORTANT] 🚧 Public Preview: This project is currently in public preview and under active development. Features and functionality may change before the stable 1.0 release. While we encourage exploration and testing, please avoid production deployments. We welcome your feedback through GitHub Issues to help shape the final release.

📄 Table of contents

• 📺 Overview
• 🔒 Security & Permissions
• 🔐 MCP Client Authentication
• Supported Tools
• Installation & Setup
  • Prerequisites
  • Environment Configuration
  • Installation
• Usage
  • Command Line
  • Service Configuration
  • Additional Command Line Options
• Zscaler API Credentials & Authentication
  • OneAPI Authentication
  • Authentication Troubleshooting
  • MCP Server Configuration
• As a Library
• Container Usage
  • Using Pre-built Image (Recommended)
  • Building Locally (Development)
• Editor/Assistant Integration
  • Using uvx (recommended)
  • With Service Selection
  • Using Individual Environment Variables
  • Docker Version
• Additional Deployment Options
  • Remote MCP Deployment (EC2, VM, etc.)
  • Azure Container Apps / Virtual Machine / AKS (Preview)
  • Google Cloud (Cloud Run / GKE / VM / ADK Agent)
  • Kubernetes (Helm Chart)
  • Amazon Bedrock AgentCore
• Using the MCP Server with Agents
  • Claude Desktop
  • Cursor
  • Visual Studio Code + GitHub Copilot
• Platform Integrations
• Troubleshooting
• License

📺 Overview

The Zscaler Integrations MCP Server brings context to your agents. Try prompts like:

• "List my ZPA Application segments"
• "List my ZPA Segment Groups"
• "List my ZIA Rule Labels"

[!WARNING] 🔒 READ-ONLY BY DEFAULT: For security, this MCP server operates in read-only mode by default. Only list_* and get_* operations are available. To enable tools that can CREATE, UPDATE, or DELETE Zscaler resources, you must explicitly enable write mode using the --enable-write-tools flag or by setting ZSCALER_MCP_WRITE_ENABLED=true. See the Security & Permissions section for details.

[!TIP] Writing effective prompts: This server exposes 402 tools across multiple Zscaler services. Most MCP clients (Claude Desktop, Cursor, etc.) use deferred tool loading and will search for relevant tools based on your prompt. For best results, be specific about the service and action in your prompts:

• Good: "List my ZPA application segments" — targets the right service and tool directly
• Good: "Show ZIA firewall rules" — clear service (zia) and action (list)
• Less effective: "Show me my devices" — ambiguous; multiple services expose device-related tools

When a service is disabled, its tools are fully removed from the server. However, the AI agent may still attempt to find related tools in other services. If you get unexpected results, refine your prompt with the specific service name (e.g. zpa, zia, zdx, zcc, zcell, zms).

🔒 Security & Permissions

The Zscaler MCP Server implements a security-first design with granular permission controls and safe defaults:

Read-Only Mode (Default - Always Available)

By default, the server operates in read-only mode, exposing only tools that list or retrieve information:

• ✅ ALWAYS AVAILABLE - Read-only tools are registered by the server
• ✅ Safe to use with AI agents autonomously
• ✅ No risk of accidental resource modification or deletion
• ✅ All list_* and get_* operations are available (110+ read-only tools)
• ❌ All create_*, update_*, and delete_* operations are disabled by default
• 💡 Note: You may need to enable read-only tools in your AI agent's UI settings

# Read-only mode (default - safe)
zscaler-mcp

When the server starts in read-only mode, you'll see:

🔒 Server running in READ-ONLY mode (safe default)
   Only list and get operations are available
   To enable write operations, use --enable-write-tools AND --write-tools flags

💡 Read-only tools are ALWAYS registered by the server regardless of any flags. You never need to enable them server-side. Note: Your AI agent UI (like Claude Desktop) may require you to enable individual tools before use.

Write Mode (Explicit Opt-In - Allowlist REQUIRED)

To enable tools that can create, modify, or delete Zscaler resources, you must provide BOTH flags:

1. ✅ --enable-write-tools - Global unlock for write operations
2. ✅ --write-tools "pattern" - MANDATORY explicit allowlist

🔐 SECURITY: Allowlist is MANDATORY - If you set --enable-write-tools without --write-tools, 0 write tools will be registered. This ensures you consciously choose which write operations to enable.

# ❌ WRONG: This will NOT enable any write tools (allowlist missing)
zscaler-mcp --enable-write-tools

# ✅ CORRECT: Explicit allowlist required
zscaler-mcp --enable-write-tools --write-tools "zpa_create_*,zpa_delete_*"

When you try to enable write mode without an allowlist:

⚠️  WRITE TOOLS MODE ENABLED
⚠️  NO allowlist provided - 0 write tools will be registered
⚠️  Read-only tools will still be available
⚠️  To enable write operations, add: --write-tools 'pattern'

Write Tools Allowlist (MANDATORY)

The allowlist provides two-tier security:

1. ✅ First Gate: --enable-write-tools must be set (global unlock)
2. ✅ Second Gate: Explicit allowlist determines which write tools are registered (MANDATORY)

Allowlist Examples:

# Enable ONLY specific write tools with wildcards
zscaler-mcp --enable-write-tools --write-tools "zpa_create_*,zpa_delete_*"

# Enable specific tools without wildcards
zscaler-mcp --enable-write-tools --write-tools "zpa_create_application_segment,zia_create_rule_label"

# Enable all ZPA write operations (but no ZIA/ZDX/ZTW)
zscaler-mcp --enable-write-tools --write-tools "zpa_*"

Or via environment variable:

export ZSCALER_MCP_WRITE_ENABLED=true
export ZSCALER_MCP_WRITE_TOOLS="zpa_create_*,zpa_delete_*"
zscaler-mcp

Wildcard patterns supported:

• zpa_create_* - Allow all ZPA creation tools
• zpa_delete_* - Allow all ZPA deletion tools
• zpa_* - Allow all ZPA write tools
• *_application_segment - Allow all operations on application segments
• zpa_create_application_segment - Exact match (no wildcard)

When using a valid allowlist, you'll see:

⚠️  WRITE TOOLS MODE ENABLED
⚠️  Explicit allowlist provided - only listed write tools will be registered
⚠️  Allowed patterns: zpa_create_*, zpa_delete_*
⚠️  Server can CREATE, MODIFY, and DELETE Zscaler resources
🔒 Security: 85 write tools blocked by allowlist, 8 allowed

Tool Design Philosophy

Each operation is a separate, single-purpose tool with explicit naming that makes its intent clear:

✅ Good (Verb-Based - Current Design)

zpa_list_application_segments    ← Read-only, safe to allow-list
zpa_get_application_segment      ← Read-only, safe to allow-list
zpa_create_application_segment   ← Write operation, requires --enable-write-tools
zpa_update_application_segment   ← Write operation, requires --enable-write-tools
zpa_delete_application_segment   ← Destructive, requires --enable-write-tools

This design allows AI assistants (Claude, Cursor, GitHub Copilot) to:

• Allow-list read-only tools for autonomous exploration
• Require explicit user confirmation for write operations
• Clearly understand the intent of each tool from its name

Responses: the API record, verbatim

A read tool returns the Zscaler API record unchanged. The server does not trim, rename, or re-declare a resource's attributes — that attribute set belongs to the API, so a field Zscaler adds in future reaches you with no server upgrade.

To keep responses small, you control what comes back rather than the server guessing. Two levers:

• --toolsets — load only the slice of tools you need, so the tool catalog stays small (see Toolsets).

• query — every list tool accepts an optional JMESPath expression applied to the results, so the agent projects exactly what it wants:

  zcc_list_devices(query="[*].{user: user, policy: policyName}")   # just those two fields
  zcc_list_devices(query="[?registrationState=='Quarantined']")    # just quarantined devices
  zcc_list_devices(query="length(@)")                              # just the count
  

  Field names are exactly what the Zscaler API returns. Omit query to get the full records.

Security Layers

The server implements multiple layers of security (defense-in-depth). The first nine apply on every transport, including stdio — they govern which tools are exposed and how dangerous calls are confirmed. The remaining HTTP-only layers (TLS, host-header validation, source-IP ACL, MCP client authentication) are described in the Network-Level Controls section further down.

1. Read-Only Tools Always Enabled: Safe list_* and get_* operations are always available (254 tools).
2. Default Write Mode Disabled: Write tools are disabled unless explicitly enabled via --enable-write-tools.
3. Mandatory Allowlist: Write operations require explicit --write-tools allowlist (wildcard support).
4. OneAPI Entitlement Filter: At startup, toolsets for products the OneAPI credentials cannot call are silently dropped (see OneAPI Entitlement Filter below).
5. Toolset Selection: Optionally narrow the registered tool surface to a specific slice (e.g. --toolsets zia_url_filtering,zpa_app_segments). See the Toolsets section below.
6. Verb-Based Tool Naming: Each tool clearly indicates its purpose (list, get, create, update, delete).
7. Tool Metadata Annotations: All tools are annotated with readOnlyHint or destructiveHint for AI agent frameworks.
8. AI Agent Confirmation: All write tools marked with destructiveHint=True trigger permission dialogs in AI assistants.
9. Human Confirmation for DELETE: Delete operations are gated by a server-initiated confirmation. On clients that support MCP elicitation (Claude Desktop, Cursor), the server asks the client to prompt a human and the answer arrives as a protocol field — so a hijacked agent cannot author the approval, and a failed round trip fails closed. Clients without that capability fall back to a cryptographic confirmation token (HMAC-SHA256, single-use, 5-minute TTL, bound to the exact tool and parameters), which stops an approval being spent on another resource, replayed or reused — but is not by itself an anti-prompt-injection control, since a hijacked agent receives and can redeem it. The fallback is also single-process; elicitation-capable clients are unaffected. See MCP protocol posture — confirmation threat model for the full analysis.
10. Environment Variable Control: ZSCALER_MCP_WRITE_ENABLED, ZSCALER_MCP_WRITE_TOOLS, ZSCALER_MCP_TOOLSETS, ZSCALER_MCP_DISABLE_ENTITLEMENT_FILTER, and the disable lists can all be managed centrally without code changes.
11. Output Sanitization: Every string in every tool result is run through a three-stage sanitizer before reaching the agent — invisible/control characters (BiDi overrides, zero-width chars, BOM, soft hyphen) are stripped, raw HTML and HTML comments are removed (via bleach), Markdown link/image syntax is neutralised so embedded URLs cannot be smuggled to the agent, and Markdown code-fence info-strings containing role-impersonation tokens (system, assistant, tool, ignore, …) are collapsed to a neutral text tag. This defends against prompt-injection payloads that an attacker — or a careless admin — might embed in editable Zscaler resources (rule descriptions, location names, label descriptions, etc.). On by default. Opt-out with ZSCALER_MCP_DISABLE_OUTPUT_SANITIZATION=true (use only for diagnostics).
12. Audit Logging: When --log-tool-calls / ZSCALER_MCP_LOG_TOOL_CALLS=true is set, every tool invocation is logged with its arguments (sensitive values redacted), duration, and a result summary.

This multi-layered approach ensures that even if one security control is bypassed, others remain in place to prevent unauthorized operations. Layers 1-12 above apply equally to stdio, sse, and streamable-http.

Toolsets

Tools are grouped into 63 named toolsets so you can load only the slice an agent actually needs (e.g. zia_url_filtering (5 tools) instead of every tool from every service (402)). Toolsets reduce the agent's context cost and improve tool-selection accuracy.

# Load just two slices
zscaler-mcp --toolsets zia_url_filtering,zpa_app_segments

# Or use the curated default-on subset
zscaler-mcp --toolsets default

# Or load every registered toolset explicitly
zscaler-mcp --toolsets all

# Equivalent via environment variable
export ZSCALER_MCP_TOOLSETS="zia_url_filtering,zpa_app_segments"

When --toolsets is unspecified, every toolset whose service is enabled is loaded (preserves the historical default).

The agent can also enable additional toolsets at runtime through the always-on zscaler_list_toolsets, zscaler_get_toolset_tools, and zscaler_enable_toolset tools.

For the full catalog (29 toolsets across all services), filter precedence rules, per-toolset agent guidance, and the complete reference, see docs/guides/toolsets.md.

OneAPI Entitlement Filter

After your toolset selection resolves, the server reads the product entitlements from the OneAPI bearer token issued for your ZSCALER_CLIENT_ID and silently drops toolsets for products the credentials cannot call. If your OneAPI client is only entitled to ZIA and ZPA, every zdx_* / zcc_* / ztw_* / zid_* / zeasm_* / zins_* / zms_* toolset is filtered out at startup — even with --toolsets all.

This prevents an agent from discovering tools whose first call would only ever return 401 Unauthorized. The filter applies on every transport, including stdio.

When the filter runs you'll see one log line at startup, for example:

entitlement filter applied: entitled services=['zia', 'zpa'], kept 12 toolset(s), removed 17 toolset(s)

The filter is non-fatal. If credentials are missing, the token endpoint is unreachable, the token doesn't decode, or the token has no recognizable product entitlements, the server logs a single WARN line and starts normally with the user-selected toolsets unchanged.

To bypass the filter (for example, while diagnosing an unusual token shape):

zscaler-mcp --no-entitlement-filter
# or
export ZSCALER_MCP_DISABLE_ENTITLEMENT_FILTER=true

Only product entitlement is honoured — not role names. The server defers per-action permission enforcement to the live API; the entitlement filter only ensures we don't advertise tools for products the client has zero access to.

Cryptographic Confirmation for Destructive Actions

Delete operations are never executed on the first call. What happens instead depends on the client:

• Clients that support MCP elicitation get an interactive prompt naming the resource, answered by a human. The AI agent never handles the approval.
• Every other client gets a cryptographic confirmation token (HMAC-SHA256) it must pass back to proceed. The token is bound to the specific operation and its parameters, is single-use, and expires after 5 minutes — so it can't be forged, replayed, or reused for a different resource.

There is no way to turn this off. Deletes are irreversible against a live tenant, so the server ships no flag or environment variable that skips the gate. If you don't want an agent deleting anything, don't allowlist the delete tools: write tools are off by default and --write-tools takes explicit patterns (see Write Mode).

Network-Level Controls (HTTP only)

The next four subsections — TLS, source-IP allowlist, host-header validation, and the .env plaintext-secret scanner — apply only to the HTTP transports (sse, streamable-http). They control who can reach the server over the network. They are independent of the tool-level controls listed in Security Layers above (read-only mode, write allowlist, toolsets, entitlement filter, HMAC confirmations), which apply on every transport including stdio.

The corresponding MCP client authentication (Bearer / Basic / OAuth 2.1) is a fifth network-level layer covered in detail in the MCP Client Authentication section further below.

HTTPS/TLS Support

HTTPS is required by default for non-localhost deployments. The server will refuse to start on a non-localhost interface without TLS certificates unless you explicitly set ZSCALER_MCP_ALLOW_HTTP=true.

When running with HTTP transports (sse or streamable-http), provide TLS certificates:

ZSCALER_MCP_TLS_CERTFILE=/path/to/cert.pem
ZSCALER_MCP_TLS_KEYFILE=/path/to/key.pem

# Optional: private key password and CA bundle
ZSCALER_MCP_TLS_KEYFILE_PASSWORD=your-key-password
ZSCALER_MCP_TLS_CA_CERTS=/path/to/ca-bundle.pem

When TLS is configured, the server automatically starts with HTTPS. This works with both public (CA-signed) and private (self-signed) certificates. Generate a self-signed certificate for testing:

openssl req -x509 -newkey rsa:2048 -keyout key.pem -out cert.pem -days 365 -nodes \
  -subj "/CN=localhost"

Source IP Access Control

You can restrict which client IPs are allowed to connect using ZSCALER_MCP_ALLOWED_SOURCE_IPS. When unset (the default), source IP filtering is disabled and deferred to upstream controls (firewall rules, AWS Security Groups, etc.).

# Allow only specific IPs/subnets
ZSCALER_MCP_ALLOWED_SOURCE_IPS=10.0.0.0/8,172.16.0.5

# Allow all (effectively disable — same as not setting the variable)
ZSCALER_MCP_ALLOWED_SOURCE_IPS=0.0.0.0/0

Supports individual IPv4/IPv6 addresses, CIDR notation, and the wildcard 0.0.0.0/0. Health-check endpoints (/health, /healthz, /ready) are exempt so load-balancer probes continue to work. Requests from disallowed IPs receive 403 Forbidden.

.env File Security Warning

When starting with HTTP transports, the server automatically scans any .env file in the working directory for plaintext secrets (values containing SECRET, PASSWORD, KEY, or TOKEN). If detected, a security warning is logged recommending the use of a secrets manager or environment variables instead.

Security Posture Banner

On startup, the server logs a consolidated Security Posture Banner summarizing the active security configuration — transport mode, host validation status, authentication mode, TLS status, and any active warnings. This makes it easy to verify the security state at a glance.

Key Security Principles:

• No "enable all write tools" backdoor exists - allowlist is mandatory
• AI agents must request permission before executing any write operation (destructiveHint)
• Every destructive action requires explicit user approval through the AI agent's permission framework
• Destructive confirmations are cryptographically bound to prevent prompt injection bypass

Best Practices

• Read-Only by Default: No configuration needed for safe operations - read-only tools are always available
• Mandatory Allowlist: Always provide explicit --write-tools allowlist when enabling write mode
• Development/Testing: Use narrow allowlists (e.g., --write-tools "zpa_create_application_segment")
• Production/Agents: Keep server in read-only mode (default) for AI agents performing autonomous operations
• CI/CD: Never set ZSCALER_MCP_WRITE_ENABLED=true without a corresponding ZSCALER_MCP_WRITE_TOOLS allowlist
• Least Privilege: Use narrowest possible allowlist patterns for your use case
• Wildcard Usage: Use wildcards for service-level control (e.g., zpa_create_*) or operation-level control (e.g., *_create_*)
• Audit Review: Regularly review which write tools are allowlisted and remove unnecessary ones
• Specific Prompts: With 402 tools and deferred loading, AI agents match prompts to tools by relevance. Use service-specific prompts (e.g., "List ZPA segments" instead of "Show my segments") for accurate tool selection

🔐 MCP Client Authentication

📖 Full Documentation: Authentication & Deployment Guide

When running the MCP server over HTTP (sse or streamable-http transports), you can enable authentication to control who is allowed to connect to the server. This is independent from the Zscaler API credentials, which control how the server authenticates to Zscaler APIs.

For HTTP transports, the server auto-detects and enables authentication when auth-related environment variables are present. For stdio transport, authentication is not applicable (the operating system's process isolation provides security).

Authentication Modes

The server supports four authentication modes, configured via environment variables: