RBAC Configuration

RBAC Configuration

Role-based access control (RBAC) defines which actions users or teams can perform in MCP Gateway. This document covers the two-layer security model, token scoping semantics, permission system, and best practices for access control.

Overview

MCP Gateway implements a two-layer security model:

1. Token Scoping (Layer 1): Controls what resources a user CAN SEE (data filtering)
2. RBAC (Layer 2): Controls what actions a user CAN DO (action authorization)

Both layers must pass for an operation to succeed.

┌─────────────────────────────────────────────────────────────────────────────┐
│                      Two-Layer Security Model                               │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   Request → Authentication → Token Scoping → RBAC Check → Operation        │
│                              (Can See?)       (Can Do?)                     │
│                                                                             │
│   ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐ │
│   │   JWT    │   │  User    │   │ Resource │   │Permission│   │ Execute  │ │
│   │  Token   │──▶│ Identity │──▶│  Access  │──▶│  Check   │──▶│ Operation│ │
│   │          │   │          │   │          │   │          │   │          │ │
│   └──────────┘   └──────────┘   └──────────┘   └──────────┘   └──────────┘ │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

Authentication Methods

Core Concepts

Subjects

Users authenticated via:

• JWT tokens (session or API)
• SSO providers (OAuth 2.0/OIDC)
• Basic authentication (development only)

Teams

Logical groups that:

• Organize users for access boundaries
• Own resources (tools, prompts, resources)
• Map from external identity providers (SSO groups)

Built-in RBAC Roles

!!! info "Default Role Assignment" New users automatically receive up to two roles upon creation:

**Admin users** (`is_admin: true`) receive:

1. `platform_admin` role with **global scope** (`scope_id` = None)
   - Grants unrestricted access to all platform resources
2. `team_admin` role with **team scope** (`scope_id` = personal team ID)
   - Grants full management of their personal team resources
   - Only assigned if personal team creation succeeds

**Non-admin users** (`is_admin: false`) receive:

1. `platform_viewer` role with **global scope** (`scope_id` = None)
   - Grants read-only access to all platform resources
2. `team_admin` role with **team scope** (`scope_id` = personal team ID)
   - Grants full management of their personal team resources
   - Only assigned if personal team creation succeeds

This dual-role approach ensures:
- Users always have appropriate global visibility (via `platform_admin` or `platform_viewer`)
- Users can fully manage their personal team resources (via team-scoped `team_admin`, when available)
- Clear separation between team-level and platform-level permissions

The `granted_by` field tracks which admin created the user for audit purposes.

!!! note "Existing Users Migration" An Alembic migration (v1a2b3c4d5e6) automatically updates existing users without roles:

**Previous behavior (before migration):**
- Admin users: Only `platform_admin` with global scope (`scope_id` = None)
- Non-admin users: Only `viewer` with team scope (`scope_id` = None)

**After migration:**

**Admin users** receive:

1. `team_admin` role with **team scope**
   - `scope_id` = user's personal team ID (from `email_team_members` table)
   - Enables management of personal team resources
2. `platform_admin` role with **global scope**
   - `scope_id` = None
   - Maintains unrestricted platform access

**Non-admin users** receive:

1. `team_admin` role with **team scope**
   - `scope_id` = user's personal team ID (from `email_team_members` table)
   - Enables management of personal team resources
2. `platform_viewer` role with **global scope**
   - `scope_id` = None
   - Provides read-only access to platform resources

**Migration behavior:**
- Only affects users without existing role assignments
- Users without a personal team still receive their global role (team_admin is skipped)
- The platform admin (configured via `PLATFORM_ADMIN_EMAIL`) is excluded to preserve bootstrap configuration
- Migration is idempotent and safe to run multiple times

Resources

Protected entities:

• Servers (MCP gateways and virtual servers)
• Tools, Prompts, Resources (MCP primitives)
• System configuration and audit logs

Token Scoping Model

Token scoping controls what resources a token can access based on the teams claim in the JWT payload. The normalize_token_teams() function is the single source of truth for interpreting JWT team claims across all enforcement points.

Token Scoping Contract

The teams claim in JWT tokens determines resource visibility. The system follows a secure-first design: when in doubt, access is denied.

┌─────────────────────────────────────────────────────────────────────────────┐
│                        Token Teams Claim Handling                           │
├─────────────────────────────────────────────────────────────────────────────┤
│  JWT Claim State          │  is_admin: true       │  is_admin: false        │
├───────────────────────────┼───────────────────────┼─────────────────────────┤
│  No "teams" key           │  PUBLIC-ONLY []       │  PUBLIC-ONLY []         │
│  teams: null              │  ADMIN BYPASS (None)  │  PUBLIC-ONLY []         │
│  teams: []                │  PUBLIC-ONLY []       │  PUBLIC-ONLY []         │
│  teams: ["team-id"]       │  Team + Public        │  Team + Public          │
│  teams: ["t1", "t2"]      │  Both Teams + Public  │  Both Teams + Public    │
└───────────────────────────┴───────────────────────┴─────────────────────────┘

!!! warning "Admin Bypass Requirements" Admin bypass (unrestricted access) requires BOTH conditions:

1. `teams: null` (explicit null, not missing key)
2. `is_admin: true`

A missing `teams` key always results in public-only access, even for admins.
An empty `teams: []` also results in public-only access, even for admins.

Return Value Semantics

Security Design Principles

1. Secure-First Defaults

   • Missing teams key always returns [] (public-only access)
   • This prevents accidental exposure when tokens are misconfigured

2. Explicit Admin Bypass

   • Admin bypass requires explicit teams: null AND is_admin: true
   • Empty teams [] disables bypass even for admins

3. Scoped Automation Tokens

   • Tokens with teams: [] are intentionally restricted to public resources
   • Use case: CI/CD pipelines, monitoring systems, public API clients

Token Scoping Flow

                                 ┌──────────────────┐
                                 │   JWT Token      │
                                 │   Received       │
                                 └────────┬─────────┘
                                          │
                                          ▼
                              ┌───────────────────────┐
                              │  Extract "teams"      │
                              │  claim from JWT       │
                              └───────────┬───────────┘
                                          │
                          ┌───────────────┴───────────────┐
                          │                               │
                          ▼                               ▼
               ┌─────────────────────┐       ┌─────────────────────┐
               │ "teams" key EXISTS  │       │ "teams" key MISSING │
               └──────────┬──────────┘       └──────────┬──────────┘
                          │                             │
                          ▼                             ▼
               ┌─────────────────────┐       ┌─────────────────────┐
               │ Check teams value   │       │ Return []           │
               └──────────┬──────────┘       │ PUBLIC-ONLY         │
                          │                  │ (secure default)    │
          ┌───────────────┼───────────────┐  └─────────────────────┘
          │               │               │
          ▼               ▼               ▼
  ┌───────────────┐ ┌───────────┐ ┌───────────────┐
  │ teams: null   │ │ teams: [] │ │ teams: [...]  │
  └───────┬───────┘ └─────┬─────┘ └───────┬───────┘
          │               │               │
          ▼               │               ▼
  ┌───────────────┐       │       ┌───────────────┐
  │ Check is_admin│       │       │ Return [...]  │
  └───────┬───────┘       │       │ TEAM-SCOPED   │
          │               │       └───────────────┘
    ┌─────┴─────┐         │
    │           │         │
    ▼           ▼         ▼
┌────────┐ ┌────────┐ ┌────────┐
│ Admin  │ │Non-Adm │ │ Empty  │
│ true   │ │ false  │ │ list   │
├────────┤ ├────────┤ ├────────┤
│ Return │ │ Return │ │ Return │
│ None   │ │ []     │ │ []     │
│ BYPASS │ │ PUBLIC │ │ PUBLIC │
└────────┘ └────────┘ └────────┘

!!! note "Key Insight" The difference between teams: null and missing teams key is critical:

- **Missing key**: Always `[]` (public-only) - secure default
- **Explicit null**: Admin bypass when `is_admin: true`, otherwise `[]`

Visibility Levels

Resources in MCP Gateway have three visibility levels:

Access Matrix by Token Type

!!! warning "Public-Only Token Limitations" Public-only tokens (teams=[]) cannot access private resources, even if the resource is owned by the token's user.

This is intentional security behavior - public-only tokens are designed for limited-scope access to public resources only. To access private resources, users must use a team-scoped token that includes their personal team.

```
# Public-only token behavior:
# ✅ Can access: visibility='public' resources
# ❌ Cannot access: visibility='team' resources (any team)
# ❌ Cannot access: visibility='private' resources (even if owned by user)
```

Enforcement Points

Token scoping is enforced consistently across all access paths:

Token Types and Use Cases

Session Tokens (UI Login)

Generated when users log in via the Admin UI:

{
  "sub": "admin@example.com",
  "is_admin": true,
  "teams": null,
  "iss": "mcpgateway",
  "aud": "mcpgateway-api",
  "exp": 1234567890
}

Behavior: Admin session tokens should set teams: null (explicit null) combined with is_admin: true to enable admin bypass (unrestricted access to all resources).

API Tokens (Programmatic Access)

Generated via the Admin UI or API for automation:

{
  "sub": "service-account@example.com",
  "is_admin": false,
  "teams": ["team-uuid-1", "team-uuid-2"],
  "iss": "mcpgateway",
  "aud": "mcpgateway-api",
  "exp": 1234567890
}

Behavior: Access restricted to public resources plus resources owned by specified teams.

Scoped Automation Tokens

For CI/CD, monitoring, or public API access:

{
  "sub": "ci-pipeline@example.com",
  "is_admin": true,
  "teams": [],  // Explicitly empty = public-only
  "iss": "mcpgateway",
  "aud": "mcpgateway-api",
  "exp": 1234567890
}

Behavior: Even admin tokens with teams: [] are restricted to public resources only. This enables creating limited-scope tokens for automation that shouldn't access team-internal resources.

Generating Scoped Tokens

Using the CLI Tool

# Unrestricted admin token (no teams key)
python3 -m mcpgateway.utils.create_jwt_token \
  --username admin@example.com \
  --exp 60 \
  --secret $JWT_SECRET_KEY

# Team-scoped token
python3 -m mcpgateway.utils.create_jwt_token \
  --username user@example.com \
  --exp 60 \
  --secret $JWT_SECRET_KEY \
  --teams '["team-uuid-1"]'

# Public-only scoped token (for automation)
python3 -m mcpgateway.utils.create_jwt_token \
  --username ci@example.com \
  --exp 60 \
  --secret $JWT_SECRET_KEY \
  --teams '[]'

Using the Admin UI

1. Navigate to Admin UI → Tokens

2. Click Create Token

3. Select team scope:

   • No team selected: Public resources only (secure default)
   • Specific team(s): Team + public resources

4. Configure additional restrictions (IP, permissions, expiry)

!!! warning "Token Scope Warning" Tokens created without selecting a team will have access to public resources only. This is the secure default to prevent accidental exposure of team resources.

Permission System

Permission Categories

Permissions are defined in the Permissions class and control what actions users can perform:

Permission Checking Flow

@require_permission("resource.action")
    │
    ▼
┌─────────────────────────────┐
│ Extract user_context        │ ← From request/kwargs
└─────────────────────────────┘
    │
    ▼
┌─────────────────────────────┐
│ Plugin Permission Hook      │ ← HTTP_AUTH_CHECK_PERMISSION can override
└─────────────────────────────┘
    │ (no plugin decision)
    ▼
┌─────────────────────────────┐
│ Admin Bypass Check          │ ← If allow_admin_bypass=True AND user.is_admin
└─────────────────────────────┘
    │ (not admin or bypass disabled)
    ▼
┌─────────────────────────────┐
│ Role Collection             │ ← Get all active roles for user
│ - Global scope roles        │
│ - Personal scope roles      │
│ - Team scope roles          │
└─────────────────────────────┘
    │
    ▼
┌─────────────────────────────┐
│ Permission Aggregation      │ ← Collect permissions from roles
│ - Include inherited perms   │   (role inheritance supported)
│ - Check for wildcard (*)    │
└─────────────────────────────┘
    │
    ▼
┌─────────────────────────────┐
│ Fallback Permission Check   │ ← Implicit permissions (see below)
└─────────────────────────────┘
    │
    ▼
  GRANT or DENY

Fallback Permissions

The system grants implicit permissions without explicit role assignment. These are not shown in /rbac/my/permissions but are effective:

!!! info "Why Fallback Permissions Exist" Fallback permissions enable basic functionality without requiring explicit role assignment:

- Users can always create and view teams they belong to
- Team owners automatically have management rights
- Users can always manage their own API tokens

Admin API RBAC

The Admin API enforces strict RBAC where even users with is_admin: true must have explicit permissions granted. This enables delegated administration - granting specific admin capabilities without full superuser access.

Key behaviors:

Example: Delegated Server Management

{
  "role": "server-manager",
  "permissions": [
    "servers.read",
    "servers.create",
    "servers.update",
    "servers.delete"
  ]
}

A user with this role can:

• ✅ Access /admin/servers/* endpoints
• ✅ View the Admin UI (has servers.* which satisfies has_admin_permission())
• ❌ Access /admin/tools/* endpoints (no tools.* permissions)
• ❌ Access /admin/gateways/* endpoints (no gateways.* permissions)

!!! warning "Platform Admin Role" The built-in platform_admin role has ["*"] (wildcard) permissions, which grants access to all operations. For delegated administration, create custom roles with specific permission sets.

Configuration Safety

Development vs Production Settings

The following configuration combinations require careful consideration:

Proxy Authentication Mode

When MCP_CLIENT_AUTH_ENABLED=false and TRUST_PROXY_AUTH=true:

• The gateway trusts the X-Forwarded-User header from upstream proxy
• No JWT validation or database verification is performed
• Only use when deployed behind a trusted reverse proxy that handles authentication

!!! danger "Security Warning" Proxy authentication mode should only be used in trusted network environments where the reverse proxy is the only entry point to the gateway. Exposing the gateway directly to untrusted networks with this configuration allows header injection attacks.

Anonymous Mode (AUTH_REQUIRED=false)

When AUTH_REQUIRED=false:

• All unauthenticated requests receive platform-admin context
• Never use in production - all users have full admin access
• Intended only for local development and testing

!!! danger "Production Warning" Setting AUTH_REQUIRED=false in production grants administrative access to all requests. This completely bypasses authentication and authorization.

Best Practices

Token Lifecycle

1. Use short expiration times for interactive sessions (hours)
2. Use longer expiration for service accounts with IP restrictions
3. Rotate tokens regularly (recommended: 90 days for long-lived tokens)
4. Revoke tokens immediately when access should be removed

Team Organization

1. Create purpose-specific teams:

   • platform-admins - Full administrative access
   • developers - Development and testing resources
   • ci-automation - CI/CD pipeline access
   • monitoring - Read-only observability access

2. Map SSO groups to teams for automatic membership management

3. Use personal teams for individual resource ownership

Scoping Strategy

Troubleshooting

Token Not Seeing Expected Resources

1. Check token claims: Decode the JWT to verify teams claim

   # Decode JWT payload (middle section)
   echo "$TOKEN" | cut -d. -f2 | base64 -d | jq .
   

2. Verify resource visibility: Check the resource's visibility and team_id

   curl -H "Authorization: Bearer $ADMIN_TOKEN" /tools/{id} | jq '{visibility, teamId}'
   

3. Check user admin status: Non-admin users without teams get public-only access

Admin Token Being Restricted

If an admin token is unexpectedly restricted:

1. Check for explicit teams claim: teams: [] restricts even admins
2. Verify is_admin flag: Must be true in JWT or database user
3. Check middleware logs: Look for "token_teams" in debug output

Inconsistent Results Between Endpoints

If REST and RPC endpoints return different results:

1. Check for caching: REST list endpoints may have cached data
2. Wait for cache TTL: Default is 60 seconds for registry cache
3. Use direct GET: /tools/{id} bypasses list cache

Bootstrap Custom Roles

MCP Gateway allows you to define custom roles that are automatically created during database bootstrap. This is useful for organizations that need to pre-configure roles before deployment.

Configuration

Enable custom role bootstrapping with these environment variables:

Role Definition Format

Create a JSON file containing an array of role definitions:

[
  {
    "name": "data_analyst",
    "description": "Read-only access for data analysis",
    "scope": "team",
    "permissions": ["tools.read", "resources.read", "prompts.read"],
    "is_system_role": true
  },
  {
    "name": "auditor",
    "description": "Compliance audit access",
    "scope": "global",
    "permissions": ["tools.read", "resources.read", "prompts.read", "servers.read", "gateways.read"],
    "is_system_role": true
  }
]

Required fields:

• name - Unique role name
• scope - Either team (team-level access) or global (system-wide access)
• permissions - Array of permission strings (e.g., tools.read, resources.create)

Optional fields:

• description - Human-readable description
• is_system_role - Set to true to prevent users from modifying/deleting the role

Available Permissions

Docker Compose Example

services:
  gateway:
    environment:
      - MCPGATEWAY_BOOTSTRAP_ROLES_IN_DB_ENABLED=true
      - MCPGATEWAY_BOOTSTRAP_ROLES_IN_DB_FILE=/app/custom_roles.json
    volumes:
      - ./custom_roles.json:/app/custom_roles.json:ro

Kubernetes/Helm Example

# values.yaml
mcpContextForge:
  env:
    MCPGATEWAY_BOOTSTRAP_ROLES_IN_DB_ENABLED: "true"
    MCPGATEWAY_BOOTSTRAP_ROLES_IN_DB_FILE: "/config/custom_roles.json"

  # Mount ConfigMap with role definitions
  extraVolumes:
    - name: custom-roles
      configMap:
        name: mcp-gateway-roles
  extraVolumeMounts:
    - name: custom-roles
      mountPath: /config

Error Handling

• File not found: Bootstrap continues with default roles only; warning logged
• Invalid JSON: Bootstrap continues with default roles only; error logged
• Malformed entries: Invalid role entries are skipped with warnings; valid entries are processed

!!! tip "Idempotent Bootstrap" Bootstrap is idempotent - running it multiple times won't duplicate roles. Existing roles are detected and skipped.

Related Documentation

• Team Management - Setting up teams and SSO mapping
• Security Features - Comprehensive security configuration
• Configuration Reference - Environment variables
• API Usage - Token usage in API calls