Core concepts

Copy

• Copy for LLM

  Copy page as Markdown for LLMs
• View as Markdown

  Open this page as Markdown
• Open in ChatGPT

  Get insights from ChatGPT
• Open in Claude

  Get insights from Claude
• Connect to Cursor

  Install MCP server on Cursor
• Connect to VS Code

  Install MCP server on VS Code

Agen for SaaS provides the infrastructure for SaaS products to safely expose their APIs and capabilities to AI platforms and agents. It turns your product into an AI-native platform, allowing customers to interact with it through AI assistants like ChatGPT, Claude, Gemini, or custom MCP agents — while maintaining security, governance, and compliance.

This page explains the core concepts that make Agen for SaaS work.

MCP Gateway

The MCP Gateway is the managed entry point that connects AI agents to your product's APIs through the Model Context Protocol (MCP).

• Hosted and orchestrated by Agen for SaaS at a unique URL (e.g., your-id.mcp-gw.frontegg.com).
• Exposes your APIs as Tools that AI agents can discover and invoke.
• Enforces authentication, access control, policies, data protection, and hooks on every request.
• Provides a Test MCP Server button to verify your gateway is operational at any time.

The gateway URL is auto-generated when you create your server and can be copied directly from the control plane or the Configuration tab.

Sources

A Source is a connection to an external API or service that provides the tools your AI agents can use. Agen for SaaS supports three source types:

Each source has a name, an API base URL, connected tools, and an enabled/disabled toggle. When you add a REST or GraphQL source, click Generate tools to automatically create tools from your API specification.

Tools

A Tool represents a single API operation that an AI agent can invoke through Agen for SaaS.

Each tool has:

• Tool name — A unique identifier (e.g., Create_a_new_expense, Get_receipt_by_ID).
• Tool description — A human-readable explanation of what the tool does (up to 2,000 characters). This description is what AI agents use to understand the tool's purpose.
• Tool API — The HTTP method (GET, POST, PUT, DELETE) and path (e.g., /api/receipts/{id}).
• Source name — Which source the tool belongs to (e.g., default, Frontegg).
• Enable toggle — Tools can be individually enabled or disabled without deletion.

Tools are auto-generated when you add a source, but you can edit any tool's name, description, method, and path after creation. You can also delete tools that are no longer needed.

Authentication

Agen for SaaS authenticates every request that passes through the MCP Gateway. Two authentication providers are supported:

You can optionally configure a custom domain for your authentication endpoint through a two-step wizard (add domain, then configure DNS records).

Access control (RBAC / ABAC)

Access control defines who can invoke which tools by mapping tools to JWT attributes.

Each access control rule (called a mapping) consists of:

• Mapping name — A descriptive label (e.g., "Admin only: Delete_expense").
• Tools — One or more tools this rule applies to.
• Mapping type — Either Map to roles (RBAC) or Map to permissions (ABAC). This choice is permanent and cannot be changed after creation.
  • Map to roles — Restricts tool access to users with specific roles (e.g., "Read Only", "Admin").
  • Map to permissions — Restricts tool access to users with specific permissions (e.g., "Read event categories").

Access control rules are evaluated against the JWT claims in every incoming request. If a user lacks the required role or permission, the tool call is blocked.

Policies and guardrails

Policies define conditional rules that control what happens when an AI agent calls a tool. Each policy evaluates request attributes and triggers an action.

A policy consists of:

• Name and description — Identify the policy's purpose.
• Policy tools — Which tools the policy applies to.
• Policy attributes — Conditional expressions using the format: IF [Attribute] [Operator] [Value] (e.g., IF IP Contains 192.168). Multiple conditions can be combined with + and.
• Policy action — The enforcement action when conditions match:

Policies can be toggled on/off, filtered by action type, and targeted at specific tools. Each policy shows its type (Global or tool-specific), action, and targeting expression in the list view.

Approval flows

When a policy triggers a Request approval action, an Approval flow defines who reviews and approves the request.

Creating an approval flow is a four-step process:

1. General settings — Set the flow name, description, and activation status.
2. Channel configuration — Enable notification channels:
   • Email — Send approval requests to approvers' email addresses.
   • SMS — Send short approval requests as text messages.
3. Approval management — Define the approval sequence using a visual flow builder:
   • Add approval steps with approvers defined by role or email address.
   • Set the required minimum number of approvers for each step.
   • Steps are executed sequentially (Start flow → Step 1 → Step 2 → ... → End flow).
4. Advanced settings — Fine-tune behavior:
   • Auto-approve timeout — Automatically approve requests that remain pending beyond a specified time.
   • Send reminder — Nudge approvers if a request stays pending.
   • Notify requester on decision — Inform the requester when their request is approved or rejected.
   • CC traditional recipients — Send copies of all approval decisions to additional stakeholders.

Data protection

Data protection policies mask sensitive information in tool responses to maintain compliance with regulations like GDPR, HIPAA, and CCPA.

Each data protection policy defines:

• Policy name and description — Identify the policy's purpose.
• Data types to protect — Select from categorized sensitive data types:
  • PHI — Protected Health Information (39 types including Medicare numbers, health IDs, insurance numbers).
  • PII — Personally Identifiable Information.
  • GDPR — EU data protection categories.
  • Additional compliance categories (PCI, CCPA, COPPA).
• Policy tools — Which tools this masking applies to.
• Policy targeting — Conditional expressions that determine when masking is applied (e.g., IF Country In United States).

Data protection policies have a status toggle and can be activated or deactivated at any time.

Hooks

Hooks allow you to execute custom JavaScript code when specific MCP events occur. There are two hook types:

Each hook includes:

• Body — A JavaScript code editor with a default template. The onEvent(eventData) function receives the event details and returns a verdict ('allow' or 'deny') along with an optional response object.
• Settings — Configure the fail methodology (e.g., "Fail Open" to allow requests when the hook fails) and timeout (default: 5 seconds).
• Test button — Validate your hook logic before activation.
• Activate/Inactivate toggle — Control whether the hook is active.

Hooks enable advanced use cases like custom validation, request transformation, dynamic tool filtering, and external service integration.

Configuration

The Configuration tab manages your MCP server's basic settings:

• Basic configuration:

  • Server name — The display name for your MCP server (e.g., "Expense Management App").
  • MCP Gateway URL — The auto-generated endpoint (e.g., your-id.mcp-gw.frontegg.com). Read-only, with a copy button.
  • Custom domain — Optionally map a custom domain to your MCP Gateway through a two-step DNS configuration wizard.

• Allowed origins:

  • Controls which origins can access your Agen for SaaS app, enforced by the Origin header.
  • Pre-populated with {{APP_URL}} and {{LOGIN_URL}}.
  • Add additional origins as needed. Changes take up to 5 minutes to take effect.

Monitoring and auditing

All actions within Agen for SaaS are logged for transparency and compliance. The monitoring system records:

• Which user or AI agent invoked which tool.
• When the request occurred.
• Which policies, access control rules, or approvals were triggered.
• The full request and response payloads.

Logs can be viewed in the dashboard, streamed to third-party platforms (Datadog, Splunk, AWS EventBridge), and filtered by time, tool, user, or action.

How the layers interact

Related topics

• Overview
• Quickstart
• Sources
• Tools
• Policies
• Hooks
• Data protection