Skip to main content

MCP Integration Guide

Overview

The Model Context Protocol (MCP) integration allows AMAS agents to connect to external MCP servers and use their tools within the AI workflows. This guide walks you through the complete setup process.

Table of Contents

  1. Overview
  2. Prerequisites
  3. Step 1: Create an Integration
  4. Step 2: Configure MCP Server
  5. Step 3: Create a Connection
  6. Step 4: Retrieve Tools
  7. Human Approval for MCP Tools
  8. Step 5: Use MCP Tools in YAML
  9. Provider-Specific Setup
  10. Troubleshooting
  11. Advanced Configuration

Prerequisites

Before setting up MCP integration, ensure you have:
  • Access credentials for your MCP provider (OAuth credentials or API key)
  • The MCP server URL (base URL where the MCP server is running)

Step 1: Create an Integration

An Integration is a company-level configuration that stores the authentication credentials for connecting to an external service.

For OAuth2 Providers (e.g., Microsoft, Atlassian)

  1. Navigate to Settings → Integrations
  2. Click “Add Integration”
  3. Fill in the form:
    • Provider: Select your provider (e.g., Microsoft, Atlassian)
    • Name: Give your integration a descriptive name (e.g., “Company Atlassian”)
    • Description: Optional description
    • OAuth Client ID: Your OAuth application’s client ID
    • OAuth Client Secret: Your OAuth application’s client secret
    • Scopes: Required OAuth scopes (comma-separated), e.g:
      • For Atlassian: read:jira-work, write:jira-work
      • For Microsoft: Mail.Read, Mail.Send, offline_access
  4. Click “Create Integration”
  5. Set the integration as default one (at least one default integration is needed for a provider)

For API Key Providers (e.g., Lexware)

  1. Navigate to Settings → Integrations
  2. Click “Add Integration”
  3. Fill in the form:
    • Provider: Select your API key provider (e.g., Lexware)
    • Name: Give your integration a descriptive name (e.g., “Company Lexware”)
    • Description: Optional description
  4. Click “Create Integration”
  5. Set the integration as default one (at least one default integration is needed for a provider)
Note: For API key providers, you don’t need to provide credentials at the integration level. The API key will be provided when creating a connection.

Step 2: Configure MCP Server

An MCP Configuration defines the MCP server details and links it to your integration.
  1. Navigate to Tools → MCP Servers
  2. Click “Add MCP”
  3. Fill in the configuration:
    • Name: Descriptive name for this MCP server (e.g., “Atlassian Jira MCP”)
    • Base URL: The MCP server endpoint
      • Example: http://localhost:8008/mcp (for local development)
      • Example: https://mcp-atlassian.yourcompany.com/mcp (for production)
    • Authentication Type: Select the auth type
      • No Authentication: For public MCP servers
      • OAuth 2.0: For OAuth-based providers
      • API Key: For API key-based providers
    • Provider: Select the provider which is linked to your created integration in Step 1
  4. Click “Create MCP Server”

Step 3: Create a Connection

A Connection is a user-level link between your account and the external service. This is where users authenticate.

For OAuth2 Providers

  1. Navigate to the MCP server configured in step 2
  2. Click “Connect” or View Details -> “Add Connection”
  3. Give the connection a unique name not used in other MCPs as well (set it as default, at least 1 default connection is needed)
  4. You’ll be redirected to the provider’s OAuth consent screen
  5. Grant the requested permissions
  6. You’ll be redirected back to Agent Studio with an active connection

For API Key Providers

  1. Navigate to the MCP server configured in step 2
  2. Click “Connect” or View Details -> “Add Connection”
  3. Give the connection a unique name not used in other MCPs as well (set it as default, at least 1 default connection is needed)
  4. You’ll be redirected to the provider’s consent screen
    • API Key: Enter your API key from the provider
  5. Click “Connect”
  6. You’ll be redirected back to Agent Studio with an active connection
Important: Each user needs to create their own connection to use MCP tools. The integration is shared at the company level, but connections are per-user.

Step 4: Retrieve Tools

Once your MCP server is configured, and you have an active connection, you can fetch/refresh tools. Afterward you can remove tools which you don’t want in the configured MCP.
In the end you click *Save Tools otherwise the MCP will be empty or with the last configured tools.

Human Approval for MCP Tools

MCP tools come from a server whose code you do not own, so there is no source-level declaration of which ones are dangerous. AMAS therefore maintains a per-server approval set: every tool in it pauses for human approval before it runs.

How the set is seeded

When you fetch or refresh the manifest, AMAS classifies each newly-discovered tool:
  • A tool the server explicitly marks read-only is left ungated.
  • Every other tool — writes, and anything with no hint at all — is added to the approval set.
The default is deliberately cautious: an unannotated tool is treated as if it writes.

How your choices survive a refresh

On a later refresh:
  • Tools you removed from the set stay removed. A refresh never re-gates them.
  • Tools you added stay added, as long as they still exist in the manifest.
  • Tools that disappeared from the manifest drop out of the set.
  • Only genuinely new non-read-only tools are gated automatically.
Removing a tool from the approval set means an agent can call it with no human in the loop. Only do this for tools you have confirmed are safe to run unattended against this server.
You can only gate tools that exist in the server’s current manifest — a name that is not in it is rejected rather than silently stored.

Relationship to YAML HITL

The approval set acts as a policy mandate, the second-strongest layer of the Human-in-the-Loop precedence ladder. An agent author cannot switch it off per instance: disable_human_approval: true is ignored against it. To gate an MCP tool from YAML instead — for example a tool referenced directly rather than through a configured server — use mandate_human_approval on the tool entry:

Human-in-the-Loop

The full precedence ladder, every hitl_config option, and how approvals are answered

Step 5: Use MCP Tools in YAML

Once your MCP server is configured, and you have an active connection, you can use MCP tools in your YAML workflows.

Tool Naming Convention

MCP servers/tools are prefixed with mcp__ followed by the MCP configuration name:

Example YAML Configuration


Provider-Specific Setup

Atlassian (Jira/Confluence)

Authentication: OAuth 2.0 Required Scopes: For example
  • read:jira-work - Read Jira issues
  • write:jira-work - Create/update Jira issues
Note: You can have a look at the client you have configured at the atlassian admin for more scopes
Special Requirements:
  • The system automatically fetches your Atlassian Cloud ID
  • If you have multiple Atlassian sites, you’ll need to configure the Cloud ID manually (Still under development)
Example MCP Server URL: http://localhost:8008/mcp Available Tools:
  • list_issues - List Jira issues with filters
  • create_issue - Create new Jira issues
  • update_issue - Update existing issues
  • get_issue - Get details of a specific issue

Lexware Office

Authentication: API Key API Key Location: Get your API key from Lexware Office settings Special Requirements:
  • Uses custom header Lexware-API-Key for authentication
  • No OAuth setup required
Example MCP Server URL: http://localhost:8009/mcp Available Tools:
  • get_invoices - List invoices with filters
  • get_invoice - Get invoice details by ID
  • get_contacts - List contacts
  • get_posting_categories - Get posting categories
  • get_countries - Get country tax classifications

Troubleshooting

”No connection found” Error

Problem: MCP tools fail with “No connection found” Solution:
  1. Ensure you’ve created a connection (Step 3)
  2. Check that the connection is set as default (if you have multiple)

“Multiple Atlassian sites found” Error

Problem: Atlassian integration fails with multiple sites error Solution:
  1. This occurs when your Atlassian account has access to multiple sites
  2. You need to manually configure the Cloud ID
  3. Contact your administrator to set the specific Cloud ID for your organization

Advanced Configuration

Multiple MCP Servers for Same Provider

You can configure multiple MCP servers for the same provider (e.g., different Jira projects):

Custom MCP Servers

If you’re running your own MCP server with some custom provider for authentication please communicate to the administrator to set up the MCP.

Last Updated: March 2026 Version: 1.0