| Type | What it does |
|---|---|
| API Connection | Calls a single REST endpoint. Simpler to configure; best for standard CRUD operations on external systems. |
| MCP Connection | Connects to a Model Context Protocol server that exposes multiple tools. Better for complex, stateful workflows or when you need a pre-built skill. |
API connections
API connections let RAUM AI call any REST endpoint using credentials you provide. Common examples include order management, payment processing, shipping carriers, and CRM systems.Add an API connection
Enter the connection details
Fill in the following fields:
| Field | Description |
|---|---|
| Name | A descriptive label, e.g. Shopify — order lookup. The AI uses this name when deciding which tool to call. |
| Base URL | The full endpoint URL, e.g. https://your-store.myshopify.com/admin/api/2024-01/orders/{id}.json. |
| HTTP method | GET, POST, PUT, PATCH, or DELETE. |
| Auth type | Bearer token, API key header, Basic auth, or None. |
| Credentials | The API key or token. Credentials are encrypted at rest. |
| Description | Plain-language description of what this connection returns. The AI reads this to decide when to call it. |
| Example output | A representative JSON response. Helps the AI parse and use the data correctly. |
Configure caching and approval
Set a Cache TTL (in seconds) to avoid redundant API calls. Enable Safe to simulate if this endpoint is read-only and can be called without customer impact. Enable Auto-approve to let the AI call it without requiring agent approval.
Supported connection categories
E-commerce
Shopify, WooCommerce, Magento — order lookups, return status, inventory checks.
Payments
Stripe, Braintree — charge lookups, subscription status, refund eligibility.
Shipping
FedEx, UPS, USPS — tracking lookups, delivery estimates, carrier status.
CRM
Salesforce, HubSpot — contact records, account health, deal status.
MCP connections
MCP connections use the Model Context Protocol to connect RAUM AI to a server that exposes one or more tools. MCP servers communicate over SSE (Server-Sent Events), making them suitable for real-time data retrieval.Built-in skill types
When you add an MCP connection, you assign it a skill type. The type helps RAUM AI understand the category of data the server provides:| Skill type | Description |
|---|---|
WEATHER | Real-time weather conditions and forecasts |
PRODUCT_AVAILABILITY | Stock levels, availability, and inventory status |
SERVICE_STATUS | System health, uptime, and incident status |
WEB_SCRAPER | Live content extraction from web pages |
CUSTOM | Any other tool — describe its purpose in the connection metadata |
Add an MCP connection
Open Connections
Click Connections in the dashboard sidebar, then click New connection and select MCP Connection.
Enter the server URL
Paste the SSE endpoint URL of your MCP server. The URL must use HTTPS in production.
Select the skill type
Choose the skill type that best describes the server’s capabilities. For servers that don’t fit a standard category, choose
CUSTOM and write a clear description.Add credentials (if required)
If your MCP server requires authentication, enter the API key or headers. Credentials are encrypted with AES-256-GCM before storage.
Configure per-tool cache TTL
After the server connects, RAUM AI retrieves the list of tools it exposes. For each tool, set a Cache TTL in seconds. A higher TTL reduces API calls for data that doesn’t change frequently (e.g., product catalog). Use
0 for data that must always be fresh (e.g., live order status).Mark tools as safe
Tools marked as Safe are simulation-safe — they can run without agent approval because they are read-only and have no side effects. Mark any tool that only retrieves data and never modifies it.
Configure auto-approval
Tools with Auto-approve enabled run without requiring an agent to review the action in Copilot Mode. Enable this for low-risk lookups where human review would slow down resolution.
How RAUM AI uses integrations
During the Act step of the pipeline, the Tool Executor agent calls the relevant connections based on what the Planner Agent has determined it needs. Here is how a typical ticket flow works with integrations:Thinker identifies data requirements
The Thinker Agent analyzes the ticket and determines which external data sources must be queried before a resolution is possible — for example, “This is a delivery inquiry, so order status and tracking must be fetched.”
Tool Executor calls the connections
The Tool Executor calls your configured connections (or MCP tools) in the correct sequence. Results are cached using the TTL you configured to avoid redundant calls on subsequent tickets.
Planner decides based on live data
The Planner Agent combines the ticket context, your procedures, and the real-time data returned by the tools to determine the correct resolution: approve a refund, send a tracking link, escalate to a specialist, or another action.
Example: order lookup
A customer submits a ticket: “My order #12345 was supposed to arrive yesterday. Where is it?”- The Thinker Agent identifies this as a delivery inquiry and flags the order lookup and shipping tracker connections as required.
- The Tool Executor calls your Shopify — order lookup connection with order ID
12345. - The Tool Executor calls your FedEx tracking connection with the tracking number from the order.
- FedEx returns a status: “Out for delivery — estimated arrival today by 8pm.”
- The Planner Agent matches the delay procedure and determines no escalation is needed.
- The Answerer Agent drafts a reply citing the tracking status and estimated delivery time.