Overview

As soon as the AI needs a connector in a conversation that your account isn't connected to yet, you can connect it right in the chat – no detour through the settings.

The AI shows a prompt with a Connect button (e.g. "Connect to Gmail", "Connect to Google Calendar", Asana, Confluence).

You authorize your account via OAuth in the window that opens.

After connecting, the AI continues the task automatically – without leaving the chat.

Connected services remain centrally manageable under Settings → Connections (disconnect, reconnect).

Already-connected connectors can be called up directly in the chat input by typing @ — same pattern as assistants and data sources.

Type @ in the chat input. A picker opens showing your connected connectors (e.g. Google Calendar, Microsoft Teams, custom MCP servers).

Select the connector you want. It is inserted as an @-mention in the message and activated for this chat.

Active connectors appear as chips below the input field — click the × on a chip to remove the connector again.

Note

The connector must already be connected under Settings → Connections or via the "+" menu → Connectors in the chat. Otherwise it won't appear in the @ picker. Per-function permissions (e.g. "read-only" on Outlook or Teams) are honored whether the connector is activated via the @ menu or the "+" menu.

In the Assistant Editor

Open the Assistant Editor

Navigate to the "Tools" section

Select desired connectors

Configure required parameters

Save the assistant

Tool Configuration

tools:
  - type: mcp
    name: database
    config:
      connection_string: ${DATABASE_URL}
      allowed_operations: ["SELECT"]
 
  - type: mcp
    name: web
    config:
      max_pages: 5
      timeout: 30

Custom MCP Servers

You can also integrate your own MCP servers:

Requirements

Publicly accessible URL
MCP protocol implementation
HTTPS encryption
Authentication (optional)

Setup

Create the server

Create it via the "Add a custom MCP server" tile (name + URL)

Add credentials

OAuth (if the server supports it) or an API key/token as a header

Set permissions

In the assistant editor

Details incl. troubleshooting: Custom MCP Server Integration

Security

Permission Model

Tool-Level - Which tools can the assistant use?
Operation-Level - Which actions are allowed?
Data-Level - Which data can be accessed?
User-Level - Who can use which tools?

Best Practices

Minimal permissions

Grant only necessary rights

Audit logging

Log all tool usage

Secure connections

Always use HTTPS/TLS

Secrets management

Store API keys securely

Rate limiting

Prevent abuse

Use Cases

Data Analysis

Assistant with Database + Charts Tools:
- Retrieve data from database
- Perform analysis
- Create visualization
- Present insights

Research Assistant

Assistant with Web Research + Context7 + Memory:
- Search web for information
- Retrieve technical documentation
- Save important findings
- Create summary

Business Intelligence

Assistant with HubSpot + Stripe + Database:
- Aggregate customer data
- Perform revenue analysis
- Identify trends
- Generate reports

Performance

Optimization

Caching - Cache frequent requests
Batching - Bundle multiple operations
Timeouts - Set appropriate time limits
Parallelization - Use independent tools in parallel

Monitoring

Tool usage statistics
Response times
Error rates
Cost monitoring

Troubleshooting

Common Issues

Tool not responding:

Check connection
Increase timeouts
Analyze logs

Tool works once, then stops working:

Send your next message in the same chat and try the tool call again – every new message rebuilds the tool connection, which often picks up a fresh authorization automatically if the previous one had expired.
If the call keeps failing, open the tool picker (@ in the message box) or go to Settings → Connections and reconnect the affected tool – its stored authorization (especially for OAuth) couldn't be refreshed automatically.
If the problem persists, contact support with the affected chat link.

Unexpected results:

Verify tool configuration
Validate permissions
Check input format

Performance issues:

Enable caching
Optimize requests
Use parallelization

Further Information

Was this page helpful?