Managing and Using MCP Server Connectors

Once you have created an MCP server connector, you can link it to your agents and monitor its operational status in real-time.

DOCUMENTATION SCOPE

This guide covers how to:

• Connect MCP servers to your AI agents.
• Monitor connector health and operational status.
• Manage and toggle individual tools within a server.
• Apply best practices for secure integration.

Adding MCP Access to Agents

To grant an agent access to external capabilities provided by an MCP server, follow these steps:

1. Navigate to Agents in your admin panel.
2. Select an existing agent or click Create New.
3. In the agent configuration, find the MCP Servers section.
4. Select your desired MCP server from the available options.
5. Save the configuration to activate the integration.

Advanced Server & Tool Management

MCP servers often expose multiple specialized functions (tools). You have granular control over which of these tools are active.

Viewing Tool Details & Descriptions

To understand exactly what capabilities a server provides:

1. Locate your server in the MCP Servers list.
2. Click the three dots (⋮) menu icon.
3. Select View Details.
4. You will see a complete list of tools exposed by the server, including their specific descriptions as defined in the server's manifest.

Enabling and Disabling Tools

You can manage the "surface area" of your agent's capabilities within the View Details pane:

• Toggle Switches: Use the toggle next to each tool to enable or disable it individually.
• Granular Control: If an agent only needs one specific function from a large server, you can disable the rest to improve agent focus and security.

Understanding Server Status

Monitor the health of your connectors directly from the dashboard using the status indicators:

Status	Indicator	Meaning
Ready	✅ Green	Server is operational. Agents can successfully make requests.
Error	❌ Red	Connectivity or configuration issue. Check endpoint and headers.

note

Platform-hosted MCP servers (running directly on our infrastructure) are coming soon.

Returning Files from MCP Tools

MCP tools can return files (documents, spreadsheets, presentations, images, etc.) that are automatically attached to the user's conversation as downloadable attachments.

No Artifact Hosting

The platform does not host files at any URL. Files returned by MCP tools are ingested as conversation attachments — there is no /artifacts/... path or downloadable link generated by the platform. The model should never output a fabricated URL for a returned file.

How It Works

When an MCP tool returns a file, the platform ingests it and attaches it to the conversation automatically. The UI renders it as a downloadable attachment, and the tool result includes an ingestion note listing the attached file(s). No additional action is required from the MCP server after returning the file.

Supported Methods

There are two ways for an MCP tool to return a file:

1. EmbeddedResource with BlobResourceContents (Recommended)

Return the file as a base64-encoded blob using the MCP EmbeddedResource type with BlobResourceContents. You must set:

• mimeType: The correct MIME type (e.g., application/vnd.openxmlformats-officedocument.wordprocessingml.document for DOCX)
• uri: A URI or filename ending with the correct extension (e.g., report.docx, slides.pptx)

{
  "type": "resource",
  "resource": {
    "uri": "file://output/report.docx",
    "mimeType": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
    "blob": "<base64-encoded-content>"
  }
}

2. Fetchable URL in Tool Text Result

Include an http:// or https:// URL in the tool's text result. The platform downloads the file and attaches it to the conversation. The remote server must provide either:

• A correct Content-Type header in the HTTP response, or
• A filename with the correct extension in the URL path

{
  "type": "text",
  "text": "Report generated successfully: https://internal-api.example.com/files/report.docx"
}

URL Limits

A maximum of 5 URLs per tool result are processed. Additional URLs are ignored.

If the URL points to a ZIP archive, the platform automatically extracts its contents and attaches individual files (up to 50 files per archive). ZIP-based document formats (DOCX, PPTX, XLSX, EPUB, etc.) are kept intact and not extracted.

Supported File Types

Category	Formats
Documents	PDF, DOCX, PPTX, XLSX, CSV, TXT, MD, RTF, EPUB, ODT, ODS, ODP
Email	EML, MSG
Images	PNG, JPEG, JPG
Audio	MP3, WAV, M4A, OGG, FLAC, AAC, Opus
Video	MP4, WebM, MOV, MKV

Size Limits

Maximum file size per ingested file: 512 MB.

Important Notes

• After ingestion, the file is attached to the conversation and visible to the user immediately — no upload callback or URL is returned to the model.
• The model should not attempt to type or fabricate a download URL. The UI handles file presentation automatically.
• The tool result includes an ingestion note confirming which file(s) were attached (filename, MIME type, size).
• Deduplication: If the same source URI is returned again in the same conversation, the existing attachment is reused instead of re-ingesting.

---

Best Practices for Management

To ensure high performance and stability, follow these industry standards:

• Principle of Least Privilege: Only enable the specific tools an agent requires via the View Details toggle.
• Health Monitoring: Regularly verify your endpoint's health, especially for self-hosted connectors.
• Credential Rotation: Update authentication headers and API keys periodically to maintain security.
• Response Testing: Use tools like curl or Postman to verify server responses before testing within an agent chat.

PRACTICAL EXAMPLES

Looking for real-world scenarios? Check out Chat with Agents - MCP Integration to see sample queries, responses, and workflow integrations.