Skip to content
Learning Outcomes
  • Register and configure external MCP servers in Foundry Connect.
  • Test connections and manage server status.
  • Use MCP tools in AI Insight conversations.
  • Approve or reject tool calls and review results.

MCP Server Management and AI Insight Integration

Foundry Connect lets you connect external MCP (Model Context Protocol) servers — such as PubMed, UniProt, or custom internal tools — and use them directly inside AI Insight conversations. When the AI determines that a tool can help answer your question, it requests permission to call it, and the result is folded into the conversation automatically.

How this relates to MCP Integration

This page covers managing MCP servers inside Foundry Connect and using their tools in AI Insight. If you want to connect an external AI client (Cursor, VS Code, Claude Desktop) to Foundry Connect's own MCP server, see MCP Integration.

Prerequisites

Before you begin, make sure you have:

  • An active Foundry Connect account with AI Insight enabled.
  • At least one AI integration token (OpenAI, Gemini, or Anthropic) configured under Account → Profile → Integration Tokens. See Integration Tokens.
  • The URL and any authentication details for the MCP server you want to register.

Accessing MCP Servers

  1. Open the Cortex section from the main navigation.
  2. Select the MCP Servers tab.

The MCP Servers page displays all servers you have registered, with columns for name, transport type, connection status, tool count, description, last modified date, and an enabled toggle.

Registering a New MCP Server

  1. Click Add MCP Server in the top-right corner.
  2. A drawer opens with the New MCP Server form.

Standard configuration

Fill in the following fields:

Field Required Description
Name Yes A display name for the server (e.g., PubMed, UniProt, Internal Tools).
Server URL Yes The endpoint URL of the MCP server (e.g., https://api.example.com/mcp).
Transport Type Yes The connection protocol. Options: HTTP Streamable, Server-Sent Events (SSE), or WebSocket.
Headers No Key-value pairs for custom HTTP headers (e.g., authentication tokens). Click Add Header to add rows.
Description No A brief description of what the server provides.

JSON configuration (advanced)

For complex setups, toggle JSON Configuration at the top of the form. This replaces the URL, headers, and transport fields with a JSON editor where you can provide the full MCP server configuration object. JSON configuration takes precedence over the standard form fields at runtime.

{
  "url": "https://api.example.com/mcp",
  "transport": {
    "type": "streamableHttp"
  },
  "headers": {
    "Authorization": "Bearer your-token-here"
  }
}

Testing the connection

Before saving, click Test Connection at the bottom of the form. Foundry Connect will attempt to connect to the server and report:

  • Connection Successful — with server name, version, tool count, and latency.
  • Connection Failed — with an error message describing the issue.
Always test before saving

Testing the connection verifies that the URL is reachable, headers are correct, and the server responds with a valid MCP handshake. This avoids saving a misconfigured server.

Once satisfied, click Add MCP Server to save.

Managing MCP Servers

Server list columns

Column Description
Name The display name you assigned.
Transport The protocol type (HTTP Streamable, SSE, or WebSocket).
Connection Current status: Connected, Disconnected, or Error. Hover over an Error badge to see the error message.
Tools Number of tools discovered from the server.
Description The description you provided.
Modified Time since the server was last updated.
Enabled Toggle to enable or disable the server.

Enabling and disabling servers

Use the Enabled toggle on each row to control whether a server's tools are available in AI Insight:

  • Enabled — Foundry Connect connects to the server and its tools appear in AI Insight conversations.
  • Disabled — Foundry Connect disconnects. The server configuration is kept, but its tools are not available.

Editing a server

  1. Click the three-dot menu (Actions) on the server row.
  2. Select Edit.
  3. The edit drawer opens with the current configuration pre-filled.
  4. Make changes and click Save Changes.

If you change connection-related fields (URL, headers, transport type) on an enabled server, Foundry Connect reconnects automatically.

Refreshing tools

MCP servers can add or remove tools over time. To update the tool list:

  1. Click the three-dot menu on the server row.
  2. Select Refresh Tools.

Foundry Connect reconnects, re-discovers available tools, and updates the tool count.

Deleting a server

  1. Click the three-dot menu on the server row.
  2. Select Delete.
  3. Confirm the deletion in the modal.

Deleting a server disconnects it, removes all cached tools, and permanently removes the configuration.

Using MCP Tools in AI Insight

Once you have at least one MCP server enabled with discovered tools, those tools become available in your AI Insight conversations.

How it works

  1. Ask a question in AI Insight as you normally would.
  2. The AI model analyzes your question and determines whether any available MCP tools would help.
  3. If the AI decides to use a tool, a tool call request appears in the conversation.
  4. You approve or reject the tool call.
  5. If approved, the tool executes and the result is incorporated into the AI's response.

Tool call approval (human-in-the-loop)

Every MCP tool call requires your explicit approval. When the AI wants to call a tool, you see a confirmation card with:

  • Server name — which MCP server the tool belongs to.
  • Tool name — the specific tool being called.
  • Tool description — what the tool does.
  • Arguments — click Show arguments to inspect the parameters being sent.

You have two options:

Action What happens
Accept The tool executes with the shown arguments. The result is returned to the AI, which incorporates it into the response.
Reject The tool call is cancelled. The AI is informed and continues without the tool result.
Approval timeout

Tool call requests expire after 60 seconds. If you do not respond in time, the call is automatically treated as rejected.

Tool call results

After a tool call completes, a result bubble appears in the conversation showing:

  • Tool name and server name.
  • Status — Accepted or Rejected.
  • Outcome — Success or Failed (for accepted calls).
  • Duration — How long the tool call took.
  • Arguments — click Show arguments to review what was sent.

Tool call results are saved as part of your conversation history, so you can review them later.

Multi-step tool use

The AI may call multiple tools in sequence within a single conversation turn. For example, it might search a database, then fetch details for a specific result. Each tool call goes through the same approval flow independently.

Supported AI Providers

MCP tools work with all AI providers supported by AI Insight:

Provider Notes
OpenAI Full support. Models like GPT-4 work well with tool calling.
Gemini Full support. Tool names are automatically sanitized for Gemini compatibility.
Anthropic Full support. Claude models support native tool use.

The AI provider is determined by the integration token you select when opening AI Insight. You do not need separate configuration for MCP tool support — it works automatically with any supported provider.

Example: Using PubMed MCP Server

Here is a typical workflow using the PubMed MCP server:

  1. Register the server: Add a new MCP server with the PubMed MCP endpoint URL and any required headers.
  2. Test and save: Click Test Connection to verify, then save.
  3. Enable the server: Ensure the toggle is on.
  4. Open AI Insight: Click the AI Insight button in the header.
  5. Ask a question: Type something like "Find recent papers on CRISPR gene editing in neurodegenerative diseases."
  6. Approve the tool call: The AI requests to call search_articles on PubMed. Review the arguments and click Accept.
  7. Review the response: The AI presents the search results, summarized with key findings and citations.

Security and Privacy

Credential encryption

All sensitive configuration — including authentication headers, API tokens, and advanced JSON configuration — is encrypted at rest. Credentials are only decrypted at the moment Foundry Connect makes a request to the MCP server.

Human-in-the-loop approval

Every MCP tool call requires your explicit approval before execution. The AI cannot silently call tools or access external data without your consent.

Server isolation

Each user manages their own MCP servers independently. Your servers and their tools are visible only to you — even platform administrators see only their own servers in AI Insight conversations.

Invocation logging

All tool calls are logged for audit purposes, including the user who approved the call, the tool name, arguments sent, result returned, execution status, and call duration.

Troubleshooting

Connection issues

Symptom Likely cause Suggested solution
Connection status shows Error Server URL unreachable or incorrect Hover over the Error badge to see the message. Verify the URL and network access.
Connection status shows Disconnected Server is disabled or was manually disconnected Toggle the Enabled switch to reconnect.
Test Connection fails Authentication headers missing or incorrect Double-check your headers and tokens.
Tools count shows 0 Server connected but exposes no tools Verify the MCP server is correctly configured and running. Use Refresh Tools after fixing.

AI Insight issues

Symptom Likely cause Suggested solution
AI never calls any tools No enabled MCP servers, or no tools discovered Check the MCP Servers page: at least one server must be enabled with tools.
Tool call times out You didn't respond within 60 seconds Re-ask the question to trigger a new tool call.
Tool call shows Failed The MCP server returned an error Check the server logs or contact the server administrator.
Wrong AI model used with MCP Integration token mismatch Ensure you selected the correct integration token in AI Insight. All providers (OpenAI, Gemini, Anthropic) support MCP tools.