Tendrl
Docs / Strand / connectors/mcp-server
MCP Server
Direction: Read / Write | Type: mcp.server
Model Context Protocol (MCP) server connector for providing external tools to AI nodes. MCP connectors are not used as standalone nodes; they are selected within an AI node's configuration to give the model access to external tools.
Operations#
| Operation | Direction | Description |
|---|---|---|
tool_call |
Read / Write | Discover and execute tools from the MCP server |
Connector Configuration#
| Field | Required | Default | Description |
|---|---|---|---|
| URL | Yes | - | MCP server endpoint (Streamable HTTP) |
| Auth Type | No | none |
Authentication method: none, api_key, bearer, or oauth |
| API Key | No | - | API key (sent as Authorization: Bearer <api_key> header, encrypted) |
| Bearer Token | No | - | Bearer token (sent as Authorization: Bearer <bearer_token> header, encrypted) |
| Custom Headers | No | - | Additional HTTP headers for requests |
| Tool Prefix | No | - | Prefix added to all tool names (avoids collisions) |
| Timeout | No | 30 |
Request timeout in seconds (5-300) |
Using MCP with AI Nodes#
- Create an MCP server connector with the server URL and authentication
- Add an AI connector node (OpenAI, Anthropic, or Gemini) to your workflow
- In the AI node configuration, select the MCP server connector(s)
- The AI model can then invoke tools provided by the MCP server
You can attach multiple MCP server connectors to a single AI node, giving the model access to tools from multiple servers. Use Tool Prefix to avoid name collisions (e.g., prefix mydb makes tool query available as mydb_query).
The URL must be reachable from Strand#
The server address is checked before the node runs, and a private or loopback address is refused with MCP server URL rejected. That covers localhost, the private ranges, and cloud metadata addresses, and it applies to the URL on the connector and to any inline server set on the node. Testing the URL in your own browser proves nothing — it has to resolve and be reachable from outside your network. See Connectors overview for the ways to expose a server you host yourself.
A refused address fails the node rather than quietly dropping that server from the model's tools, so a typo cannot leave the model running with fewer tools than the workflow asked for.
Health Monitoring#
MCP server connectors include health monitoring with a circuit breaker:
- Test Connection: verifies the server is reachable and lists available tools
- Circuit Breaker: auto-disables the connector after 3 consecutive failures
- Reset: manually re-enable a disabled connector from the connector manager
Example#
Connector Setup:
{
"url": "https://mcp.example.com/mcp",
"auth_type": "api_key",
"api_key": "your-api-key",
"tool_prefix": "myapi",
"timeout": 60
}
AI Node Configuration:
- Select the MCP server connector in the "MCP Servers" dropdown
- The AI model will automatically discover and can invoke tools from the server
Related#
- MCP Integration Overview: Full tool list, client setup, and orchestration patterns
- OpenAI | Anthropic Claude | Google Gemini
- Connector Output Structure
