Connectors and MCP
A Connector exposes tools that an agent can call. Use this page to choose the integration type and understand readiness. For field-by-field setup, import/export and connection management, follow Configure Connectors.
Choose a connection type
| Type | Use it when | Requirements |
|---|---|---|
| Built-in Connector | The application already provides the required data source or operation | Enable it for the intended agent; some services also require credentials |
| Local MCP server | A tool must run on this computer or read locally available research data | An installed launcher, server and permitted input paths |
| Remote MCP server | A service exposes its tools through a hosted MCP endpoint | The actual MCP endpoint, supported HTTP/SSE transport and any required authentication |
Molecule provides local structure rendering; it is not a remote database. Choose a source in Scientific databases, or use Scientific viewers to inspect molecular and sequence files. Selecting a Docker launcher does not install a container engine or server image.
Distinguish connection from availability
Check these stages in order. Passing one does not establish that the next has succeeded.
| Stage | What to check | Next action |
|---|---|---|
| Configuration | Correct launcher or endpoint and required authentication | Complete the Connector form |
| Connection | The server responds and its tools are discovered | Inspect the connection result and actual errors |
| Agent access | The resource is enabled for Main or assigned to the intended Specialist | Check that agent's capability bindings |
| Operation | The chosen tool accepts the request and returns the required data | Use the operation reference for built-in inputs |
| Research result | Returned identifiers, source and scope match the task | Inspect the response before using it as evidence |
An enabled switch is not a successful service call. A metadata response is not a downloaded full text or count matrix. A custom row may open its configuration editor; built-in details provide the tool specification.
Choose the next step
| Task | Main instructions |
|---|---|
| Add/edit a server, import several servers or transfer configuration | Configure Connectors |
| Bind API keys or OAuth credentials | Service credentials |
| Implement a small server and call its tools | Create a custom tool |
| Look up database operation fields | Connector operation reference |
| Manage connections from scripts | CLI or SDK |
Diagnose the failing stage
For connection failures, identify whether the launcher, transport or authentication failed. If connection succeeds but a call fails, check agent availability and that tool's arguments before changing the server configuration. Report the actual returned error using Troubleshooting.
MCP discovery is performed during connection. host.mcp("server", "tools/list", {}) is not an application tool call and can return unknown tool. The custom example also surfaced an internal unknown-sample failure as connector_unavailable; that response alone does not distinguish an invalid sample from a transport failure.