Skip to main content

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

TypeUse it whenRequirements
Built-in ConnectorThe application already provides the required data source or operationEnable it for the intended agent; some services also require credentials
Local MCP serverA tool must run on this computer or read locally available research dataAn installed launcher, server and permitted input paths
Remote MCP serverA service exposes its tools through a hosted MCP endpointThe 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.

StageWhat to checkNext action
ConfigurationCorrect launcher or endpoint and required authenticationComplete the Connector form
ConnectionThe server responds and its tools are discoveredInspect the connection result and actual errors
Agent accessThe resource is enabled for Main or assigned to the intended SpecialistCheck that agent's capability bindings
OperationThe chosen tool accepts the request and returns the required dataUse the operation reference for built-in inputs
Research resultReturned identifiers, source and scope match the taskInspect 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

TaskMain instructions
Add/edit a server, import several servers or transfer configurationConfigure Connectors
Bind API keys or OAuth credentialsService credentials
Implement a small server and call its toolsCreate a custom tool
Look up database operation fieldsConnector operation reference
Manage connections from scriptsCLI 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.