MCP integrations: remote HTTP servers as skills

Prime Agent's MCP integration path connects to remote HTTP servers and wraps them through McpIntegration as Python-callable capabilities. Built-in services and custom server endpoints use this networked transport.

This path does not wire stdio MCP servers into the same integration. A server that is available only over stdio requires a different connection path rather than an HTTP configuration guess.

Built-in MCP providers can use login flows, while custom services may rely on OAuth or bearer credentials appropriate to the remote server. Those credentials remain part of the local Prime Agent trust boundary.

Authentication grants the remote tool whatever account scope the service exposes. Review that scope and avoid assuming that successful login makes a third-party tool or its responses trustworthy.

Custom MCP tools are discovered from the server at runtime rather than fixed permanently in Prime Agent's source. Discovery supplies the current schemas that are then exposed through the Python-callable integration.

Because a remote service can change its advertised tools or schemas independently of Prime Agent releases, workflows should verify the discovered interface instead of relying indefinitely on an old capture.

An MCP call crosses from the local agent to a third-party server and may send task data covered by that tool's arguments. The remote service, its schema, and its authentication handling all expand the effective trust boundary.

Use only reviewed services and limit credentials to the access the workflow needs. Python-callable convenience does not make remote behavior local, isolated, or stable.