src/mcp/convert.rs

This file holds the mapping between loopctl types and MCP wire types, in both directions, behind the mcp feature.

It also holds the safety validation on the way out.

Key items

Inbound you get bridge_content(&ContentBlock) -> ToolContentPart, which covers text, images, and embedded resources as text notes, and nothing is silently dropped.

bridge_result(...) -> ToolOutput maps isError to is_error and carries structuredContent as a JSON-stringified text part, and an error with no content becomes McpError::EmptyToolError.

Outbound you get tool_schema_to_mcp(schema, is_read_only) -> Option<McpTool>, which does the tool to name rename.

Validation runs there, and your schema must compile under jsonschema with external references refused through RefusingRetriever. So there is no IO during your tools/list, and the root must be a typed object.

Anything else gives you None plus a warning, and the tool is omitted.

dispatch_result_to_call_tool(...) turns every tool outcome into a tool-level result, hard failures included.

not_found_error(requested, available) builds the METHOD_NOT_FOUND payload you send back.

Behavior notes

Read-only forwarding is consistent with the spec’s defaults, so your tool gets readOnlyHint: true plus destructiveHint: false, and a non-read-only tool gets no annotations at all.

The validation is the security boundary for the schemas you serve. A $ref pointing at a URL or file path would make your server do IO mid-listing, so it is refused instead.

MCP’s name convention, ^[a-zA-Z0-9_-]{1,64}$, is the embedding’s responsibility, since loopctl imposes no rules on you.

For the deep dive, see MCP.