René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Response format schema for prompts","articleBody":"**Is your feature request related to a problem? Please describe.**\nCurrently, I am developing an MCP server that manages prompt templates for parameterized invocation by clients. The current implementation follows a pattern like `promptsHolder.getPrompt({name: \"myPrompt\", arguments: {...}})`, which returns an array of messages to send to the LLM agent.\n\nSome of my prompts require a specific output format from the LLM. At present, I rely on few-shot prompting techniques to enforse desired structured output. However, this approach sometimes leads to inconsistent results, such as not expected enums values, or markdown-style quotation.\n\n**Describe the solution you'd like**\nI propose that the MCP server should be able to specify a JSON schema for the expected response. The schema definition would take place on the server side, while the client would handle the final adaptation for specific host models. A common option to consider could be [response_format](https://platform.openai.com/docs/guides/structured-outputs?api-mode=chat). \n\nA potential implementation might look like this:\n\u003e /server/api/getPrompt/myPrompt\n```\n{\n  \"jsonrpc\": \"2.0\",\n  \"id\": 2,\n  \"result\": {\n    \"description\": \"Code review prompt\",\n    \"options\": {\n      \"response_format\": {\"type\": \"json_object\", \"schema\": __jsonSchema__},\n    }\n    \"messages\": [\n      {\n        \"role\": \"user\",\n        \"content\": {\n          \"type\": \"text\",\n          \"text\": \"Please review this Python code:\\ndef hello():\\n    print('world')\"\n        }\n      }\n    ]\n  }\n}\n```\n\n**Describe alternatives you've considered**\n- few shots prompting, that works in most cases, but misfires time to time \n- tool calling, when the tool with structured params definition is provided to the model — this is common workaround, and even anthropic [recommended](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview#json-output) to use it as claude does not accept a schema parameter, but it still a workaround which add unnecessary complexity\n\n**Additional context**\nThere is related discussion https://github.com/modelcontextprotocol/specification/issues/97 and even same question is [raised](https://github.com/modelcontextprotocol/specification/issues/97#issuecomment-2539110532), but that topic is generally consider another feature.\n\nI understand that such feature significantly increases API-surface. And at least raise some questions. \n\nServer can't be sure for 100% that client or host implement that feature. So it need to proceed validation on own side and even raise errors. Should it be reflected in protocol somehow? Like a requirement from the client to inform the server about the available options.\n\nWould appreciate your thoughts about this feature. Would it be on roadmap or better to consider alternative options?","author":{"url":"https://github.com/gleb-lobastov","@type":"Person","name":"gleb-lobastov"},"datePublished":"2025-03-28T10:46:30.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":6},"url":"https://github.com/249/modelcontextprotocol/issues/249"}