An MCP server that gives AI assistants the ability to enforce structured output: extract JSON from messy LLM text, gate it against a shape spec, and produce the retry feedback message when the model returns the wrong shape.

Built on top of @mukundakatta/agentcast. Works with Claude Desktop, Cursor, Cline, Windsurf, Zed, and any other MCP client.

Pull a JSON value out of messy LLM output. Tries the whole text, then a fenced ```json ``` block, then the largest balanced {...} / [...] substring. Returns the parsed value plus which strategy succeeded.

{
  "text": "Sure, here you go:\n```json\n{\"answer\": 42}\n```\nLet me know!"
}

→

{
  "value": { "answer": 42 },
  "found": true,
  "source": "fenced_json"
}

source is one of whole, fenced_json, fenced_plain, balanced_substring, or none.

Validate a parsed JSON value against an agentcast shape spec. Spec maps field name to type: string, number, boolean, array, object. Suffix with ? for optional.

{
  "value": { "name": "ada" },
  "shape": { "name": "string", "age": "number" }
}

→

{
  "valid": false,
  "error": "missing required field 'age'"
}

Given an attempt history, produce the validation-error feedback message agentcast appends to the conversation when the model returned the wrong shape. Codifies the "validation error as feedback" pattern for non-Node MCP clients that want to drive the same retry loop manually.

{
  "attempts": [
    { "text": "{\"name\":\"ada\"}", "error": "missing required field 'age'" }
  ],
  "expected_shape": { "name": "string", "age": "number" }
}

→

{
  "feedback": "Your previous response did not match the required shape. Error: missing required field 'age'\n\nTry again. Respond with ONLY valid JSON that fixes the error above.\n\nExpected shape: {\"name\":\"string\",\"age\":\"number\"}"
}

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "agentcast": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/agentcast-mcp"]
    }
  }
}

Same shape, in the appropriate mcp.json for your client. Most clients auto-discover via npx -y @mukundakatta/agentcast-mcp.

npm install -g @mukundakatta/agentcast-mcp
mcp-agentcast        # listens on stdio

When an LLM is supposed to return structured data, it sometimes wraps the JSON in prose, fences, or hallucinated fields. Standard JSON.parse throws. Hand-rolled regex misses nested structure. This MCP server gives any model driving an agent a real handle on (1) pulling JSON out of the response, (2) checking it matches the expected shape, and (3) building the exact retry prompt that nudges the model to fix it on the next turn.

MIT.