MCP server for agentvet: validate LLM-generated tool-call args before execution. Returns a typed error with an LLM-friendly retry hint when the args don't match.
Three focused validation tools with clear purposes and complete input schemas. All tools have descriptions (50-150 chars) and properly typed parameters. However, descriptions lack LLM-optimization guidance (WHEN to use, dependencies, examples of success/failure). Output schemas are not documented, callers cannot predict response structure. Error handling is basic (returns JSON with 'valid', 'error', 'retry_hint' fields) but lacks recovery guidance for edge cases. No parameter descriptions for nested objects (e.g., 'shape' in validate_tool_args lacks field-level guidance). Tool composition is sound, each does one thing, but lacks integration hints.
Given a tool name, validation error, and attempted args, build the canonical LLM-facing retry feedback message. Uses agentvet's ToolArgError.toLLMFeedback() formatting so the wording matches what runtime callers see.
Sanity-check a tool definition for common mistakes that hurt LLM tool-use accuracy: missing description, vague description, no required fields, schema fields without descriptions, non-snake_case names.
Validate a tool-call args object against a small shape spec. Returns { valid, error?, retry_hint? } where retry_hint is a ready-to-send LLM feedback message describing exactly what was wrong.
Output schemas not documented. Callers cannot predict response structure (e.g., validate_tool_args returns {valid, error?, retry_hint?} but this is not in the tool definition). LLMs need explicit output documentation to plan downstream actions.
Parameter 'shape' in validate_tool_args lacks field-level guidance. Description says 'Shape spec mapping field name to type' but does not explain how to construct it, what happens if a field is missing, or how optional fields are marked. Example is helpful but insufficient.
Descriptions lack WHEN-to-use guidance and integration hints. E.g., validate_tool_args does not explain: 'Call this after an LLM generates tool args but before executing the tool.' This forces LLMs to infer the workflow.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Error handling returns raw validation error strings without recovery guidance. E.g., 'missing required field: user_id' does not tell the LLM what to do next (ask user? retry with different args? call a lookup tool?).
Tool 'lint_tool_definition' accepts a 'tool' parameter with optional 'description' and 'inputSchema' fields, but the schema marks only 'name' as required. This allows callers to pass incomplete tool definitions, which lint_tool_definition then warns about, but the tool itself should enforce minimum structure.