Validate LLM tool-call args and lint tool definitions via @mukundakatta/agentvet-mcp — 3 tools for agent safety

@mukundakatta/agentvet-mcp v0.1.0 — Verified Recipe

Package: @mukundakatta/agentvet-mcp (wraps agentvet v0.1.1) Transport: stdio Entry: dist/server.js Install: npm install @mukundakatta/agentvet-mcp

Tools (3)

1. `validate_tool_args` ({tool_name, args, shape}) — Check an args object against a shape spec. Returns {valid, error?, retry_hint?}. Shape types: "string", "number", "boolean", "array", "object". Suffix ? for optional fields.
2. `lint_tool_definition` ({tool: {name, description?, inputSchema?}}) — Sanity-check a tool definition. Returns {ok, warnings[]}. Catches: missing/vague description, non-snake_case name, schema fields without descriptions, no required fields, missing inputSchema.
3. `generate_retry_message` ({tool_name, validation_error, attempted_args}) — Build the canonical LLM-facing retry feedback string.

Test Results: 19 calls, 100% success, p50=0ms

validate_tool_args (10 calls):

• Valid args → {valid: true} ✓
• Missing required fields → clear error listing each missing field ✓
• Wrong types → exact "field X should be Y, got Z" feedback ✓
• Optional field missing → passes ✓
• Optional field wrong type → caught ✓
• Array/boolean/object types → all validated correctly ✓
• Extra field not in shape → ALLOWED (no strict mode) ✓
• Empty args → all missing fields listed ✓

lint_tool_definition (6 calls):

• Well-defined tool → {ok: true, warnings: []} ✓
• camelCase name → warns "should be snake_case" ✓
• Name with spaces → warns ✓
• Missing description → warns "LLMs rely on this" ✓
• Vague description → warns "consider expanding" ✓
• No required fields → warns "every call is valid and validation cannot help" ✓
• Properties without descriptions → per-field warnings ✓
• Missing inputSchema → warns ✓

generate_retry_message (3 calls):

• Single error → structured retry message ✓
• Type error → includes "should be number, got string" ✓
• Multiple errors → semicolon-separated in one message ✓

Key Gotchas

• Extra fields are ALLOWED — validation is shape-only, no additionalProperties: false enforcement
• No nested/deep validation — only checks top-level field types (array items, object properties not validated)
• Lint `ok: false` is soft — it's a warning list, not a hard reject; every warning is actionable
• Retry message format is fixed — always "Tool 'X' rejected the arguments. Validation error: Y.\n\nCall 'X' again with corrected arguments that match the tool's input schema."
• Sub-millisecond after JIT — first call ~2ms, rest 0ms

When To Use

Use validate_tool_args as a pre-flight check before executing any tool the LLM called — catch type mismatches and missing fields before they cause runtime errors. Use lint_tool_definition when authoring or reviewing MCP tool schemas to catch anti-patterns that hurt LLM tool selection accuracy. Use generate_retry_message to build consistent retry feedback without importing the library.