Agentvet MCP server

agentvet-mcp

MCP server for @mukundakatta/agentvet. Lets Claude Desktop, Cursor, Cline, Windsurf, Zed, or any other MCP client validate LLM-generated tool-call args before execution and produce LLM-friendly retry messages when something's wrong.

npx -y @mukundakatta/agentvet-mcp

Three tools:

• validatetoolargs — check args against a small shape spec; returns { valid, error?, retryhint? } where retryhint is a ready-to-send LLM feedback message.
• linttooldefinition — sanity-check a tool definition for common mistakes that hurt LLM tool-use accuracy.
• generateretrymessage — given a validation error, build the canonical LLM-facing retry message using agentvet's ToolArgError.toLLMFeedback() formatting.

Add to your client

Claude Desktop

Edit ~/Library/Application Support/Claude/claudedesktopconfig.json (macOS) or %APPDATA%\Claude\claudedesktopconfig.json (Windows):

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

Same shape for Cursor (~/.cursor/mcp.json), Cline, Windsurf, Zed.

Tool examples

validatetoolargs:

{
  "tool_name": "send_email",
  "args": { "to": "a@b.com" },
  "shape": { "to": "string", "subject": "string", "body": "string" }
}

Returns:

{
  "valid": false,
  "error": "missing required field: subject",
  "retry_hint": "send_email rejected your args: missing required field: subject. Please call again with the corrected arguments."
}

linttooldefinition:

{
  "tool": {
    "name": "BadName",
    "inputSchema": { "type": "object", "properties": { "x": { "type": "string" } } }
  }
}

Returns warnings about non-snake_case name, missing description, missing field descriptions, and no required fields.

generateretrymessage:

{
  "tool_name": "send_email",
  "validation_error": "missing required field: subject",
  "attempted_args": { "to": "a@b.com" }
}

Returns the canonical retry feedback string the runtime callers see — so you can prepare retry text outside the live agent loop.

Why a separate MCP server

@mukundakatta/agentvet is a zero-dependency JavaScript library. This MCP server makes its validation primitives accessible from any MCP-aware AI assistant. Useful for quickly auditing a registry of tools, or asking the assistant "is this args object valid for my send_email tool?" without leaving the chat.

For runtime arg validation in your agent loop, use @mukundakatta/agentvet directly inside your Node process (it wraps your tool fn and throws ToolArgError synchronously).

Sibling MCP servers

Part of the agent-stack series:

• @mukundakatta/agentfit-mcp — Fit it.
• @mukundakatta/agentguard-mcp — Sandbox it.
• @mukundakatta/agentsnap-mcp — Test it.
• @mukundakatta/agentvet-mcp — Vet it. (this)
• @mukundakatta/agentcast-mcp — Validate it.

License

MIT