Lightweight MCP server for DNS-based agent service discovery over UDP. No registry needed.
mcp-www has 4 tools with inconsistent quality. Tool naming follows verb_noun patterns (lookup_, browse_, call_) which is good. However, descriptions vary significantly in quality and completeness. Input schemas are present but parameter descriptions are minimal or missing. The server lacks error handling guidance, output schema documentation, and parameter constraint details. Tool composition is reasonable but parameter descriptions do not explain dependencies or constraints. No tool annotations (readOnlyHint/destructiveHint) are present despite clear semantic distinctions (call_remote_tool is WRITE, others are READ_ONLY). The code shows raw URL fetching with minimal input validation or actionable error messaging.
Browse a domain for MCP server information. First attempts to fetch server card from .well-known/mcp.json, then falls back to MCP initialize handshake on DNS-advertised servers. Includes IDN homograph detection and DNS TXT record lookup.
Browse an MCP server by direct URL. Attempts to fetch server card from .well-known/mcp.json on the hostname, then falls back to MCP initialize handshake. Checks for DNS TXT records for the hostname.
Call a tool on a remote MCP server. Initializes a session with the server, then invokes the specified tool with provided arguments.
Looks up _mcp TXT DNS records for a domain and parses them into key-value pairs. Returns all discovered records and checks for IDN homograph attack indicators.
Parameter descriptions are minimal or absent. 'domain' param in lookup_mcp_domain has a basic description 'The domain name to look up (e.g., 'example.com')' but does not explain: what format is expected? Should it be normalized? What happens with IDN domains? The 'tool_name' param in call_remote_tool has no constraint explanation, should it match a specific pattern? Can it contain special characters?
Output schemas are not documented. Code shows inspectMcpServer() returns a complex object with 'serverInfo', 'protocolVersion', 'instructions', 'tools', 'resources', 'prompts' fields, but the tool definition does not declare this structure. LLMs cannot predict what fields exist in the response or plan downstream operations. browse_domain and browse_url both call inspectMcpServer() but their tool definitions do not describe the returned structure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 41 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite semantic clarity: lookup_mcp_domain, browse_domain, browse_url are all READ_ONLY (safe for repeated calls); call_remote_tool is WRITE (destructive). Code assigns 'Risk: READ_ONLY' and 'Risk: WRITE' as metadata but these are not present in the MCP tool registration. Without annotations, LLMs cannot determine which tools are safe to retry or have side effects.
Error handling does not guide recovery. Code throws raw errors like 'Server returned 500' or catches 'ENODATA' without actionable guidance. Per pattern:recovery-guide, errors should tell the LLM what to do next. E.g., 'DNS lookup failed for _mcp.example.com, domain may not support DNS-based MCP discovery. Try browsing the server directly via fetch_server_card or initialize handshake.' Current errors are silent or unhelpful.
'arguments' parameter in call_remote_tool is type 'object' with description 'Arguments to pass to the tool', no schema constraint, no validation of what fields are allowed, no documentation of how nested objects or arrays are handled. LLMs will guess at structure and likely send invalid payloads.
Descriptions for browse_domain and browse_url are overly verbose (200+ chars) and bury key details. Per pattern:tool-description baseline, descriptions should be 10 - 1024 chars; these are at the upper end and try to document too much in one description. Should be split into core purpose + detail params. E.g., browse_domain: 'Browse an MCP server discovered via DNS or direct lookup. Returns server info, tools, resources, and prompts.' Additional detail about IDN detection and fallback can be in param descriptions.
Pagination and result limits are not documented. If inspectMcpServer() returns 'tools' array with many items, there is no limit or pagination guidance. Per pattern:paginated-result, tools returning lists should document limits. E.g., 'Returned tools are limited to first 100 items; use pagination cursor if available.'
Input validation is weak. lookupMcpDomain() normalizes domain but does not validate format (e.g., reject invalid TLDs, check for empty strings). call_remote_tool() does not validate 'url' format or 'tool_name' before sending to remote server. Per pattern:tool-gateway, all agent-provided input should be sanitized early with clear error messages.