A dependency-light Go port of the reference Python mcp-server-fetch tool. Fetches a URL, optionally extracts the page's main content as Markdown, and returns text ready to hand to an LLM.
Single tool 'Fetch' has a well-structured input schema with 9 optional parameters, each with type and description. Tool description is clear and actionable (156 chars). However, output schema is not documented, the description mentions wrapping content in a prefix but does not specify the response structure (fields, types). Parameter descriptions are generally good but lack format constraints (e.g., MaxLength range, URL validation rules). No error handling guidance provided. Naming is verb-based ('Fetch') but generic; 'fetch_url' or 'fetch_webpage' would be more specific. Security: SSRF guard mentioned in description is good, but no explicit documentation of what happens on auth failures, timeouts, or malformed URLs.
Fetches the URL, extracts/keeps the content, applies start_index/max_length paging, and returns the text wrapped as "<prefix>Contents of <url>:\n<content>". Outbound connections are restricted to public IPs by the SSRF guard in the dialer.
Output schema not documented. Description mentions wrapping content but does not specify response fields, types, or structure. LLMs cannot plan downstream operations without knowing what fields to extract.
Parameter constraints missing. MaxLength, StartIndex, and numeric fields lack min/max bounds. URL parameter has no format validation rule (e.g., must be http/https). LLMs may pass invalid values.
No error handling guidance. Description does not explain what happens on network errors, timeouts, invalid URLs, or auth failures. LLMs have no recovery path.
Tool name is generic. 'Fetch' does not clearly distinguish this from other HTTP tools. 'fetch_webpage' or 'fetch_and_extract_content' would better signal the extraction and markdown conversion behavior.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Pagination not addressed. If content is truncated (MaxLength), the tool supports StartIndex for paging, but no guidance on detecting truncation or iterating. Response should include a 'truncated' flag or 'total_length' field.