Loading ToolBench

Reading the current corpus…

clean-fetch MCP server · D · ToolBench

clean-fetch

A clean, trustworthy WebFetch replacement for Claude Code. Extracts main content using Trafilatura, removes known prompt injection patterns, supports FQDN whitelist, and provides SSRF protection.

2026-07-28+STDIO1 tools

52/100

Definition Quality53

Protocol Readiness50

SupportabilityNot measured-

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

Single tool 'fetch_clean' has a clear verb-noun name and addresses a real security need (SSRF protection, injection filtering). However, the tool description in code is incomplete (truncated at docstring start), parameter descriptions are minimal, and output schema is undocumented. The 'mode' parameter lacks an enum constraint despite having only two valid values ('main', 'full'). No error handling guidance is visible, the tool returns a string but does not document what happens on failure (timeout, invalid URL, network error). The implementation shows good security practices (SSRF checks, injection pattern masking) but the tool definition itself lacks the rigor needed for reliable LLM invocation.

Tools (1)

fetch_cleanread onlysource verified53/100

Fetch a URL and return clean Markdown content. Use this instead of WebFetch for all web content retrieval. Extracts main content with high accuracy, removes injection patterns. Blocks access to private/internal network addresses.

Issues (5)

CRITICALtool-description ↗

Tool description is incomplete/truncated in source. The docstring for fetch_clean() ends abruptly after 'Fetch a URL and return clean Markdown content.' without explaining return format, error cases, or when to use 'mode' parameter.

fetch_clean

HIGHconstrained-input ↗

'mode' parameter accepts only 'main' or 'full' but is defined as a free-form string with no enum constraint. LLMs will hallucinate invalid values like 'summary', 'extract', 'minimal'.

fetch_clean

HIGHtool ↗

Output schema is undocumented. Tool returns a string but does not specify: Is it always Markdown? What happens on timeout (15s)? On invalid URL? On SSRF block? On extraction failure? LLMs cannot plan error recovery without this.

fetch_clean

HIGHrecovery-guide ↗

No error handling guidance. The code validates URLs and blocks private IPs, but the tool definition does not explain what error messages to expect or how to recover. A bare string return makes it impossible to distinguish success from failure.

Recommendations

Complete the fetch_clean() docstring. Add: 'Returns: Markdown string of extracted content. On error, returns error message starting with "Error:". Raises timeout after 15s. Blocks private/internal IPs (127.0.0.0/8, 10.0.0.0/8, 192.168.0.0/16, etc.).'
Convert 'mode' parameter to an enum: {"type": "string", "enum": ["main", "full"], "default": "main", "description": "'main' extracts article body only (faster, cleaner). 'full' includes sidebars, related articles, and supplementary sections (more comprehensive but noisier)."}
Document output schema explicitly: 'Returns a Markdown string. On success, contains extracted article text. On error, returns a string starting with "Error: " (e.g., "Error: access to private/internal addresses is not allowed (192.168.1.1)."). On timeout, returns "Error: request timed out after 15s."'
Add parameter descriptions for 'url': 'The URL to fetch (http or https only). Must be publicly accessible. Private/internal addresses (127.0.0.0/8, 10.0.0.0/8, 192.168.0.0/16, 172.16.0.0/12, 169.254.0.0/16) are blocked for security.'
Enhance tool description to ~150 chars: 'Fetch a URL and extract clean Markdown content. Removes HTML noise, script/style tags, and known prompt injection patterns. Blocks private/internal network addresses (SSRF protection). Returns Markdown or error message. Use instead of raw WebFetch for untrusted web content.'

Score history

Overall score trend

First recorded score · v2 rubric

52/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	52	2026-07-28+	v2