Loading ToolBench

Reading the current corpus…

ClawForge MCP server · D · ToolBench

ClawForge

github.com/jackjin1997/ClawForge

A collection of modular skills and tools for Claude Code, including domain-driven design modeling, news extraction, GitHub trending analysis, LangSmith dataset/trace management, and evaluation utilities.

<=2025-11-25STDIO16 tools

50/100

Definition Quality65

Protocol Readiness37

Supportability35

Spec posture

Inferred effective spec <=2025-11-25·current baseline 2026-07-28

Static source inference · medium confidence · detected: Logging

Deprecated protocol patterns detected

Summary

ClawForge is a collection-focused MCP server with 16 tools across GitHub trending, LangSmith dataset/evaluator/trace management, and news extraction. Definitions show moderate quality but with significant gaps: parameter descriptions are present for most tools, but many descriptions are terse (15-40 chars), schema completeness varies, and error handling is not visible. News extraction tools lack output schema documentation. LangSmith tools have reasonable parameter sets but descriptions are often generic. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are declared. Tools are well-scoped (single responsibility), but descriptions don't consistently explain WHEN to use them or what prerequisites exist.

Tools (16)

deletedestructiveauthsource verified73/100

Delete an evaluator by name from LangSmith.

exportwriteauthsource verified70/100

Export LangSmith dataset to a local file.

extract_news_from_neteaseread only50/100

Extract news content from Netease News (网易新闻) articles.

extract_news_from_sohuread only50/100

Extract news content from Sohu News (搜狐新闻) articles.

extract_news_from_tencentread only50/100

Extract news content from Tencent News (腾讯新闻) articles.

extract_news_from_toutiaoread only50/100

Extract news content from Toutiao (今日头条) articles.

extract_news_from_wechatread only50/100

Extract news content from WeChat public account articles.

fetch_trendingread only50/100

Fetch and basic extract GitHub Trending repos. Takes a span parameter (daily, weekly, monthly) and returns GitHub trending repositories.

Issues (6)

HIGHresponse-shaper ↗

News extraction tools (extract_news_from_*) lack documented output schemas. Descriptions state 'Extract news content...save to JSON' but do not document the structure of returned data (fields, types, nested objects). LLMs cannot plan downstream processing without knowing what fields to extract.

extract_news_from_wechatextract_news_from_toutiaoextract_news_from_neteaseextract_news_from_sohuextract_news_from_tencent

HIGHtool-annotations ↗

No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are declared in any tool. This is a current MCP spec requirement for production tools. Example: 'delete' and 'upload' tools should declare destructiveHint=true and writeHint=true respectively so clients can warn before execution.

deleteuploadexport

MEDIUMtool-description ↗

Generic and terse tool descriptions. 'list_datasets' → 'List all LangSmith datasets...' (47 chars) lacks WHEN to use it and what it enables downstream. Best practice: 'Retrieve all available datasets from LangSmith with names, IDs, and counts. Call this first to discover dataset names before querying with show() or exporting with export().' (description avg in rubric: 194 chars).

Recommendations

Document output schemas for all news extraction tools. Example for extract_news_from_wechat: 'Returns JSON object with fields: {title: string, content: string, author: string, publish_date: ISO8601, source_url: string, images: string[]}'. Make this visible in the tool definition so LLMs know how to process results.
Add tool annotations to all tool definitions: readOnlyHint=true for GET-like tools (fetch_trending, list_datasets, show, view_file, structure, list, recent, trace, extract_news_*); destructiveHint=true for delete; readOnlyHint=false + writeHint=true for export and upload. This enables client UIs to show warnings.
Rename 'list' evaluators tool to 'list_evaluators' to avoid name collision with 'list_datasets'. Follow verb_noun pattern consistently.
Expand tool descriptions to 150-200 characters and include WHEN guidance. Example: 'show' → 'Display a sample of examples from a LangSmith dataset. Use this to preview data before evaluation or export. Supports JSON and pretty-print formats; call list_datasets() first if you need to find dataset names.'
Add recovery hints to parameter descriptions where user input is required. Example: 'dataset_name (string, required): Name of the LangSmith dataset. If unknown, call list_datasets() first to discover available datasets.'
Document constraints on news extraction save_path parameter: 'Absolute or relative file path where extracted JSON will be saved. Directory must exist or be creatable. Default: ./data/'
Add idempotentHint=true annotations to read-only and safe tools. This helps agents and clients understand which tools are safe to retry without side effects.

Spec posture evidence

Inferred effective spec: <=2025-11-25.

Relies on Logging (deprecated) - log to stderr or use OpenTelemetry

Score history

Overall score trend

↑ 4 points across a rubric change (v1 → v2)

50/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-22	D	50	<=2025-11-25	v2
2026-03-09	F	46	-	v1