Loading ToolBench

Reading the current corpus…

websearch-mcp MCP server · D · ToolBench

websearch-mcp

Self-hosted MCP server for deep internet research — web search, page scraping, Python sandbox, and search suggestions. No API keys required.

2026-07-28+STDIO4 tools

59/100

Definition Quality85

Protocol Readiness50

Supportability21

Spec posture

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

Static source inference · medium confidence

Current-spec patterns detected

Summary

Server has 4 well-named tools with complete input schemas and descriptions. All tools follow verb_noun naming (search_internet, fetch_page, execute_python, get_search_suggestions). Descriptions are detailed (150-250 chars) and explain WHAT, WHEN, and HOW. Input schemas use Zod with proper types, enums, and constraints. However, output schemas are not documented, responses are JSON-stringified without field-level documentation. Error handling returns structured error objects with codes and messages, but lacks recovery guidance (e.g., 'try search_internet with a simpler query'). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Missing parameter descriptions for some edge cases (e.g., wait_for CSS selector behavior).

Tools (4)

execute_pythonwritesource verified80/100

Execute Python code in an isolated Docker sandbox with network disabled, capability dropping, and resource limits (512MB RAM, 1 CPU, 60s timeout). Auto-detects and installs missing packages.

fetch_pageread onlysource verified85/100

Fetch and extract content from a web page using Playwright with stealth mode. Supports snippet, readable, and full extraction modes. Content is truncated to 50,000 characters.

get_search_suggestionsread onlysource verified85/100

Get autocomplete suggestions for a partial search query from SearxNG. Returns up to max_suggestions suggestion strings. Falls back to simple query expansion when SearxNG autocomplete is unavailable.

search_internetread onlysource verified90/100

Search the internet using SearxNG meta-search engine. Returns search results with title, URL, snippet, source engine, and score. Supports speed, balanced, and quality modes. Deduplicates by URL and limits same-domain results to 2. Filters low-trust domains when website-weight-config.json is present.

Issues (5)

HIGHresponse-shaper ↗

Output schemas not documented. Tools return JSON-stringified results without field-level documentation. LLMs cannot plan downstream calls or extract specific fields reliably.

search_internetfetch_pageexecute_pythonget_search_suggestions

HIGHtool ↗

No tool annotations (readOnlyHint/destructiveHint/idempotentHint). execute_python is clearly destructive (writes to Docker, runs arbitrary code) but lacks annotation. Agents cannot reason about side effects.

execute_python

MEDIUMrecovery-guide ↗

Error messages lack recovery guidance. Errors return code + message but do not suggest next steps (e.g., 'SearxNG unavailable, retry in 30s or check config').

search_internetfetch_pageexecute_python

MEDIUMtool-description ↗

fetch_page wait_for parameter lacks detail on CSS selector behavior (e.g., timeout if selector never appears, partial match rules).

Recommendations

Document output schemas for all tools. Example: search_internet returns {results: [{title, url, snippet, source, score}], total_count, deduped_count}. Include field types and meanings.
Add tool annotations: mark search_internet, fetch_page, get_search_suggestions with readOnlyHint=true; mark execute_python with destructiveHint=true and idempotentHint=false.
Enhance error messages with recovery guidance. E.g., 'SearxNG connection failed (ERR_SEARXNG_UNAVAILABLE). Verify SearxNG is running at configured URL. Retry in 30s or check config/logs.'
Clarify fetch_page wait_for behavior: 'CSS selector to wait for (max 10s). If not found, returns page content anyway. Use for dynamic content that loads after page render.'
Document execute_python artifact format and retrieval: 'If save_artifacts=true, output files are saved to /tmp/artifacts/ and returned as {filename, size_bytes, content_preview}. Artifacts persist for 1 hour.'

Score history

Overall score trend

First recorded score · v2 rubric

59/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	59	2026-07-28+	v2