type: reference
status: active
timestamp: 2026-06-28
tags: [mcp, blocklist, web-tools, reference, services]
status: active
timestamp: 2026-06-28
tags: [mcp, blocklist, web-tools, reference, services]
Broken / unreliable MCP servers — skip list
MCP servers that failed during 2026-06-28 testing; skip list, re-evaluate quarterly
Broken / unreliable MCP servers — skip list
Tested 2026-06-28 during compression-research session. Each MCP listed below failed in some way. Do NOT add these to .mcp.json, ~/.claude.json, .agents/*/mcp.json, or any other config without re-testing first.
Removed from Smithery toolbox 2026-06-28
These 3 were uninstalled from the chirag127 Smithery toolbox via smithery mcp remove. They will not appear in agent tool lists anymore. Do not re-add without verifying the underlying issue is resolved.
| MCP | Why removed | Re-add condition |
|---|---|---|
exa (Exa Search) | Free-tier exhausted; needs paid API key. Violates no-card-on-file rule. | Only if user funds a paid Exa key |
ghostrouter-ghostrouter-web | Crypto micropayment paywall (USDC per call). Hard veto. | Never re-add |
pinkpixel-dev-web-scout-mcp (Web Scout / DDG) | Returns empty results consistently for valid queries that worked on other engines. | Test 5 queries return non-empty; if all pass, re-add |
Re-add via: smithery mcp add <name> after verifying.
Web search MCPs — broken / conditional (still in toolbox)
| MCP | Failure mode | Verdict | Re-eval condition |
|---|---|---|---|
mcp__chirag127__tavily_* | ”Connection not found: tavily” until ToolSearch loaded schemas; even then rate-limited on first call | Conditional — works after ToolSearch load + once-per-day rate window | Try after 24h cooldown |
mcp__chirag127__brave_brave_web_search | ”Connection not found: brave” — same ToolSearch pattern | Conditional — works after ToolSearch | Try after 24h |
Web search MCPs — working
| MCP | Notes |
|---|---|
mcp__chirag127__keenable-web-search_search_web_pages | Default. Returned 9-10 quality hits per query; filters by date; arxiv + blog mix |
mcp__chirag127__linkupplatform-linkup-mcp-server_linkup-search | Deep search worked; returns 97KB+ payloads with full content from multiple sources |
mcp__chirag127__oevortex-ddg-search_web-search | Returned 8 hits cleanly; fast |
mcp__searxng__searxng_web_search | Failed earlier with JSON-parse error but reportedly works; retry |
mcp__fetch__fetch | Reliable for single-URL markdown extraction |
mcp__chirag127__parallel-search_web_search_preview | Untested this session — try next |
mcp__chirag127__axel-belfort-web-search_web_search_query | Untested |
mcp__chirag127__apify_apify--rag-web-browser | Untested |
mcp__chirag127__ref-tools-ref-tools-mcp_ref_search_documentation | Docs search — different role, not generic web |
Security flags — working but with known vulnerabilities
| MCP | CVE / Advisory | Risk | Mitigation |
|---|---|---|---|
open-websearch (npm pkg in ~/.claude.json) | CVE-2026-42260 — SSRF | Bracketed IPv6 literals + non-resolving hostnames bypass isPrivateOrLocalHostname check in fetchWebContent tool. SSRF non-blind (response streamed to MCP caller). | Kept active per user decision 2026-06-28. Don’t use it on untrusted URLs. Patch when upstream releases fix. Check npm view open-websearch versions quarterly. |
Order to try web search (fastest to most-thorough)
mcp__chirag127__keenable-web-search_search_web_pages— defaultmcp__chirag127__oevortex-ddg-search_web-search— fast fallbackmcp__chirag127__linkupplatform-linkup-mcp-server_linkup-search(depth: deep) — when need comprehensivemcp__fetch__fetch— when you have a specific URL- Retry tavily / brave after a ToolSearch load + 24h cooldown
Other broken / abandoned MCPs noted
| MCP | Issue |
|---|---|
DesignSync | Disconnected mid-session (system-reminder confirmed) — server died |
Re-evaluation cadence
Test quarterly:
- Try the conditionally-broken ones (tavily, brave, exa) with fresh API keys
- Try pinkpixel-ddg to see if results came back
- Try searxng with a fresh JSON-parse retry
If any becomes reliable for 5+ test queries, remove from this list and add to the working set in .agents/claude/rules/never-use-native-web-tools.md.