MCP Server Discovery - October 20, 2026 (Night Sweep)

Night delta about five hours after the October 19 evening capture. The mcp.so /feed page 1 carried four cards posted inside the window (iGaming Tools at 36 minutes, a Patsnap patent and literature search at one hour, Coinrule at two hours, twocents at five) and the rest carried their evening dispositions. mcpservers.org /all pages 1 to 3 served 90 cards, and the useful signal this time was positional: everything stacked above the catalogued drift card was added since that capture, which separated genuinely new listings from the ones held a few hours earlier.

Four entries were catalogued, and the difference from the last several sweeps is that every one of them was verified by running it rather than by reading its listing. Two were run end to end on this machine (a Python proxy and a Node server), one was probed to its authentication boundary and its OAuth metadata read, and one was checked against its package registries and repository state.

New Servers Catalogued

Server	Category	What it does
Azquo MCP	Data and Analytics	Official hosted server at https://data.azquo.ai/mcp for Azquo, a finance-owned store where agents keep what they learn. Names live in hierarchies rather than rows, so a new dimension is a new lens instead of a migration, and every individual figure carries per-cell provenance back to its source file, the user and the date. Several agents and people work in one database with every write attributed to both the agent and the person behind it; conventions an agent agrees are stored in the database for the next agent. Capabilities, in the vendor's wording because no tools/list is reachable without a credential: create databases, import raw files, design import templates, query in AQL, browse and extend hierarchies, build and run reports, audit any figure through every transformation, rehearse a change on a copy, see a name's history and restore a deleted one. Verified: initialize returned HTTP 401 Unauthorized : no key with WWW-Authenticate: Bearer resource_metadata="https://data.azquo.ai/.well-known/oauth-protected-resource", and that document returned {"resource":"https://data.azquo.ai/mcp","authorization_servers":["https://data.azquo.ai"],"bearer_methods_supported":["header"],"scopes_supported":["mcp"],"resource_name":"Azquo"}. Evaluation access on request, price unpublished.
New Term Radar MCP	SEO	Local stdio Python server (pip install "new-term-radar[mcp]", command new-term-radar-mcp) that reads eight signals for names that are starting to surge before keyword tools list them: new words on a list of X accounts, Hugging Face trending, GitHub breakouts under 14 days old at 200+ stars a day, word surges across watched sitemaps, new domains ranking on 20+ keywords, Reddit asks, TrustMRR verified revenue and your own Search Console first impressions. Three tools - recent_leads free and instant, keyword_table at about $0.05 a name plus $0.09 a call, run_radar paid and measured in minutes - and a keyword table that sizes the generic demand behind each new name and decides prioritize, watch or skip. PyPI 0.4.0, MIT, Python 3.10+; repository created October 12, 2026, pushed October 18, 1,218 stars, 436 forks. The worked example is checkable: Niko1221/Strata, created 2026-09-24, 7,700 stars nine days later, searched from 9/26 while no keyword tool listed it. About $15 a month on the example config.
GLP Audit Proxy MCP	Compliance	Local stdio proxy (pip install mcp-glp-audit-proxy, 0.1.0, MIT, Python 3.11 to 3.13, alpha) that sits between an MCP host and its upstream servers and writes every mediated call, including denials and failures, as one RFC 8785 canonical-JSON line in an append-only SHA-256 chain with a .head sidecar. Role separation is enforced - operator calls upstream tools, approver approves, qa reads the trail, admin changes policy and cannot call upstream - and auditing cannot be switched off: audit_enabled: false is denied and logged. Installed and run here: eleven entries for one session, glp_verify_chain returned chain ok entries=11 head=6669451c..., the verify CLI returned chain ok entries=12 shown=12 with a per-entry listing, and the tool surface differed by role exactly as documented (operator with no upstreams: tools: []; qa: glp_verify_chain, glp_search_audit, glp_export_audit). Coverage is mapped clause by clause to OECD GLP No. 17, FDA 21 CFR Part 11 11.10(e), EU GMP Annex 11, MHRA and PIC/S PI 041-1, WHO TRS 1033 and EU AI Act Articles 12 and 19. Community approval is not an electronic signature; WORM storage and signing are Pro.
Creavores AEO Audit MCP	SEO	Local stdio Node server (npx -y creavores-aeo-mcp, npm 1.1.0, MIT, Les Creavores) that scores one page for AI-search readiness: crawlability 30, structure for AI extraction 35, E-E-A-T trust 35, plus separate llms.txt and AI-bot access checks. Eight tools verified by tools/list, including check_agentic_readiness (aligned with Lighthouse's Agentic Browsing category) and simulate_ai_citation (which passage an engine would extract, and why). Two live calls on a public marketing page returned score: 100, verdict: "excellent", issueCount: 0, seven AI crawlers checked and none blocked, and an llms.txt compliance score of 100 with 13 sections and 37 links parsed. Filed as two stars rather than a gap-closer: Ranki already carries AEO scoring in this catalog and Prognosite carries SEO and AEO intelligence, so this is packaging and two angles at zero cost, no key and no rate limit.

Sources

• mcp.so /feed page 1 via the extraction path that returned all 30 cards with submit ages: four posted inside the window (iGaming Tools 36 minutes, Patsnap patent and literature search 1 hour, Coinrule 2 hours, twocents 5 hours), then Reelier 16 hours, TaskerArmy and Capital.com 18 hours, and the previous day's cohort behind them
• mcpservers.org /all pages 1, 2 and 3 via the r.jina.ai reader proxy, 90 cards parsed, position relative to the catalogued drift card used to separate new listings from those held in the evening sweep
• Vendor documentation for every catalogued entry: the Azquo listing and its OAuth protected-resource metadata, the New Term Radar README on main, the GLP proxy README, PyPI metadata and config.example.json on main, and the Creavores npm metadata plus the live server itself
• Live probes: HTTP initialize against https://data.azquo.ai/mcp (401 with the authentication challenge and resource metadata), and GET https://data.azquo.ai/.well-known/oauth-protected-resource (200, quoted above)
• Live probes against the GLP proxy, installed into a clean virtual environment and run twice: initialize and tools/list as an operator principal with no upstreams (tools: []), then as qa (three trail tools), then glp_verify_chain (chain ok entries=11 head=6669451c...), then the mcp-glp-audit-proxy-verify CLI with --print-head (chain ok entries=12 shown=12 plus the full listing)
• Live probes against Creavores: tools/list on protocol 2025-06-18 (serverInfo creavores-aeo-mcp 1.1.0, eight tools with schemas), audit_aeo and check_llms_txt tool calls returning the figures quoted above
• Registry and repository metadata: npm creavores-aeo-mcp (versions 1.0.0 to 1.1.0, created 2026-07-10, modified 2026-10-12), PyPI new-term-radar 0.4.0 and mcp-glp-audit-proxy 0.1.0, GitHub API for zhuyansen/new-term-radar and antonio-castellon/mcp-glp-audit-proxy
• Not run this cycle, and not claimed: the twelve-sitemap canonical refresh, the chatmcp/mcpso issue window, the Glama, Smithery and PulseMCP listing probes, and mcpservers.org /all page 4

Also Identified (Not Catalogued)

mcp.so feed window (4 cards, all held): iGaming Tools (structured iGaming reference data: slot specs and RTP variants, provider catalogues, regulators, industry news and jobs - one vertical, reference rather than operations); Coinrule (official MCP for a crypto trading cloud account - trading); Patsnap patent-literature-search-mcp (natural-language search across a commercial patent and literature database with Markdown bibliographic output - one vendor's IP index, the closest call of the four and worth revisiting if an IP class ever opens); twocents (twocents.page/mcp with share_page and get_feedback - a genuinely novel data source, human feedback on a shared page, but two tools and no published pricing or documentation to verify).

Held from the feed, previously carried: TaskerArmy Agent is already catalogued as taskerarmy-agent-mcp. MailStream (physical postcards and letters, address verification, mailing lists, campaign tracking) held as a print-and-post vendor surface with no verifiable tool list. Sugra API (1,500+ endpoints across 36 data domains from 160+ primary sources) held against the aggregator class. endoflife.ai (485 products, EOL dates, support status and a 0-100 risk score, free) held as IT asset hygiene. Velarion (executive compensation and governance from SEC proxy filings, nine tools) held as a single-purpose finance vertical. Vibgrate (51 tools over drift scores, CVEs and EOL runtimes) held against security hygiene. Reelier (record a tool-call workflow, replay at zero tokens, diff runs) held as agent dev infrastructure.

mcpservers.org /all pages 1 to 3 (90 cards). The closest calls this cycle, all held:

• AdsCove - hosted server over Shopify, Meta Ads, Google Ads and TikTok Ads, 323 tools and 14 workflows, writes approval-gated with keys server-side. Held because the ads class is saturated in this catalog and CorpusIQ ships its own cross-source ads connector. The approval-gated write model is the right one and is worth pointing at when the product team next revisits write scopes.
• Kirat - a multi-model AI council for Claude Desktop and Claude Code: independent positions from ChatGPT and Gemini, a vote with the dissent kept, and code review where each finding is checked by a second model family, bring your own keys, paid one-time. Held as agent infrastructure.
• Velocity (AI-centric project management), Pinafly (a UAT link a coding agent reads, fixes against, and a person marks verified), Elementor MCP Createur, GanttFather, Product Page Extractor, Castov, Founden, rostrum, Chargezen SmartyTap, mercadopleis, check.tattoo, RedirX, AgentLeak, Animation Machine, FlowX and SiteDoctor carry their October 19 evening disposition and were not re-litigated.
• Brik, attest, Agent Connectors, Synmerco, Jobven, WellFred, Chinaready, NobetciEczane, Verifai, MotionVault, EasyGerber, Latchkey, YouTubeTranscriptAPI, nohumans.directory, Cotlet, Rush, Cordex, Fantopy, Pirate Skills, XposedOrNot, Runlane, FinTarget, Arcdeck, Remembite, Latent Spa, RugShield, FeatQ, Rankly, InkRoom, Porch, Sato Hub, CaptchaNo, the DuckDuckGo reference card, ShortSync, ActAs, FramerToNextJS, People Remover, Shotkiln, SocialDataX and ConnectMachine were read and grouped by class; none closed a gap.

The disposition rule that keeps earning its keep: if the tool reads no customer data and answers only questions about itself, it is a sales surface and it is held. Six of the new cards in this capture were built exactly that way - a vendor's own overview, pricing and calculator behind an MCP endpoint. It is a legitimate acquisition channel and a poor reason to catalogue something.

CorpusIQ Listing State

Not probed this cycle. The listing state carries the October 19 evening capture: mcpservers.org LISTED at https://mcpservers.org/servers/corpusiq-io (proven through the reader proxy, since a direct origin fetch returns 403 and must never be read as a delisting), mcp.so NOT LISTED at /servers/corpusiq-io with quarantine holding and no resubmit.

Result

Catalog moved from 942 servers (+828 guides) to 946 servers (+832 guides). Four guides written, all four passing the local documentation gate (frontmatter, title length, description length, no em or en dashes, FAQ headings). Three of the four were executed rather than read: the GLP proxy installed into a clean virtual environment, run under two roles and its chain verified; the Creavores server called twice against a live page; Azquo probed to its authentication boundary with its OAuth metadata read. Held honestly: the sitemap refresh, the issue window, the three listing probes and /all page 4 did not run, and Azquo's tool surface remains unverified because no tools/list is reachable without a credential.