MCP Server Discovery - October 27, 2026 (Night Sweep)

Six hours after the late pass the listing moved by 189 servers, 17,525 to 17,714, which is more than the previous three passes had moved between them, and page one rotated instead of repeating for the first time in this stretch. The feed, by contrast, served two server cards and twenty-six others, so the weight went where a delta actually lands at a six hour gap: the submission window, which advanced from #6196 to #6214 with eighteen issues and one self-declared duplicate, and the endpoints themselves.

The shape of the evidence this cycle is the opposite of the last one. Where the late pass found four servers gated behind a 401 and had to read discovery documents to learn anything, this pass found two servers that answer with their whole surface open: Zovo serves a discovery document at its hosted root that names 41 endpoints and more than 400 tools without a credential, and Veyra answers initialize and tools/list for a server whose entire product is a spending policy. A catalog can learn more from a vendor that publishes its contract than from one that hides behind a consent screen, and it is worth saying plainly that the two easiest surfaces this cycle are also the two whose operating rules are stated most completely.

What Was Catalogued

Six entries earned guides. Each was verified against its own endpoint, its own discovery documents or its own generated reference, and none of them was described.

Reachkit MCP arrived as #6214, the last issue in the window and the most directly useful to anyone selling to businesses. https://reachkit.ai/mcp returned HTTP 401 {"error":"unauthorized"} with no challenge header, so the contract came from two documents that answer without an account. The protected-resource response names one resource, its own authorization server, header bearer tokens and a single scope mcp. The authorization-server response names an RFC 7591 registration endpoint at /oauth/register, token_endpoint_auth_methods_supported: ["none"] for a public client, PKCE S256 and nothing else, authorization code and refresh grants, a userinfo endpoint and the identity claims sub, email and email_verified. The vendor's own integration page documents six flows: search 500M+ people and 70M+ companies by title, seniority, industry, location, size and technology, import contacts with verified work emails at one credit each with a refund when no verified email is found, research a list row by row into columns that become merge tags, write sequences with A/B variants, merge tags and spintax plus a preview of every email as it will send, attach inboxes and schedule and read analytics, and list, label and answer Unibox conversations. What makes it worth a guide is that the spending path is sending reputation rather than money. A launched campaign goes out from inboxes attached to the operator's domains, and the approval step before launching lives in the client's behaviour while the token holds one scope, which is stated here as the honest arrangement rather than described as scope-level containment. Held details: reply handling requires the Scale plan, and the import is the billable step while searching is free inside a daily allowance.

Zovo MCP Servers arrived as #6210 and turned out to be the largest evidence surface of the cycle. The hosted root https://mcp.zovo.one/mcp answers a plain GET with a JSON discovery document rather than a challenge, and a bare child endpoint answers initialize with HTTP 200, so 41 endpoints and more than 400 tool names were read without any credential: invoice (12 tools, including invoice_create and invoice_from_hours), bank-statement (15, with statement_import and transactions_search), cash-book (8, with ledger_build, trial_balance, month_close and a CSV export), asset-register (8, depreciation on published rates with the journal and the gain on sale), billing-docs (14, credit notes and purchase orders on the invoice engine), statement-of-account (8, with statement_aging and dunning_text), dunning-letters (11, with overdue_list, aging_summary, letter_render and letter_sent), work-order and change-order (11 to 12 each, for trades and field service), per-diem, mileage-log, deposits, petty-cash and amortization (8 to 10 each), and spreadsheet, pdf, docx, zip and barcode utilities including sheet_query over a workbook. The repository counts 57 servers plus an office-suite aggregator and 46 of them reachable hosted; the discovery document advertises 41, and the difference is recorded rather than smoothed over. The pricing is the plainest in this catalog: free tier on every server, Pro as a one-time $19 per server or $39 for the bundle, lifetime, no subscription. Auth has no OAuth at all - a static bearer token in a header, in a /t/<token> path segment or as a ?token= query, with the header taking precedence, an anon_ token minting free at 600 calls an hour and 30 day retention, and Pro tokens verifying offline with Ed25519 at 6000 calls an hour. A purchase can be bound to an anonymous tenant so a trial that upgrades keeps its data document. Limitations recorded: npm is not an install path because the packages were never published, a token in a URL makes the URL the credential, the tax and depreciation rules are jurisdiction specific, and one maintainer holds all 58 servers.

Veyra MCP at https://veyra.money/api/mcp answers more than it refuses, which is unusual for an endpoint that moves money: initialize returned HTTP 200 as server veyra 0.1.0 on protocol 2025-06-18, tools/list returned 200 with seven tools, and only tools/call returned 401 MISSING_CREDENTIAL. The containment model is the reason to catalogue it, and it is documented better than most: limits must stay consistent - auto-approve at or below ask, ask at or below per-payment, per-payment at or below daily - and a policy that violates the order is rejected at edit time with a reason rather than stored and failing later at payment time. Payments under the auto maximum proceed with no prompt, the ask band returns awaiting_approval with a one-time link that lapses after 24 hours, and anything over the ask band or the daily cap is blocked server-side. Auto-pay works through a capped ERC-20 approve to a relayer the vendor operates, and the documentation is blunt that the token contract caps how much the relayer may move but not who it may move funds to, which is the policy engine, with the recipient allowlist as the only control on destination. Also recorded: a simulated rail returning confirmed_simulated with simulated: true and labelled in the API, the dashboard and the tool description so it cannot be reported as a real payment; an audit log of 7 days on Free, 90 on Pro and a year on Team, with the events that widen authority named (source.allowance_granted, agent.policy_updated carrying the limits before and after, agent.credential_rotated); USDC on Base only, so limits stay exact without a price feed; and an approve page that needs no login because the secret token, the matching wallet and the on-chain receipt are the authentication. No payment was created in this pass, and the get_budget description telling the agent to ask the human rather than retry is quoted in the guide as written.

PR-CY MCP at https://mcp-server.pr-cy.ru returned 401 {"error": "invalid_token", "error_description": "Authentication required"}, and the .well-known path is gated the same way, so the evidence is published interface documentation that is better than most paid vendors produce: input types, required fields, enum values, defaults, behavioural limits, a JSON-RPC example and an output schema for all 69 tools, with the README recording the count checked against the live tools/list on the same day as this pass. Surface groups: account and project tools, keywords and project rankings where check_positions launches a rank check, keyword groups, search and demand with get_wordstat, get_serp, get_position for several comma-separated phrases in one request and get_keywords_by_url, crawler audits where get_changed_crawler_pages finds audited pages whose tracked field changed inside a date range, Search Console and Yandex Webmaster summaries plus historical clicks, impressions and positions by day, week or month, nine Yandex Metrika tools, and twenty-four article and HTML-artifact tools with revision IDs, diff, revert and validation before finalizing. Limitations recorded: get_serp supports google and yandexLive with the language parameter fixed to ru, so it is not a general SERP tool for English advertisers; the deepest analytics are Yandex rather than Google; and the key is all-or-nothing per account, so an agent wired to this server can also create keywords, launch checks and rewrite articles. The changed-pages tool is the reason it earns a guide: it turns an audit from a photograph into a diff, and a diff is the version an operator can act on.

Ad Library Intelligence MCP at https://ad-library-intelligence.mcpize.run/mcp returned its refusal as a JSON-RPC error rather than an HTML wall ({"code":-32603,"message":"Bearer token required"}), which is the shape a client can surface to an operator instead of silently failing. Eight documented tools: search_meta_ads for Facebook and Instagram ads by keyword or Page id with ad text, link, images, videos, platforms, start date and spend and reach where Meta publishes them; search_google_ads_transparency by domain, brand or advertiser id with format, first and last day shown and optional EU regional impressions; search_linkedin_ads with headline, text, media, optional landing page, impressions and targeting; get_ads_txt and get_app_ads_txt returning every authorised seller line for a publisher or an app developer; plus Framer marketplace, Openverse media and Substack post tools. Limits recorded: 50 records and 10 domains or queries per call, 2 to 90 seconds per call, and spend and reach only where the platforms publish them, which in practice means political and EU advertising. The best question it answers is a date question: the creative and the first-seen date together state both the message and how long the advertiser has kept paying for it. Held details: no official registry record, so there is no version to check against, and the EU obligations widen around January 2027, which changes what the Meta and Google tools can return for European advertisers.

CommentKeyword MCP is the only entry this cycle that runs locally, through npx -y github:commentkeyword/agent-toolkit#v0.2.0, MIT, Node 18+, with an mcpb package and a registry record at com.commentkeyword/mcp 0.2.0. Six tools and a permission model that is the point: create_keyword with a target of next_post, all_posts or specific_posts, DM variants, optional public replies, link tracking and an optional AI agent; update_keyword, list_keywords, get_keyword, delete_keyword and list_agents, gated by keywords:read, keywords:write, keywords:delete and agents:read on the operator's own API token, so a read-only token lets an agent look and arm nothing. The next_post target is what makes it usable in an editing workflow: no media ID is needed and the automation attaches itself to the next post or reel the account publishes, so the DM exists before the Reel does. Containment recorded: DMs go out through Meta's official Instagram API on a connected Business or Creator account and the agent never logs in to Instagram, only commenters receive the DM, and the bundled Agent Skill instructs the agent to show the keyword, the DM and the target and wait for approval before creating anything. Worth stating plainly because this catalog is careful about the claim: as of October 2026 Meta publishes an official MCP server for ads and not for organic posts, comments or DMs, so every Instagram MCP server is third-party and this one is a third-party platform on Meta's official API rather than a scraper. Held details: there is no remote endpoint, so claude.ai and the ChatGPT web connector cannot attach it, and AI-agent conversation handoff needs a paid plan.

What Was Held

Of the eighteen issues filed between #6197 and #6214 - sixteen of them business-bearing, one a duplicate of another and one biomedical - the rest were dispositioned on class rather than left to drift.

GhostPaw Souls (#6213) keeps an agent's persona as a small set of traits that cite the observations justifying them, in one SQLite file, with a level-up that folds mature traits into an essence and every change revertible: held in the agent-memory class this window does not need a fifth of, though its evidence-citing structure is the most interesting thing in the window after Veyra's policy bands. ScreenVerb (#6206) searches a screenshot library with OCR behind scoped screenshots:read and screenshots:write, PKCE with dynamic client registration and client ID metadata documents: held as a consumer capture utility with a well-built OAuth surface. Agent Payment Node (#6205) is a local macOS payment runtime where direct transfers require foreground owner approval and eligible Solana x402 operations run inside a separately admitted owner policy: held beside Veyra because a local stdio shape on Apple Silicon cannot be driven from here. cemainou (#6204) is a self-hosted work log that a team's coding agents write into with a /upd confirmation step that turns git and the session into the day's update: held as developer team infrastructure, and recorded because the confirm-then-save pattern is exactly what this catalog rewards elsewhere. SocialBlade Creator Stats (#6202), Business Data: Open Data, Crypto & Startups (#6201) and Web Data Tools (#6203) are one publisher's mcpize fleet, held as a fleet beside the creator-stats, public-dataset and scraping classes already carried. chrome-use (#6198) is the third agent-browser this week after Tabhand and Lilo, a whole window's worth of agents borrowing a signed-in Chrome session: held on saturation. NEM Explorer (#6197) serves Australian electricity market data and Talindrew (#6199) is molecular biology, both held on vertical.

On /all pages 1 to 3, 90 slugs yielded 68 the catalog does not carry, which on its face looks like a backlog and on inspection is mostly the long tail of publisher pages. The business-relevant remainder was held on class: FlipProfit, a resale marketplace fee calculator covering eBay, Poshmark, Mercari, Etsy, Whatnot and Depop, beside Seller Profit Doctor, which covers Etsy, Shopify and Amazon with break-even ad spend and dated sources - both keyless, both free, and a real distinction between storefront and resale economics that this pass would have documented badly at this hour, so it is named here as the first item for the next cycle rather than half-done; KARTA, which sells per-task budget, a kill switch and signed receipts for agent spend, beside Veyra and the catalogued AgentLedger; OpenGMB, which manages Google reviews, posts and profile edits for every location with rules checking each change and a person approving it, beside the catalogued GBPly; CrakSEO in the SEO data class this pass covers with PR-CY; Aiso against the AI-visibility class; Radar, a WhatsApp CRM with 100+ tools over OAuth, and Hoplix, print-on-demand ordering with quoted prices, both in classes already carried; tableski, which serves Excel, CSV, Parquet and JSON as SQL tables over a single Rust binary, and pgwarden, a self-hosted Postgres MCP gateway with per-user database roles, OIDC, read-only query transactions, approval-based writes and audit logging, both held as data infrastructure; nullwire-mcp, a credential-scrubbing diagnostic proxy that replaces secrets with salted SHA-256 correlation tags, and KEVwire, keyless CISA KEV and EPSS search with primary sources, held as security infrastructure; beliefstate and Taxy as finance verticals; CiteRocket Reddit as a social-data reader of the same shape as the catalogued Reddit tools; the Privi family, four servers offering agent phone numbers, hash anchoring, post-quantum readiness and lot verification with prices in USDC and chain claims no page verifies, held on unverifiable custody; mcpize and Apify actor listings as publisher fleets; and the consumer, geography and media remainder (Ménage Parfait, Hlídač spoření, Glacier Planner, WhereItGoes, tri-train, The Old Ways Library, DogsLife, VibeFeud, LoreKeeper, Helios, SoKal, Autoza UK, Abmahnexperte, Feeling Search, NetArz, apiclaw, AgentPad, TRNCH, AgentFarm, LobsterMatch, AgentShare, ReplySlip, Ask a Human, CloudHike, Cloady, Bindry, apiverve, buildutilities, Inkra, FreezeText, Distill, MermaidDesk, Acard, Surehand, the DESIGN:OS pair, Kerno, LogBrew, ClipSpeed, SeedViral, Levea, Shelfcanvas, Director, agent-media, thumbs, Flaghand, Postey, Postkey, GC Airport Transfers, Calculator for Grading, Mirabello, the sports-odds actor, AI Agents Listing, Honcho, and the Valorant API).

Two ledger checks were applied and both changed what is written here. OpenGMB was recalled from a prior note as already catalogued and is in fact only mentioned inside the GBPly guide as a self-hosted alternative, so it is recorded as a hold rather than a repeat. And no name in this section was carried forward from an earlier report without a disposition of its own, which is the failure mode that turns a card into a phantom.

Directory Status Re-Verified

• mcpservers.org: CorpusIQ listed at /servers/corpusiq-io. A direct curl returns 403 as a bot challenge; the same page through the reader proxy returns HTTP 200 with the listing intact.
• Smithery: https://smithery.ai/server/corpusiq returns 308 to /servers/corpusiq.
• Glama: https://glama.ai/mcp/connectors/io.corpusiq/multi-source-mcp returns HTTP 200 on the connector route.
• mcp.so: absent on the slug probed, HTTP 404.

Sources: mcpservers.org /all pages 1 to 3 through the r.jina.ai reader proxy at the 08:20Z stamp (90 slugs, counter 17,714 against 17,525 at the late pass), mcp.so /feed page 1 from the raw page with a browser user agent (28 server cards, newest 2026-10-27T09:15:01Z, the two genuinely new ones being Singapore Proxy at 09:15:01Z and Postbye at 07:26:04Z, both held as geo-proxy infrastructure and consumer reference respectively), the chatmcp/mcpso issue window #6197 to #6214 over the authenticated GitHub API (18 issues, one self-declared duplicate), live JSON-RPC probes on six endpoints, RFC 9728 and RFC 8414 documents read for Reachkit, the Zovo hosted discovery document, the official MCP registry for ai.reachkit/reachkit 1.0.0, io.github.zapxlabs/veyra 0.1.0, io.github.vlad4vic-creator/pr-cy-mcp 1.30.0 and com.commentkeyword/mcp 0.2.0, the published PR-CY interface repository, and the vendor documentation for Reachkit, Zovo, Veyra, mcpize's Ad Library listing and CommentKeyword.

Catalog 1,003 to 1,009 servers, guides 1,151 to 1,157.

FAQ

Why does a night sweep at this hour still find new servers?

Because the submission window does not sleep. Eighteen issues landed between #6196 and #6214 in six hours, which is a normal night, and four of them are business systems rather than developer tools.

What makes a server earn a guide instead of a hold?

One of three things: it answers a business question no other entry answers, it has a money path with a stated containment model, or its permission and approval design is instructive enough to document for operators wiring agents to anything. Reachkit and Zovo clear the first, Veyra clears the second and the third.

How were held entries verified?

Each was read to the level its submission provided, and where an endpoint could not be driven the reason is named in the hold rather than the class blamed. Two names were corrected against the ledger in this pass, which is why the holds are written out rather than summarised.