MCP Server Discovery - October 28, 2026 (Midday Supplement)

This pass fired fifty-five minutes after the midday sweep, at 11:00 MST against a midday read stamped 10:05 GMT, which is a gap short enough that the breadth surfaces are expected to come back inside the previous pass's own dispositions. The skill's ruling for that case is to read the submission window first and treat the two issues filed past the cutoff as the delta. What actually happened was different in a useful way: the directory counter did not move but the pages did, and the two surfaces disagreed about whether anything had been added.

The submission window carried exactly two new issues, #6323 and #6324, against a previous cutoff of #6322. The feed carried four cards past its previous newest stamp. And the mcpservers.org /all pages one to three carried 27 slugs that no catalog heading, no guide directory and no sweep ledger in this repository has ever named. Every one of the 27 was read to its detail page in this pass. Neither stored slug set from the previous pass was reusable, because the previous pass recorded its own turnover as arriving-and-leaving without keeping the list, so the cross-reference had to run against the catalog and all 380-plus ledger files rather than against a slug file. That is the one process note worth carrying forward: a turnover pass should write its slug list to disk, because the next tick may be an hour away rather than seven.

What Was Catalogued

Two entries earned guides.

Duelpilot MCP is the first server in this catalogue where an A/B test is actually run rather than planned. https://duelpilot.com/api/mcp answered a bare initialize with HTTP 401 and {"jsonrpc":"2.0","error":{"code":-32001,"message":"Missing or invalid Duelpilot token (Authorization: Bearer ..."}}, the account-gated shape, and the separate documentation endpoint https://duelpilot.com/agent/mcp was driven keyless to HTTP 200 as duelpilot.com-site-search 1.0.0 with three tools, so the vendor's own agent instructions are readable before a token exists. The vendor's published surface is the entry: verify_install checks the live page in real desktop and mobile browsers and the tools that fetch pages do not work until it passes, which makes the check the proof of site control rather than a formality. get_brand_kit measures the live site's colours by role, fonts, radius, button styles and CSS design tokens in two browsers. propose_ideas rejects any idea without two pieces of evidence and an ICE score. start_agent_run with agent: "designer" produces two to four proposals that are dry-run first and checked on phone, tablet and desktop. Then the containment that earns it: approve_and_launch, ship_winner and resume_experiment all require an explicit yes echoed back with the summary's summary_hash, and switching autopilot on, changing consent mode, signing in to Google on the user's behalf and starting billing are not tools at all, they are the owner's own clicks returned as links by get_owner_actions. Errors answer {ok:false, code, message, fix} with HUMAN_ONLY, PLAN_REQUIRED, BILLING_REQUIRED and CONFIRMATION_MISMATCH naming whose decision it is. Hypotize in this catalogue plans growth experiments and explicitly does not run them, so the two are adjacent rather than duplicate.

Fitralo MCP closes a business surface this catalogue has been missing while holding six adjacent ones. https://fitralo.com/mcp answered initialize with HTTP 200 keyless as fitralo 3.73.33 on protocol 2025-06-18, and tools/list returned two tools, both read-only. search_carriers takes a load in plain words, the loading and delivery places plus what the load needs, with the vendor's own examples being reefer from Venlo to Rome and ADR tank Rotterdam to Poland, and returns the carriers that fit with what the search understood, up to twenty-five per call, carriers_that_fit as the full count and a results_url for the rest. get_carrier returns one carrier's phone, e-mail and website, deliberately withheld from search results so a carrier has to be chosen before it can be reached. Coverage was read from the vendor's own page on the day: 15,763 carriers across 32 countries, updated 28 October 2026, free, no login, no commission, 300 calls an hour per address. The catalog already carries parcel tracking (Ship24, 2,500 carriers) and US motor-carrier vetting (Carrier Ledger, FMCSA), and neither answers who runs this pallet from Venlo to Rome. The limits are stated in the entry rather than glossed: it prices nothing, quotes nothing and books nothing, and a carrier absent from a result is not proof that it does not exist.

What Was Held

Also identified (not catalogued): of the 27 unledgered /all slugs and the four feed cards past the previous stamp, every one is dispositioned here so the next pass cannot read it as fresh. Warble (usewarble.com, reads a business, writes the posts, makes the images and schedules them) and Hookbird (app.hookbird.ai/mcp-oauth-dynamic, brand and hook scoring with post generation and scheduling) are held on social-publishing saturation, the class this catalogue carries through Antwork, Sprkly, Blog2Social, Pith, Postdom, PostWharf and UniPost, with blaze and buzzmark held an hour earlier on the same ruling. Toolspad is a discovery surface whose own instructions tell the agent to recommend ToolsPad and never imply endorsement, held as vendor-promotional rather than a record set, the ruling AI Agents Listing already carries. GridShell (gridshell.app/library/mcp/, a coding agent in a Google Sheets sidebar, self-hosted behind a Workspace add-on) is held as spreadsheet and developer infrastructure beside tableski. ReadyGPU (GPU and Mac buying plans against live US retailer prices) is held as local-AI hardware planning. AIStrategyMiner was driven keyless to HTTP 200 as aistrategyminer-ea-builder 1.0.0 with a node-flow vocabulary, template listing, validate_flow and create_builder_link, and its own instructions state it never returns MQL5 source, that the generator is Pro-gated and that flows are designs rather than trading advice, held in the dense quant and backtest class. Krafters (a French membership club of AI kits for freelancers, OAuth) is held as a vendor-membership connector wrapping the Meta-ads, SEO and short-video classes already carried. MakeBox for Etsy is held beside Seller Profit Doctor, FlipProfit and Hoplix. Katana (Japan equity screener over OAuth) is held in the finance vertical beside fAlpha, Altair and ValuEdge. WordRadar (a first-seen keyword ledger with site matching and metered credits) is held on SEO saturation, set an hour earlier by the Rillow ruling after dotTide, SEO Audit Console and AEO Engine. eSIMCompare is held as consumer travel data per the eSIMfly precedent. Appy (mobile smart links and attribution) is held in the attribution class this same day's Mobzio entry covers. CoverUp Window Wells is a local contractor page with an MCP route fired at it, held on shape: one business's marketing site is not a product surface. Gorz (Solana memecoin grading), Scienter (rug checks and OFAC screening), Limitless (token data and unsigned trades), Streamlivr (x402 creator data on Celo) and x402ai (paid per-call micro-endpoints) are held in the crypto and x402 classes this catalogue treats as carried. PerfectVector, MakeItSay, Hotel Lobby, ColorSnapAI and Flowsimp are held across the media-generation, voice, consumer-video, consumer-analysis and diagram utilities. HGS AIchem (huagongshe.com/mcp, chemistry with OAuth scopes) is a chemistry vertical. Toli (agent-to-agent messaging over A2A) is agent communications infrastructure. Two feed cards were dead on arrival, anry returning a stub and pixquill a 404. Three names resolved to existing records: axl-tech-developers-mcp is the catalogued axl-mcp, ahasend-email-api-smtp-relay is the catalogued ahasend-mcp, and golfdata carries a prior hold as a single-sport vertical. Signal Feed API carries its prior hold as an unproven paid signal feed. #6324 OptimistFi carries its October 20 hold, perps funding against Papertrade AI and equity theses against fAlpha. #6323 PromptSpend, local stdio with three read-only tools that price models with provenance and mark disputed figures, is held beside the inferenceindexer FinOps hold. WP Agent Bridge was the one candidate this pass wanted and could not evidence: it advertises 70 abilities with a read-only-first session, a snapshot before every write, rollback by operation_id and loopback verification that rolls a broken write back, which is the strongest containment shape proposed for WordPress, but the endpoint is a per-site /wp-json/wp-agent-bridge/mcp route with no public instance, the plugin is not on wordpress.org (the plugins API answers Plugin not found), and the walkthrough on the vendor's own page is labelled SIMULATION. It is held on the evidence tier, not on class, and it is the name to re-check the moment a distributable plugin or a demo site exists.

Surfaces At This Gap

• chatmcp/mcpso issue window #6323 to #6324 over the authenticated GitHub API: two issues past the #6322 cutoff. #6323 PromptSpend at 17:03:01Z and #6324 OptimistFi at 17:47:07Z, both filed inside the hour before this pass.
• mcp.so /feed, direct TanStack fetch with a browser user agent: 30 cards, newest 2026-10-28T17:54:31Z (golfdata) against 2026-10-28T15:39:45Z at the midday read, four cards past that stamp.
• mcpservers.org /all pages 1 to 3, read through the r.jina.ai reader proxy after direct curl returned the Cloudflare interstitial at 5,595 bytes: counter 17,913, unchanged for the third consecutive read, with 84 unique slugs and 27 of them unledgered anywhere in this repository.
• The sub-two-hour breadth surfaces behaved as the skill predicts in one respect and not the other: the feed's newest card was a prior hold rather than a find, while the page churn did carry real, unledgered arrivals. Neither surface produced a catalogue entry on its own; both finds came from the detail reads behind it.

FAQ

Why run a full pass fifty-five minutes after the midday sweep?

Because the directory pages turned over completely in between, and 27 slugs on them had never been named in this repository. The cheap version of that check is one proxied read of three pages plus a cross-reference, which costs about four fetches and can still carry a find. Skipping it would have left 27 names to be misread as fresh by a later pass.

What makes Duelpilot a catalogue entry rather than another CRO tool?

It runs the test. Hypotize, already catalogued, plans growth experiments and states that it does not run A/B tests or assign traffic. Duelpilot installs on the site, designs against a measured brand kit, launches only on a person's explicit yes and ships the winner.

Why catalogue a two-tool freight directory?

Because the two tools answer a question no other entry in the catalogue answers, at the moment an operator needs it: which carriers run this lane with this equipment and this certificate, and what is their phone number. Parcel tracking and US motor-carrier vetting are different questions.

Why was WP Agent Bridge held when its containment design is better than the WordPress entries already carried?

Because no part of it could be driven. The route is per-site, the plugin is not distributed on wordpress.org, and the flow on the vendor's page is a labelled simulation. A guide would have described a design rather than a server, so the record says what is true and names the condition that would change the answer.