MCP Server Discovery - October 31, 2026 (Night Sweep)¶
This pass fired at 21:00 MST, four hours after the afternoon read. The mcp.so homepage recentServers block returned 132 parsed cards against 140 at the afternoon read, which is the same block-count variance recorded twice this week and not a change at the source. The newest createdAt was the reliable reading, as usual: 2026-11-01T00:15:52Z against 2026-10-31T17:06:19Z at the previous pass, so six cards sat above the afternoon stamp.
The submission window moved much further. The afternoon pass closed the window at #6511; tonight it ran to #6530, which is eighteen submissions between 19:06Z and 03:53Z, every one read with its body rather than triaged from a title.
Three shipped. All three answered a live endpoint, and two of them answered without any credential at all, which is the test this pass applies before a guide.
New: the accessibility scan nobody runs until a form asks¶
Lazynext Accessibility Checker MCP is remote Streamable HTTP at https://checker.lazynext.com/mcp, keyless, initialize answering HTTP 200 in 0.19 seconds with serverInfo accessibility-checker at 1.0.0 and tools/list returning all four tools. list_rules returned 75 WCAG criteria, each tagged static or rendered.
Two live scans were made on our own homepage, and the result is the argument for the entry. Mobile viewport: score 46 on score_model weighted-v1, 28 issues across 12 distinct criteria, section 508 non-conforming against WCAG 2.0 AA with 5 clauses implicated (302.1, 302.2, 302.3, 302.7, 302.9), a benchmark at the 54th percentile of 26 sites, and a persisted report at checker.lazynext.com/report/c036dcc1-4e6. Desktop viewport: the first call timed out at a 60-second client limit and the retry returned in 61.1 seconds with score 46 and 29 issues.
The findings are specific in a way most audit tools are not. Two wcag-2.5.3 failures where an aria-label replaces the visible text instead of containing it, so voice control cannot address the control by what it says. One wcag-1.4.10 on a <div> with a fixed 1000px width that cannot reflow at 320px. Seven wcag-1.1.1 findings that are all filename alt text: "Shopify", "HubSpot", "eBay", "HighLevel", "Odoo", "ActiveCampaign". Those are content edits, not engineering projects, which is how an operator without an accessibility specialist gets from a procurement form to a work list.
Gap recorded. The vendor site lists API, MCP and A2A as Pro features at $9 per month, with the free tier at 3 rendered scans a day, and yet the MCP endpoint accepted no key and labelled both scans plan: "free". Whether that is a launch window or a plan gate that is documented but not enforced is not stated where this pass could read, so it is recorded as an unresolved discrepancy. scan_html, get_report and the site crawl were not exercised, the benchmark cohort is unpublished, and the repository is source available rather than open source.
New: what a creator post costs, before anyone answers an email¶
distribution.tech MCP is remote Streamable HTTP with a guest tier at https://app.distribution.tech/api/mcp/guest that needs no key: initialize HTTP 200 in 0.34 seconds, serverInfo distribution at 1.37.0, and six tools with full schemas.
The guest surface returns counts rather than names, and the counts are the reason to read it. inventory at 04:02:35Z reported 1,005 approved X creators, 86 bookable now and 255 posts carrying a current price. resolve_audience on a described profile (a B2B AI data layer, ecommerce and SaaS operators, US) resolved to three segments and narrowed that to 792 creators, 63 bookable and 187 priced posts, thin: false. The 213-creator gap between the whole network and one audience is exactly the number a founder cannot get from a DM.
guest_status states its own limits: 60 calls per 10 minutes per client address and 3 plan_quote_guest calls per 7 days, which makes this a coverage tool rather than a price research tool unless you are careful with your three quotes.
The buying half is documented at /docs/agents as 22 tools, and the design is what earns the guide: a key carries a spend mandate, signing a plan produces a payment link inside that mandate, brands are reviewed before their agent can buy, creator names and two alternates arrive before signing, swaps are allowed before signing, and every placement returns evidence. An agent that can propose a plan and stop at an approval is a procurement process a founder can audit.
Gap recorded. plan_quote_guest failed with unavailable: The network could not be read just now. Try again. so no all-in price was obtained, and that failure consumed one of the three weekly guest quotes. Nothing was purchased and no brand approval was sought, so create_campaign, approve_plan, evidence and the mandate enforcement are read rather than driven, and the creator side of the payout split was not read at all.
New: pay data that reproduces its own published number¶
GWIX MCP is remote Streamable HTTP at https://api.gwix.ai/mcp, and the live server answered a keyless initialize with HTTP 401, JSON-RPC -32001, Authentication required: sign in with OAuth or send a GWIX API key as a Bearer token. A probe on /v1/mcp returned a different 401 shape, an RFC 7807 problem body with a request_id, which tells you the public API and the MCP server sit on different routers and both are closed.
The vendor publishes a keyless sample at https://api.gwix.ai/mcp/sample, and that is where this pass measured. initialize HTTP 200, tools/list returning 27 tools in 37,580 bytes. get_data_status for the US reported release: "2026-W41", last_complete_week: "2026-04-12", and a sample that covers the US, UK and Germany at 6 months of lag with history from 2024-10-14, market level only, no company or industry splits.
Then the verification that matters. get_pay for SOC 47-2111 in metro mtr_midland_tx for 2026-Q1, the exact example GWIX publishes in its own free sample, returned 74 ads with pay and percentiles of 38,215 / 48,368 / 66,781 / 76,526 / 94,694 USD per year. The median matches the vendor's published figure to the dollar. A data product that reproduces its own advertised sample on a third-party call is a data product whose other rows can be trusted to be consistent, which is a stronger signal than any marketing page.
The metadata is where the argument lives: 13-week quarters of whole Monday to Sunday weeks ending on the last Sunday on or before the quarter's last day; each country's currency converted at the period average with fx_rate returned per row; occupation codes crosswalked to US SOC 2018 with the rule that each SOC is a separate row set and never pooled. That is the machinery that makes a pay figure checkable by somebody who was not in the room.
Gap recorded. job_offer_check, the tool closest to a decision, returned a CloudFront 504 after 30.07 seconds while the row tools answered in 0.26 seconds, so the written-analysis path is unverified and the cheap path is fast. No credential exists, so the credit accounting, the report tools and get_account_usage were not driven. The 247-country coverage claim is the vendor's; the sample proves three. The credit cost of a typical question is the number an operator needs and this pass could not produce it.
Read and held¶
The mcpservers.org /all page came back through the r.jina.ai reader proxy with 23 unique slugs and the counter at 18,577 against 18,402 at the afternoon read. Every slug matched at least one existing file in the tree, so the page carried no delta; the counter moved because the catalogue is growing, not because the newest page changed. Slugs read: ravenwallet-pro-docs-ai, cymetica-com-mcp-v1-stock-vectors, theluckystrike, reason-machines, auscaster, www-reslyce-com-flow-mcp, yir-ai-mcp-server, kinti, screenifystudio, liyupi, gmapscrawl-com-google-maps-scraper-mcp, www-repeatpost-com-mcp, compi-ai, secfeedapi, anantharajuc, aiimage-com, asms-ai, reinventa-es-asistentes-ia, apex-byte-technologies, renewroster-com-docs-mcp, whiteboardanimation-ai-mcp, nitrozenio and docs-safeprompt-dev-mcp.
Of the six new mcp.so cards above the afternoon stamp, five are recorded here with the reason, because a card that is not dispositioned becomes work for the next pass:
corridome(2026-11-01T00:15:52Z) is a US private-equity and M&A register, firms, funds, deals, advisers and people, each answer with the record behind it, and its own card states it requires a Corridome membership. Held on access rather than quality:corridome.comanswered 200, no keyless surface is offered, and the deal and fund entries here already cover the class from the filings side.gwix(18:49:44Z) shipped, above.browse-bridge(23:07:27Z) shares selected Chrome tabs with an agent athttps://mcp.browsebridge.com, held in the browser automation class this catalogue carries heavily.listing-698e96(19:17:31Z) is Tabulya, a Russian-language spreadsheet with AI, Python and SQL, held as a localised product rather than an operator surface this catalogue can apply.coinvello(19:10:08Z) is the same arrival as#6513, dispositioned below.myimage2life-8891ac(18:02:28Z) turns a chat image into a printed product and is consumer retail.
The issue window, dispositioned¶
Eighteen submissions between #6513 at 19:06:26Z and #6530 at 03:53:43Z. Three shipped, fifteen recorded so the next pass cannot read them as fresh.
CoinVello (#6513) is non-custodial crypto payments for agents: checkout sessions and invoices priced in fiat or crypto, webhooks, prepaid credits, test payments, across Bitcoin, Ethereum, Base, Arbitrum, Optimism, Polygon, BNB Chain, Avalanche, Solana and Tron, MIT licensed, npx @coinvello/mcp local stdio, keys refused unless COINVELLO_ALLOW_LIVE=1, and the card says nothing in the server can move funds. Held on saturation: the agent payments class here already carries wallets, x402 readers, USDC rails and Solana transfers, and this one needs a key before it answers anything.
ILens (#6514) decompiles .NET assemblies one member at a time, Windows x64 via winget, held as developer tooling outside the operator surface.
Clearcote MCP (#6515) drives a Chromium fork with about 20 tools and hands the same browser to Playwright, Puppeteer, browser-use or Crawl4AI through get_cdp_endpoint, with the fingerprint set in C++ rather than by a stealth plugin. It is the strongest hold of the window on capability and it is held on class: browser automation is one of the deepest slices in this catalogue, and the CDP handoff is the specific feature to test if that ever changes.
Fittest (#6516) is materials science, three foundation interatomic potentials producing a trust verdict against a DFT threshold on non-periodic molecules of seven heavy atoms or fewer. Remote, keyless and genuinely novel, and it answers a question no business operator asks.
CSI MCP (#6517) drives SAP2000 and ETABS through their COM API on Windows only, with 67 of 67 rebuilt verification examples matching published values, which is unusually rigorous evidence for a submission. Held on platform and vertical: structural engineering on Windows, and the catalogue carries no construction slice for it to join.
recall (#6518) serves Claude Code's own markdown memory files to other agents with [[wiki-links]], breadth-first recall trimmed to a token budget, secret redaction and a --read-only mode. Zero dependencies, MIT. Held on saturation in the agent memory class.
conductor (#6519) is 23 tools where every one has an input schema, an output schema and a mandatory verifier that checks the result in code before the client sees it, with deterministic jobs run as plain code and model-backed jobs run hermetically with retry and visible failure. It is the most interesting reliability entry of the window and it is developer tooling; the verification contract is the part to revisit if this catalogue grows a reliability slice.
Betterloop (#6520) ranks fixes from real users, rage clicks, dead buttons, JavaScript errors and slow pages, with a free plan and a Bearer key. Not measured because the key is required, and the analytics class is deep.
Wordy Page (#6521) publishes a page or a multi-site to wordy.page/your-name/ in one call, free, with OAuth 2.1 and PKCE or a personal key, and an agent mode that takes custom HTML and CSS with outgoing links only. Held on saturation: publishing entries here number in double figures.
nana800 (#6522) is a no-code automation builder positioned against Zapier, Make and n8n with 44 tools and ruble payments, keyed. Held on locale.
Frank (#6524) runs AI-moderated voice and chat interviews with your customers and returns themes with verbatim evidence, 14 tools, OAuth 2.0, publishing gated on the account owner's separate approval because it is the only action that spends credits. The endpoint answered 401 with Authentication required in 0.27 seconds, which is a measured refusal rather than an absent surface. Held on saturation: four customer-research entries already sit here, and the promotion test is whether themes come back as quotable evidence per interview rather than as a summary.
ytranscript (#6525) returns YouTube transcripts through one tool, including on auto-dubbed videos, with AI speech-to-text for videos without captions, free at 50 units a month. Held on saturation in the transcript class.
SKV (#6526) answered a keyless initialize at https://skv.network/mcp with serverInfo skv-memory at 1.30.0 and returned more than 20 tools, including a per-turn turn(user_message, draft) duty, a paged constitution, a shared "cube" library with voting, a court for cube version disputes and an isolated sandbox_run with 10 runs a minute. It is memory plus governance, which is a rarer combination than plain memory, and it is held for one pass rather than dropped: the promotion test is whether the verification seal and the court are enforceable or self-reported, and whether turn can be driven without first installing a constitution.
Shipvo (#6527) opens a React Native app in a simulated phone with 18 tools and a hosted endpoint at https://sim.shipvo.dev/mcp that answered HTTP 200 with serverInfo shipvo 0.0.6; the tools that open or act need a key. Held as developer mobile testing.
NexusBro (#6528) is a free keyless website audit at https://nexusbro.com/api/agents/mcp, initialize HTTP 200 with serverInfo nexusbro 1.1.0 and two tools, audit_website (crawl up to five pages, a 0 to 100 health score, top issues, a fix prompt and a report link) and audit_checks. Held on saturation: eight SEO audit entries already here, and this one adds breadth rather than a new question.
VocaLingo (#6530) summarises and dubs short video, clones voices and translates on-screen text across 15 tools. Held in the saturated media generation class.
Count¶
The guide tree held 1,073 server entries before this pass and holds 1,076 after it, across 1,466 files under hermes/mcp. Sources: the mcp.so homepage recentServers block read directly (132 parsed cards, newest 2026-11-01T00:15:52Z), mcpservers.org /all through the r.jina.ai reader proxy with X-Return-Format: html (23 unique slugs, counter 18,577), the chatmcp/mcpso issue window #6513 to #6530 read with bodies in eighteen authenticated calls, a keyless initialize, tools/list, list_rules and two rendered scan_url calls on https://checker.lazynext.com/mcp (200, 4 tools, 75 rules, scans of 12.97s and 61.1s), a keyless initialize, tools/list, guest_status, inventory, resolve_audience and one failed plan_quote_guest on https://app.distribution.tech/api/mcp/guest (200, 6 tools), a keyless initialize on https://api.gwix.ai/mcp (401) and on /v1/mcp (401), a keyless initialize, tools/list, get_data_status and get_pay on https://api.gwix.ai/mcp/sample (200, 27 tools) with one job_offer_check failing at a CloudFront 504, a keyless initialize on https://skv.network/mcp (200) and on https://sim.shipvo.dev/mcp (200), a keyless initialize and tools/list on https://nexusbro.com/api/agents/mcp (200, 2 tools), a 401 on https://api.hifrank.ai/mcp, a 404 on https://api.gwix.ai root, and HTTP reads on gwix.ai, gwix.ai/mcp, gwix.ai/pricing, corridome.com, coinvello.com/ai, checker.lazynext.com, distribution.tech, distribution.tech/docs, distribution.tech/docs/agents and distribution.tech/pricing, all 200.