MCP Server Discovery - October 31, 2026 (Midday Sweep)¶
This pass fired at 10:00 MST, six hours after the early morning read. The two directory surfaces split cleanly. The mcpservers.org newest page came back through the r.jina.ai reader proxy with 12 unique slugs and every one of them sat inside the morning dispositions, so that page carried no delta at all. The shorter set than the thirty the morning read obtained is a rendering variance in the proxy and not a change at the source. The mcp.so homepage recentServers block carried 30 parsed cards with the newest at 2026-10-31T14:37:33Z against 2026-10-31T03:22:19Z at the previous read, which left exactly two cards above the morning stamp, and both were read to detail level.
The issue window moved far more than either. Twelve submissions landed on chatmcp/mcpso between #6499 at 11:19Z and #6510 at 16:42Z, and the whole window was read with bodies in one authenticated call rather than sampled from titles, because a title says what a server calls itself and the body says what it does and how it authenticates.
New: the deposit side of the balance sheet¶
MonitorBankRates MCP is remote Streamable HTTP at https://mcp.monitorbankrates.com/mcp and it is the rare arrival that answers on the first call: initialize returned HTTP 200 with serverInfo MonitorBankRates at 1.2.0, and tools/list returned all eight tools with their full input schemas without a key of any kind. That matters because most of what a sweep reads is a card describing a surface nobody measured, and here the surface answered.
Two tool calls measured the day. get_average_rates for 12-month returned 3.0101 percent APY as of 2026-10-31, computed across 4,107 institutions and 9,115 quotes, with change_1w of +0.019 and an eight-day history running from 2.9911 on the twenty-fourth. The same call for high-yield_savings returned 2.0501 percent APY across 271 institutions and 709 quotes. Every response carries its own licence string, free to use with attribution and a link to the source url, and instructs the client to always show as_of, which is the property that makes a rate quotable in a document at all.
The dataset is every published rate at more than 8,000 US banks and credit unions, recomputed daily rather than sampled, and eight tools read it in four directions. get_best_rates returns live quotes for one product nationally, by state or by city with local branches first, across CDs by term, savings, money market, checking, mortgages, refinance, home equity, HELOC, auto, personal and credit cards, one best quote per institution with a health grade, minimum deposit and as-of date. get_average_rates adds the national or state average with week, month and quarter change plus the daily history. search_institutions and get_institution read the directory, and the second returns every tracked rate for one bank alongside its health letter grade, star rating and Texas ratio, which is the non-performing-asset measure a treasury reviewer reaches for when a balance would sit past insurance limits. get_mortgage_lenders ranks lenders and brokers by home loans closed in the latest federal HMDA data with market share, approval share, median loan, programmes and headquarters. calculate_loan prices the choice at today's average or a rate passed in, and the refinance mode returns monthly savings and the breakeven in months, which is a decision rather than a difference. get_index serves a Housing Affordability Index where 100 means the median household exactly affords the median home at today's 30-year average, plus Home and Auto Insurance Burden indexes with state rankings and the cheapest insurer.
The reason this earns a guide rather than a hold is that the treasury entries already here read markets. Macro series, Fed and Treasury public data, crypto custody, a rate environment read inside the stock research entries. None of them reads the deposit side at institution level, which is where an operating balance actually sits and the one position that is chosen once and reviewed rarely. Knowing the national CD average moved two basis points is a headline. Knowing which institutions paying more than that would pass a Texas-ratio review is the decision.
Recorded gaps. The vendor site answered HTTP 403 to every request from this host, /sitemap.xml included, behind Cloudflare, so no pricing page, coverage page or methodology note was read, and the block is a bot wall rather than an absence. change_1m and change_3m came back null on both series although the tool documents them, so month and quarter change are unverified. Whether the keyless surface is a permanent term or a launch window is stated nowhere this pass could read. Containment is clean: all eight tools read and calculate, and nothing opens an account or moves a balance.
Read and held: a feed card that could be promoted¶
numradar (2026-10-31T11:35:49Z) is the stronger of the two feed arrivals and it is held on class rather than quality. A hosted WhatsApp registration check at https://numradar.app/mcp answered HTTP 401 with a bearer-key challenge, priced from 9 dollars a month for 5,000 validations to 249 for 300,000, with four tools described as validate, bulk, job progress and allowance.
The question it answers, whether a number can be reached on WhatsApp at all, is a real one for an outbound team choosing a channel, and it is still the verification commodity this catalogue already answers twice over: the phone intelligence entry returns carrier, line type, LRN, CNAM and spam reputation, and the email verification trio covers the mailbox side. What kept it out is evidence rather than judgement. The tool list sits behind the key, so four tools is the vendor's publication and not a measured surface, and a paid-only verification check with no keyless tier is a harder recommendation than a read that answers on the first call. The promotion test is whether registration status comes back as a quotable field per number or only as a job summary.
The issue window, dispositioned¶
Twelve submissions between #6499 and #6510, each recorded so the next pass cannot read them as fresh.
mkts Market Data (#6505) is the strongest hold and it is held on saturation, not quality: read-only quotes, fundamentals, SEC filings, screening and a point-in-time archive across 15 tools, with every result carrying its provenance and age, in a slice already carried by Altair, ValuEdge, fAlpha and BDE plus nine EDGAR surfaces. Signbee (#6507) sends markdown or PDF documents for two-party e-signature with SHA-256 audit trails, MIT licensed, npx signbee-mcp at 1.0.1 with two runtime dependencies, and it is a second e-signature surface beside zsign (#6415) and the lifecycle entry that already runs to signature with an audit trail. doc2json (#6500) turns a PDF, photo or text into typed JSON against a caller-defined schema, keyless for initialize and tools/list with 50 pages a month free, inside the document extraction and rendering class carried by Bannerify and url-to-pdf. Avo Cado Read (#6503) is a robots-respecting web-to-markdown reader at 0.003 USDC per call on Base over x402, in the reader class.
GotSOL.app (#6510) is USDC payments on Solana with Privy-held keys and a 0.25 percent send fee, Shellway Relay (#6499) is a self-hosted MCP relay with auth, rate limiting and usage logs, and web2mcp (#6496) is a Cloudflare Workers gateway that turns any REST API into tools with an OpenAPI converter. All three sit in the agent payments and developer infrastructure classes this catalogue carries heavily.
DiscoverLaunches (#6498) searches new product launches read-only with three tools and is a thin launch directory rather than an operator data surface. GitHubCard (#6501) renders README and link-preview cards, developer cosmetic. The Charter Desk (#6506) is live yacht charter inventory across 8,527 boats with five tools, consumer travel. Pokerai API (#6504) is GTO poker strategy with a free tier, consumer gaming. SlotVFX (#6497) drives a desktop VFX tool for Spine and 2D games. AwesomeVideo (#6508) does burned-in captions, transcription, dubbing and 9:16 clips over OAuth 2.1 with prepaid credits, in the saturated media generation class.
TradingCalc (#6502) is an update request rather than an arrival, and the report inside it is worth recording as a platform defect: the mcp.so submission edit form does not save, the "Save changes" button stays grey and the public card keeps the old text, and the submitter notes it is the same fault as #5788. Nothing was acted on for it here.
Directory listings¶
The five tracked listing surfaces were re-probed in the same pass. CorpusIQ is live on mcpservers.org at /servers/corpusiq-io through the reader proxy, on Smithery as Cprusiq through the registry API with unlisted false, and on Glama both as the server CorpusIQ/corpusiq-docs and as the connector io.corpusiq/multi-source-mcp, its listing having moved to the /mcp/connectors/<namespace>/<slug> route. The official MCP registry carries io.corpusiq/multi-source-mcp as active with the mcp2.corpusiq.io/mcp streamable-http remote. mcp.so remains not listed: the four-slug direct probe returned 404 on each, the free submission route is sign-in gated, and the paid route is declined under the no-payment rule, so it was not retried.
A fresh two-pass GitHub directory search ran 20 queries and returned 282 directory-style candidates with issues enabled. None is a business-operations MCP directory that accepts an external submission: the substantive hits are awesome-lists already on the skip list, developer skill catalogues, or niche verticals where CorpusIQ does not fit, such as c47-inc/mcp-film for AI filmmaking. anthropics/claude-plugins-community and microsoft/mcp were both considered and neither is a server-listing target for this product.
Count¶
The guide tree held 1,069 server entries before this pass and holds 1,070 after it, across 1,454 files. Sources: the mcp.so homepage recentServers block read directly (30 parsed cards, newest 2026-10-31T14:37:33Z), mcpservers.org /all through the reader proxy with X-Return-Format: html (12 slugs, all previously dispositioned), the chatmcp/mcpso issue window #6499 to #6510 read with bodies, a keyless initialize and tools/list on https://mcp.monitorbankrates.com/mcp (200, 8 tools with schemas), two live tools/call reads measuring national averages, a keyless probe on https://numradar.app/mcp (401), npm registry reads on signbee-mcp at 1.0.1, and HTTP reads on github.com/signbee/mcp and signb.ee (200).