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

Ten hours after the midday supplement the listing moved by 192 servers, 17,721 to 17,913, which is the largest single gap measured on this surface since the late pass of October 26 and roughly twenty times what the midday window carried. Page one rotated completely. The feed added three cards after its own cutoff, and the submission window advanced from #6243 to #6271 with twenty-seven issues and no #6247.

The shape of this window is back office. Three of the six finds are systems where a wrong number becomes a filing, a book entry or a client document: a Brazilian fiscal server whose every draft waits for a person, a Swedish accounting server that hides its write tools in an environment variable, and a hosted document service whose containment is a two-scope OAuth split. The other three answer measurement questions the catalog was short of: which businesses AI engines actually name, whether a citation resolves to the thing it claims, and what a person's own accounts look like behind an approval gate that lives on a server rather than in a prompt.

What Was Catalogued

Six entries earned guides. Each was read to the level its own surface allowed, and where a surface could not be driven the reason is named in the entry rather than the class blamed.

SendNF MCP filed as #6267 and is the first server in this catalog to put a tax invoice path behind a mandatory human draft. https://api.sendnf.com/mcp returned HTTP 401 with a Portuguese refusal naming the header it wants, which is the correct shape for a fiscal system, so the surface is the vendor's published table of 36 tools and the detail that makes it catalogueable is create_note_draft: it builds the NFS-e or NFC-e, validates it exactly as the issuance path would, and leaves it in pending_approval with an approval link and a preview PDF that carries no fiscal value, even when the account lock is switched off, expiring after seven days. Around that: the key prefix decides reality (nt_test_ is homologation with no fiscal value, nt_live_ is production, and account_overview reports which mode the key is in), preview_nfse validates against the official layout or the municipal provider and computes IBS and CBS before anything is transmitted, issue_nfse waits about twenty-five seconds for authorized, rejected, blocked or pending, correction runs by substitution rather than deletion (get_substitution_draft then substitute_document pointing at the original), import_invoice reads the XML of a note the company already issued and completes the whole company setup from it, upload_certificate takes the A1 certificate, create_onboarding_link makes the company owner supply their own credentials, and exports build a per-company and per-month zip of XML and PDF plus a spreadsheet with the fiscal key, on a one-hour signed URL with seven-day retention. Limitations recorded: Brazil only and the vendor writes in Portuguese, setup needs an A1 certificate and municipal credentials that how_to_get documents rather than fetches, and no call was executed because the endpoint refuses everything without a key.

bokio-mcp answers the founder question rather than the accountant question. Local stdio through npx -y bokio-mcp, npm bokio-mcp 0.5.0, MIT, Node 22 or later, 60 tools of which 36 are read-only and 24 are not registered at all unless BOKIO_ALLOW_WRITES=1. The containment is the shape this catalog rewards: write tools are absent from the tool list rather than gated inside a call, so an agent that asks for a posting tool gets a tool-not-found and no prompt can restore it. The second half is the arithmetic: computed tools run over the SIE export and return finished numbers in whole ore for balances at a date, a ledger for one account, VAT and profit and loss for a period, open items, suspected double bookings and a bank statement reconciliation, so the same question returns the same number instead of a freshly reasoned one. Read surface from the vendor table covers journal entries, invoices, credit notes, customers, suppliers, supplier invoices, items, uploads, bank payments (read only), chart of accounts, fiscal years and the export itself. Limitations recorded: Sweden only, no hosted endpoint so a browser client cannot attach it, one company per instance with the token in the environment, Bokio's 5,000 requests per company per month and 200 per minute, and no call executed because the server needs a Bokio token to start.

Recued MCP is the governance entry, and the reason it earns a guide is that it answers the permission question outside the agent. A server the owner runs, AGPL-3.0, install by one script, MCP on port 7717, sitting between agents and mail, calendar, files, contacts and CRM records. Each agent connects through a contract that lists the tools it may use; reads run immediately, writes and deletes wait for the owner's approval by default, and every call is recorded at that one gate. Before the model sees warehouse data a PII alias step replaces names, emails, organisations, phones and addresses with stand-ins while keeping the region and which people belong to the same organisation, so relationships survive the swap. Encryption at rest with a 24-word recovery key the owner holds and nobody can reset. The documentation is also honest about the boundaries: it states that Recued is not a coding agent, not a chat app, not a runner for SKILL.md files, and not built for long multi-step questions, which is the correct scope statement for a gate. Limitations recorded: the owner runs the server and holds the backups and the key, AGPL-3.0 matters to anyone embedding it, approval fatigue is the real failure mode so the contract has to stay narrow, PII aliasing is best effort in the vendor's own words, one star and first publication this week, and no call executed because the endpoint only exists on an installed instance.

Receipts Index MCP is the AI visibility entry and it is catalogued for the unit rather than the size. Local stdio through npx -y receipts-index-mcp, MIT client, no auth, documented as read-only and unable to pay. Driven on the free surface with no key: a GET on https://api.receiptsindex.com/v1/index returned row_count 2,714, first_date_utc 2026-07-29, last_date_utc 2026-10-27, a page of 100 with next_start 101, and rows carrying only a row id and a UTC date. Each row is one check on one engine on one UTC date for a fixed local buyer question, holding the businesses the engine named and the sources it showed, and that content sits behind x402 routes priced at 0.27 USDC per row and 1.81 USDC per day on Base. A POST to the same path returns HTTP 405 method not allowed, recorded because it makes a healthy HTTP dataset look like a dead MCP endpoint to a naive probe. Limitations recorded: 2,714 rows over ninety days from one publisher is a sample and not a panel, US local businesses only, the content is paid and no payment was made, one methodology so comparability with another vendor is unestablished, and x402 puts a spend-capable path on the host if automated. The publisher's repository describes the corpus as customer zero with losses included, which is the reason the claim can be acted on.

Metis Check MCP covers the case where a citation is offered as proof. https://mcp.metisfoundry.com/mcp answered a bare initialize with HTTP 401 and the body Sign in to Metis to connect, so the surface is the vendor documentation, and what it documents is the part most tools in this class round away: four states, settled true, settled false, not decidable and unreachable, where the last two are the ones a boolean throws away. Identifiers are checked at the registry that issues them (DOI, PubMed, arXiv, LEI, ORCID, ROR, ISBN, SEC CIK), each report is signed so anyone can recheck it, auto-check is off by default and depends on the client calling the tool, checks stop at the monthly allowance and never start an overage payment, and the connector receives only the identifiers submitted rather than a conversation. On the Verified plan a named researcher verifies a specific report in real time and files a signed addendum with the name and date, or covers every report for eight hours. The vendor states the scope of that human step precisely: the researcher confirms the identifier resolves to the work or entity the citation names and what the registry says, not whether the surrounding argument is true. Limitations recorded: five checks a month free, identifiers only, registries are the authority and not the truth, and no call executed without an account.

DocsAutomator MCP is the document entry, and it earns a guide because it removes generation rather than improving it. https://mcp.docsautomator.co/mcp returned HTTP 401 with a body naming all three auth paths and a discovery pointer, and the discovery document is precise: resource https://mcp.docsautomator.co/mcp, authorization server https://api.docsautomator.co, and exactly two scopes, docsautomator:read and docsautomator:write. Two scopes is a better containment surface than one because the write scope is what sends for signature. The deterministic claim is the point: the template is a Google Doc or Word file the customer already owns, the data comes from Airtable, SmartSuite or Google Sheets with linked fields expanded and line items nested three levels deep, and the agent's job is to choose the template, choose the record and trigger the render, so layout, tax id and footer cannot drift between runs. The lifecycle runs to signature in one primitive with an audit trail, and the vendor promotes a team approvals step before documents go out, which is the ordering that matters for anything client-facing. Context recorded rather than claimed as proof: 1,200+ business customers and 4.9 out of 5 across 15+ G2 reviews. Limitations recorded: no render was executed because every path needs an account, the data sources are a named list, a wrong template produces a consistently wrong document, the write scope is not per-template, and e-signature jurisdiction was not settled by anything read in this pass.

What Was Held

Of the twenty-seven issues filed between #6244 and #6271, twenty-one were dispositioned on class, on shape or on an endpoint that could not be driven, and each is named here so the next pass cannot read it as fresh.

B2B Quant & Regulatory Compliance Suite (#6257) is the most interesting hold and the reason is structural rather than editorial: it advertises 25 stateless B2B servers and 30 n8n workflows for SEC EDGAR, SAM.gov, GLEIF LEI, ClinicalTrials.gov, CISA KEV, ICANN RDAP, EU VIES VAT and Brazilian central bank data, but its remote endpoint is the Apify MCP multiplexer with a list of actors in the query string, so what a client connects to is Apify with an actor selection rather than a server with a contract. Held as a publisher fleet behind an account, with the compliance and procurement classes recorded as already carried. JustClaw-It (#6254) indexes 1.4M code components with a quality score and the licence and attribution on every result, serving only permissive and weak-copyleft code, free search with metered pulls: held as coding-agent infrastructure with an unusually good licence model, and worth a guide in a window where the six are not all back office. AstraEye Labs (#6262) serves finished AI research reports on US stocks with the method and sources shown and a member-set daily credit limit, held in the finance class this window does not need a fourth of. Record Meeting (#6250) searches speaker-labelled meeting transcripts across twelve tools and is held on class, which already carries meeting capture with transcription. Nabu (#6249) is a self-hosted read-only library of 2.47M ancient-text documents with 14 research tools and a licence class on every passage: held as a scholarly vertical, recorded because the licence tracking is better than most commercial datasets manage. Connect IQ MCP (#6245) covers Garmin Connect IQ store listings, builds and the simulator, held as developer tooling. Memory.fit (#6246) is 103 tools of athlete training, sleep and recovery data across eleven wearables with an undo on every write: held in the consumer health class. Notifario (#6244) monitors uptime from Moscow, Kazan and Helsinki with 14 read and 12 write tools, held as infrastructure monitoring already carried. Studiofinder (#6260) is a boutique fitness directory over OpenStreetMap under ODbL, Editmamei (#6261) is a Photoshop surface, Databricks Field Guide and TechFabric (#6263, #6264) are one publisher's documentation reader with a free anonymous endpoint, AIR Bandenservice (#6271) orders tyres in Haarlem with payment on Mollie, Bacon Services (#6270) is pay-per-call image tooling over x402, and SendNF's neighbour from the previous window, Agent Bounty Feed, was filed four times (#6252, #6253, #6255, #6256, three of them prefixed docs:), which is recorded as a filing pattern rather than four submissions. BountyRadar (#6258) and Jobbi Agent Gigs (#6259) carried a body that said only what they improve and fixed an unrelated issue number, so nothing could be read and the shape is named in the hold. how.is (#6265) is already catalogued on October 25 with a full guide, and OpenSERP (#6251) is already a body entry, so neither was rewritten.

On /all pages 1 to 3, 90 slugs were read and most are the long tail of publisher pages. The business-relevant remainder was held on class or on shape: CrakSEO now lists ten separate servers across SERP, keyword research, AI visibility, backlinks, URL indexing, rank tracking, Search Console, site audit and local SEO, all hosted with OAuth or a scoped key and credits charged per data call, held as credit-metered resale in the SEO class this catalog carries with its own sourcing; GORZ grades Solana memecoins on chain data, held in the crypto class; Placepoint serves Norwegian and Swedish property and company data across three remote servers, held as a regional data vertical; Compr estimates comps, ARV, repair cost and max offer for US real estate investors, held as a property vertical; AXL is already catalogued at axl-mcp; OwlMem is shared memory across assistants with a vault, held in the saturated memory class; Cadabra generates video, image, voice and music with a price shown before every paid run, held in the media generation class; PlateQuery, PostCode Check, searchnapply, dayNumber, Hotel Lobby, La Demeure des Dix, 313 Automations, SERVIQO Agents and the consumer and utility remainder were read and held on class, with toli, CloudGPU and Datapixel-class daemon and compute entries held as developer infrastructure. Two ledger checks changed what is written here: CheckMyVIN and OpenSERP were both confirmed as existing body entries rather than new, and no name in this section was carried from an earlier report without its own disposition.

Directory Status Re-Verified

• mcpservers.org: CorpusIQ listed at /servers/corpusiq-io. A direct curl still returns the Cloudflare challenge, and 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 02:04:34 GMT stamp (90 unique slugs, counter 17,913 against 17,721 at the midday read), mcp.so /feed page 1 from the raw page with a browser user agent (30 server cards, newest 2026-10-27T22:39:16Z, three filed after the prior cutoff at 22:39:16Z, 19:43:38Z and 19:22:29Z), the chatmcp/mcpso issue window #6244 to #6271 over the authenticated GitHub API (27 issues, #6247 absent), live probes on four endpoints (SendNF 401, Metis Check 401, DocsAutomator 401 plus its RFC 9728 document, Receipts Index 200 on a read-only GET with a 405 on POST), unautharied repository metadata and README reads for diblaze/bokio-mcp, recued/recued, sendnf/sendnf-mcp and mrcrtr1979-droid/namebeam-visibility-index, and the npm registry record for bokio-mcp 0.5.0.

Catalog 1,017 to 1,023 servers, guides 1,165 to 1,171.

FAQ

Why does a ten hour gap beat a six hour one by this much?

Because the submission window is bursty and the listing follows it. Twenty-seven issues arrived in this gap against twenty-seven across the whole midday window and one in its supplement, and 192 servers were indexed against 7. Neither number predicts the next window, which is why each pass measures rather than extrapolates.

What makes an entry 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 anyone wiring agents to anything. SendNF and bokio-mcp clear the second and third, Recued and Metis Check clear the third, and Receipts Index and DocsAutomator clear the first.

Why catalogue a dataset with 2,714 rows and a paid content route?

Because the unit is right. A visibility claim that cannot be reduced to engine, date, naming and sources is unverifiable, and this is the only surface in the window publishing its rows in that shape, including the weeks the publisher itself was not named.

How were entries without a drivable endpoint verified?

Against their own published contracts. SendNF, Metis Check and DocsAutomator each returned a 401 that named what it wanted, and for those the evidence is the tool table, the discovery document or the vendor reference, with the absent call named as the gap rather than described as a capability.