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

Nearly nine hours after the night pass the listing had moved by seven, 17,714 to 17,721, the smallest move measured on this surface all week, with page one rotated again. Breadth cost three fetches and no more, so the weight went to the two surfaces that do move at this gap: the feed, which carried seven cards after the night cutoff, and the submission window, which advanced from #6214 to #6242 with twenty-seven issues and no #6237.

The shape of this window is unusually practical. Four of the six finds answer initialize with a refusal that states its own terms, two answer with the entire tool surface open and no key at all, and one pair completes a comparison the night pass named and deferred rather than half-documenting: where a resale seller and a storefront seller actually lose their margin.

What Was Catalogued

FlipProfit MCP is the item the night pass named first for this cycle, and it was drivable without a credential. https://mcp.fliprofit.com/mcp returned HTTP 200 on initialize as fliprofit-mcp 1.2.0, tools/list returned three tools, and compare_platforms was executed live for a $50 sale with a $10 item cost and $7.50 shipping to return thirteen platforms ranked by profit. The spread is the finding: Vinted nets $32.50 on no fee, Poshmark nets $22.50, and the difference between the best and worst platform is 44 percent of the profit on an identical sale. get_fee_changes adds dated fee changes each carrying the platform's own source URL and the date FlipProfit last verified it, published also as an Atom feed, and every error is RFC 9457 application/problem+json with a stable code. Held caveats: US rates only, compare_platforms uses each platform's default options so a category or seller tier must be passed before a margin is quoted, and fee rules in force on a past date plus change reminders sit behind Fee Watch at $6 a month or $60 a year.

Seller Profit Doctor MCP completes the pair from the storefront side. https://sellerprofitdoctor.com/api/mcp answered initialize with HTTP 200 as sellerprofitdoctor-free-tools 1.5.0 and tools/list returned sixteen tools, one more than the vendor page documents, covering Etsy, Shopify and Amazon: a check-up that returns a grade word with level 1 to 5, per-sale fee and profit breakdowns, break-even ROAS for Etsy Ads and for Meta and Google, dimensional-weight billing, Shopify plan fit, Amazon referral and fulfilment fees, and the fee table the numbers use. What makes it trustworthy is what it refuses: every result lists the defaults it filled in, a check-up with too few answers returns no grade and a grade.missing list instead of inventing one, Amazon figures that depend on measured size and stock carry estimate: true and sit outside the loss total, and the paid full-report surface is a separate endpoint so a free-wired assistant cannot run up a bill.

SpendQuery MCP answers the federal question. https://api.spendquery.com/mcp returned HTTP 200 on initialize with instructions that are a method rather than a pitch, and tools/list returned thirteen tools against the twelve named in the vendor llms.txt. run_plan is why it earns a guide: the answer arrives with status ok, clarify, unsupported or error, a one-paragraph answer, the query and the SQL, so a figure can be re-run instead of accepted, and an out-of-coverage question is declined rather than guessed. The data behind it is USAspending prime awards FY2023 through FY2026 dated through Sep 30, 2026, open SAM.gov notices, the FAR and DFARS as codified in the eCFR, SBA size standards under 13 CFR 121.201, SBIR and STTR awards since 1983, and state, county and city bids where those governments publish them. The line that matters is in the server's own instructions: amounts are obligations, money committed, not outlays.

BugHerd MCP is the official server from an established product, and the endpoint says so: https://mcp.bugherd.com/mcp returned HTTP 401 on initialize, the correct shape for something that requires an account. The published surface is get_task_details, list_project_tasks, list_my_tasks, list_all_tasks, update_task, add_comment, create_task, create_project and update_project, and what travels with each task is the point: the screenshot or recording, the URL, the CSS selector, the client comment, the attachments and the technical details captured when the client clicked, plus console output and replay where the plan includes them. The risk is named rather than smoothed: add_comment writes to a surface a paying client reads, so an agent that posts "fixed, redeployed" before the deploy is live has told a client something untrue in the thread they watch. Permissions mirror the connecting user, the organization chosen at connect time bounds the reach, and connected clients can be revoked.

byondr Receipts MCP is the second keyless surface of the window. https://byondr.co/receipts/mcp returned HTTP 200 as "Receipts & invoices to spreadsheet" 0.1.0 and tools/list returned two tools with full schemas. Each field comes back ok, check or missing, doubtful documents are flagged with a reason instead of guessed, repeated receipts are marked as duplicates, and exports are CSV, line-item CSV, QuickBooks CSV or a zip of the originals renamed by date, vendor and total, with QuickBooks currency and date format as explicit inputs. Retention is the shortest this catalog records for a document service: files are deleted after 24 hours and never used for training. Limits: 25 documents per call, 10 MB per file, PDFs up to 10 pages, two tools only.

WatchCron MCP is the only entry that could not be probed, and the report says so plainly: mcp.watchcron.com does not resolve and watchcron.com/mcp refuses a POST with HTTP 405, so the evidence is the published npm package watchcron-mcp 1.0.0 with its visible tool reference. Nine tools: list_monitors, get_monitor, create_monitor which returns the ping URL to add to the job, update_monitor, pause_monitor, delete_monitor, get_pings, get_incidents and list_channels. The model is a dead man's switch, the only monitor shape that catches a job failing silently: the job sends one HTTP request when it finishes and silence past the grace period is the alert, so nothing exposes a port. Risks recorded as rules: a monitor that never receives a ping reads healthy, a grace period tighter than the job's runtime pages until the team stops reading the pages, and delete_monitor takes the history with it.

What Was Held

Of the twenty-seven issues filed between #6215 and #6242, one is catalogued above and the other twenty-six were read and held with reasons rather than bulk-classed. AdRails (#6220) is a resubmission of a server this catalog already carries with a guide, so it was checked against the ledger and dropped instead of re-documented. SummaBooking and Marcora are official listings whose endpoints returned named refusals (Sign-in required. and Bearer token required), both held as the first items for the next pass because booking administration and brand context are real operator systems that deserve a driven surface, not a card. docket arrived on the feed as projects, tickets, time and budgets for agencies with access mirroring the user's own, held at tagline level as the third carry. GetCourier issues email inboxes to agents with SPF, DKIM and DMARC attached and zero-width text stripped, held on agent-email saturation beside Voidmail and Agent Inbox even though its untrusted-data marking is the best written of them. Borro lends stock on Robinhood Chain with writes returning unsigned transactions, held on unverified custody; Trickl is a round-up payments platform on sandbox keys, held in the payments class; DalalOS and Lava are Indian and Chinese market verticals; the Helix GTM fleet is six prompt-template servers from one publisher with no data source behind them, held once as a fleet; Atlas Data, the itulo EU and Nordic suite and Polish MCP Servers are Apify actor distribution, held as one class. Thingery, pcprice.watch and Totally Tarot Calculators are consumer surfaces; RELTO is real-estate virtual staging whose compliance verification is the most interesting part of a vertical this catalog does not carry; Benchlight, DomainSpace, Doply, BANKSapi Docs, CloudGPU and PROCESIO were held on vertical or on class, and byondr Image Tools is held as media tooling while its receipts sibling is catalogued.

On the feed, seven cards were filed after the night cutoff and each was read: getimg, Aguasto Research, Talkform, My Privacy Agent, Hackshop, Mcp-Warden and docket. On /all pages 1 to 3, 85 unique slugs were cross-referenced three ways against headings, guide directories and every sweep ledger: 51 carried a named prior disposition, 5 were prose dispositions, 19 matched only a generic token in a heading and were re-checked on their real names, and 10 were triage-new. Two of those ten, Lilo and thumbs, were already carried as agent-browser holds in the night pass. Three of the 19 had no ledger entry in any sweep and are named here for the first time: Caisey (MSP device and session control with technician approval on every command), Tech and Toast (an AI-to-human handoff utility) and the juergenkoller desktop family (PDF content search over local files and Apple Mail, plus Nemeton controlling macOS virtual machines), the last held as macOS desktop tooling with the caveat that a local bridge over a user's own mail store should be a deliberate install rather than a directory click.

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: listed and not unlisted in the registry API (?q=corpusiq, public slug Cprusiq); https://smithery.ai/server/corpusiq returns 308 to the plural route.
• 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.
• Anthropic Claude directory: still not listed, no change to the October 4 correction.

Sources: mcpservers.org /all pages 1 to 3 through the r.jina.ai reader proxy at the 17:02Z stamp (85 unique slugs, counter 17,721 against 17,714 at the night pass), mcp.so /feed page 1 from the raw page with a browser user agent (30 server cards, newest 2026-10-27T16:33:59Z), the chatmcp/mcpso issue window #6215 to #6242 over the authenticated GitHub API (27 issues, #6237 missing), live JSON-RPC probes on seven hosted endpoints plus four URL attempts for the stdio-only entry, a FlipProfit tools/call executed against live rates, the npm registry record and repository README for watchcron-mcp 1.0.0, the vendor documentation and llms.txt files for FlipProfit, Seller Profit Doctor, SpendQuery, BugHerd and byondr, and the official MCP registry entries com.sellerprofitdoctor/free-tools and co.byondr/receipts.

Catalog 1,009 to 1,015 servers, guides 1,157 to 1,163.

FAQ

Why did a nine hour window only add seven servers?

Because the listing surface is not the submission surface. The counter measures what mcpservers.org has published, and at 17,721 it had barely moved, while the chatmcp/mcpso window filed twenty-seven issues in the same stretch. The delta lives in the window and in the endpoints, not in the counter.

What made four servers drivable without a key?

They were built for machine callers rather than for sign-ups. FlipProfit, Seller Profit Doctor, SpendQuery and byondr all return initialize and the full tools/list unauthenticated, which is a deliberate design choice: the operator can evaluate the contract before handing over a credential, and the catalog can document tool names from the server itself rather than from a marketing page.

How are holds different from skips?

A hold names the reason and the evidence level, so the next pass can pick it up without re-reading. A skip would be a name with no disposition, which is how a card gets carried three cycles and then catalogued by accident.