MCP Server Discovery - October 26, 2026 (Late Sweep)

Eleven hours after the midday pass the listing had moved by three: mcpservers.org served 17,525 servers at the 04:01Z reader stamp against 17,522 at 13:32Z, with page one in the identical order. A counter that moves by three across eleven hours is not a surface worth re-reading for breadth, so this pass spent its weight where a delta actually lands at this gap: the submission window, which advanced from #6176 to #6196 with no gaps across twenty issues, the feed, which rotated fresh cards in at the top, and the endpoints themselves.

The shape of the evidence this cycle was unusual. Four of the five finds are gated on initialize with HTTP 401, which means the capability had to be read from each server's own discovery documents rather than from a tool list. That turned out to be the more informative read: an RFC 9728 protected-resource document states the resource identifier and the scopes an agent can hold, and an RFC 8414 authorization-server document states whether a client can register itself, whether PKCE is mandatory and whether a device flow exists. A server that publishes six separate scopes is a different risk than one that publishes a single coarse one, and that difference is visible before anyone signs in.

What Was Catalogued

Five entries earned guides. Each was verified against its own endpoint or its own generated reference, and none of them was described.

PickFu MCP is the find of the cycle, and it arrived on the feed rather than in the issue window. https://mcp.pickfu.com/mcp returned HTTP 401 authentication_required with a resource_metadata pointer in the body, so the contract was read from two documents: the protected-resource response names the resource, an authorization server and identity scopes only (openid, email, profile, offline_access), with a WorkOS jwks_uri; the authorization-server response names authorize, device authorization and introspection endpoints on auth.pickfu.com, client_id_metadata_document_supported: true, PKCE S256 and grants for authorization code, refresh and device code. The vendor's own MCP reference documents the tool surface: save_survey, publish_survey, stop_survey with unused responses refunded, get_survey returning AI summaries plus per-question vote totals and a winner, get_survey_responses with explanations, demographics and stored translations, export_survey_results in CSV, PDF and PNG, list_available_targeting and list_available_reporting, build_poll_url, template tools, an upload-session flow for local files, generate_image, and a feature-gated synthetic-audience set that requires an approved plan before it runs. The money path is the reason to catalogue it. publish_survey draws the account's credit balance and then charges the saved card, returning HTTP 402 only when no usable card exists, and the documented flow keeps a poll an unpublished draft, prices it, shows the targeting and waits for approval before that call. Every response carries a written explanation from a real respondent, and the vendor explicitly warns against adding an open-ended "why" question because it is already collected and costs a question slot. Catalogued because the consumer-research class was empty in this catalog and because validated product, packaging, name and price decisions are the most expensive things a small brand gets wrong.

AXL MCP at https://app.axl.tech/mcp answered a GET with the clearest refusal read this cycle: HTTP 405 and this MCP server is stateless - it exposes no GET SSE stream and no DELETE session termination. Send JSON-RPC over POST. initialize without a token returned 401. The protected-resource document names one resource, one authorization server and scopes_supported: ["mcp"]; the authorization-server document publishes a registration endpoint (/oauth/register, an RFC 7591 public client), token_endpoint_auth_methods_supported: ["none"], PKCE S256, authorization code and refresh grants, and a revocation endpoint. Vendor-stated surface: roughly 1,100 account operations behind a small fixed tool set the agent navigates itself, covering courses and LMS, webinars, community, appointments, landing pages and checkout, CRM and pipelines, payment subscriptions, email and nine-channel messaging, all returning a link to the created record on a successful write. What separates it from the other all-in-ones in this catalog is the containment, stated plainly by the vendor: 21 sensitive operations are removed from the AI surface in the server rather than filtered in a prompt (billing writes, token minting, permission changes, admin creation), the token is an ordinary admin token with one coarse scope, and the documented route to a narrower connection is an API key with read rights. Limits are honest too: concurrency rather than rate limiting with 429 and Retry-After plus an X-Concurrency-Limit header to read, uploads that leave the MCP surface for one raw PUT so an MCP-only agent cannot finish one, and long responses truncated near 200,000 characters.

Customermates MCP at https://customermates.com/api/v1/mcp returned HTTP 401 {"error":"unauthorized"}, and its protected-resource document answers with jwks_uri on its own host, RS256 resource signing and identity scopes. The CRM class this catalog carries is thin, and this is the first entry in it whose core is open source. AGPL-3.0 with a Docker Compose path, so the vendor leaves the data path entirely, while the MCP endpoint stays the same. The permission model is the reason it earns a guide: every operation runs inside exactly one workspace, a per-workspace URL /api/v1/mcp/w/<WORKSPACE_ID> accepts only that workspace's credentials, and the failures are named rather than generic - 406 without the Accept header, 403 workspace_mismatch for another workspace's key, 403 insufficient_permissions when the user's role lacks Read access All on the API and Webhooks row, 403 plan_required for a cloud workspace awaiting its first plan, and a resource_metadata pointer for a token from the wrong workspace so the client re-authorizes instead of silently writing. The tool reference reads as a full CRM rather than a reader: get_workspace_context, discover_record_types, query_crm_records, search_crm_records, read_crm_record, mutate_crm_record, preview_crm_deletion, import and export, query_crm_measure, messaging and send_email with drafts, calendars, tasks and routines, Knowledge Base pages, roles, teams, webhooks and trash, plus LinkedIn sales lookups and social engagement. Toolsets can be narrowed with ?toolsets=, and the server sends workflow instructions that the vendor is explicit are not a server-enforced confirmation gate.

Rus.sh MCP at https://rus.sh/mcp returned 401 on initialize, and its authorization-server document is the most granular small surface read this cycle: six scopes split by object and by verb - account:read, links:read, links:write, analytics:read, qr:read, qr:write - with a registration endpoint, PKCE S256, protected_resources: ["https://rus.sh/mcp"] and a public client. Sixteen vendor-documented verbs: read account, links, QR codes and analytics for each; create a link with a custom address, title and tags; update, archive, restore, pause and resume links; archive and restore QR codes. Containment is documented rather than implied: per-connector permission approval with individual scopes untickable, an email on every connection, immediate revocation, no permanent delete anywhere (archive is restorable), and no path to API keys, plan, billing or account settings from an assistant. Held to ★★ because attribution depth is unpublished - the retention window, the bot filtering and the click-to-unique-visitor question are not stated on the pages read - and because new links consume a monthly plan limit.

App Store Connect MCP is a local stdio Go binary, so there was no endpoint to probe; the evidence is a generated tool reference in the repository, and it is the most precisely documented surface in this pass: 76 tools, each tagged read, write or destructive, with a confirm mode of none, flag, plan or human, across sixteen toolsets of which six are on by default. The revenue side is what matters here: sales_report returns Sales and Trends aggregates - units by category, proceeds per currency, top-N, with an optional full CSV link; analytics_report sums an app's newest analytics segments by dimension on a 24 to 48 hour lag; analytics_setup is the prerequisite that creates the ONGOING report request; insights_summary answers "how is my app doing" in one call with each section marked ok or unavailable with the reason; and analytics_funnel_report walks impressions to page views to downloads with conversion by source and territory, sessions, crashes, installs against deletions and a previous-period comparison. Releases are a workflow, not a bag of REST calls: release_status preflights a version, release_prepare applies a plan, release_submit submits with IAPs, release_publish releases, release_phased drives a seven-day rollout. Anything public or hard to undo is destructive-class with a human gate through elicitation, and raw API calls validate against a pinned spec first. MIT, brew install idoo/tap/app-store-connect-mcp with the tap answering 200, and registry io.github.idoo/app-store-connect-mcp at 0.13.0. This is the account-side bookkeeping complement to AppSigma, which reads the public store as a market; the two answer different questions and neither replaces the other.

What Was Held

Of the twenty issues filed between #6177 and #6196 with no gaps, and the thirty feed cards, the rest were read and held on class or on a dead endpoint.

argon-recon (#6177) was held on a failed endpoint rather than on class: https://argon.kenetic0x00.com/mcp returned a Cloudflare 1033 tunnel error at probe time, so attack-surface and CISA KEV recon stays unverified and un-catalogued until the host answers. Headroom (#6196) watches Windows RAM, commit charge and disk and returns plans rather than making changes, held as developer desktop tooling; Web Search API Eval (#6194) benchmarks You.com, Tavily, Exa and Parallel on thirty agent queries, held as developer procurement infrastructure; LogBrew (#6190) reads issues, logs, traces, spans and metrics for a coding agent, held as developer observability beside Watchcron and Staygreen; Testera (#6182) is test management with cases, plans, runs and coverage, held as QA infrastructure; amem (#6181) is local memory for coding agents, held in the developer memory class; AI Agents Listing (#6183) is a directory of agents and servers, held because a directory is not a record set. Canvas LMS (#6180, #6179) arrived twice, local and remote editions of one publisher, held once as an education vertical: 209 student tools over grades, deadlines and submissions are a student's records, not an operator's. peptides.farm (#6188) serves FDA Orange Book and CMS datasets, held as a biomedical vertical; Probative (#6195) publishes US lawyer advertising rules, bar filing requirements and fee caps for all fifty states and DC, cited and keyless, held as a legal-marketing vertical; GC Airport Transfers (#6193) books fixed-price OAuth transfers, held on travel geography; dayNumber (#6191) logs a founder's daily build journey, held as consumer productivity. Hyperresearch Answers (#6186) returns one checked answer per question over OAuth, held against the AI-visibility and answer class this window already carries through CitedBy, geoLint, LogNorm and Cite42. ai-reviewing (#6178) is a public database of AI-written reviews with a one-click approval URL, held as an unverifiable user-generated corpus. Waspes (#6184) generates sites, contact forms and Roblox assets, held in the site-builder class already carried on the listing.

On the feed, trulata (five hours old) reads, finds, updates and adds leads in a service-business CRM and lists published blog posts, and is held beside the CRM class this same sweep catalogues twice in Customermates and AXL, because a third CRM reader adds no new question. relaylink sends briefings between assistants, pidgeon and agent-inbox are agent mailboxes, odilem2m reaches a business by phone number and ask-a-human-expert sells paid human calls - all held in the communications and marketplace classes already carried by nextmailhq, wuapi and sendlettersonlinecouk. blaze and buzzmark are social scheduling and ghostwriting, held in the scheduling class this catalog carries heavily; cardboard-ai, pixelflow, drop-media, promptwise and agent-media sit in the media and media-generation classes, with cardboard-ai and pixelflow already dispositioned at midday; mythex, xhostd and aoci-code are codebase, hosting and deployment infrastructure; awardtape sells DoD and DHS contract award data at $0.02 a call, held as a paid public-dataset vertical; cp0x is DeFi referral data and cardnexus a payment-card comparison, held as crypto and finance-comparison; pugi grades a site's headers and TLS, held as agent-security infrastructure; testwell-lab-tests prices cash-pay lab tests, held as a healthcare vertical; lorekeeper is tabletop campaign lore, held as consumer; leviathan-systems is vendor marketing content. how-is, leadpages and more-ai-fortnox carried Oct 25 dispositions and were not re-read.

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: https://smithery.ai/server/corpusiq returns 308 to /servers/corpusiq, which answers HTTP 200.
• Glama: https://glama.ai/mcp/connectors/io.corpusiq/multi-source-mcp returns HTTP 200 on the connector route that replaced the moved 301 path.
• mcp.so: absent on the slug probed, HTTP 404.

Sources: mcpservers.org /all page 1 through the r.jina.ai reader proxy at the 04:01Z stamp (30 cards, counter 17,525 against 17,522 at the midday read, page one order unchanged), mcp.so /feed page 1 from the raw page with a browser user agent (30 cards, newest 2026-10-27T02:25:04Z, rotated cards at five to eleven hours old), the chatmcp/mcpso issue window #6177 to #6196 over the authenticated GitHub API (20 issues, no gaps), live JSON-RPC probes on four endpoints (all four returning 401 on initialize, one also returning a structured 405 with its own stateless explanation), RFC 9728 and RFC 8414 discovery documents read for all four, the official MCP registry for com.pickfu/mcp 1.0.0, io.github.idoo/app-store-connect-mcp at 0.13.0 and io.github.benjiwagner/customermates, the generated docs/TOOLS.md in the App Store Connect repository, the idoo/homebrew-tap repository answering 200, and the vendor documentation for PickFu, Customermates, Rus.sh and AXL.

Catalog 998 to 1,003 servers, guides 1,146 to 1,151.