MCP Server Discovery - October 21, 2026 (Evening Sweep)¶
Evening delta eight hours after the midday supplement. The two surfaces behaved differently, and that is the finding worth recording: mcp.so /feed page 1 returned thirty cards and every one of them had already been read and dispositioned in a prior cycle, while mcpservers.org /all newest-first carried two business-relevant slugs that no sweep record contained.
That is the second time this week the feed has repeated itself at a short same-day gap while /all carried the fresh material. The working rule stands: at gaps under about twelve hours, read /all pages 1 to 3 and treat /feed page 1 as a re-read of prior dispositions rather than a source of new cards.
New Servers Catalogued¶
| Server | Category | What it does |
|---|---|---|
| UniPost MCP | Marketing | LinkedIn profiles, X and TikTok scheduling from an agent, with scheduling and approving as separate tools so nothing publishes without a person. https://mcp.unipost.ai/mcp, OAuth or workspace API key. Probed live: unauthenticated POST initialize returned HTTP 401 with the server's own message, {"error":{"code":-32001,"message":"Sign in to UniPost, or create an API key in UniPost → API keys."}}. Seven tools named in the vendor's reference: list_accounts, list_posts, schedule_post, approve_post, cancel_post, get_analytics, get_x_allowance. schedule_post always creates pending_approval and returns an approval link; a post created by another person's agent cannot be approved through the tool at all, and TikTok posts must be approved on the vendor screen because their privacy settings live there. Pricing is per account and graduated, two accounts free, $3 for accounts 3 to 10, $1 for 11 to 30 and $0.50 above, with yearly at ten times the monthly base. The X credit model is the operationally important part: a plain post costs 1 credit, a post containing a link costs 13, the free allowance is 10 a month and a paid workspace gets 50 per X account for its first ten and 20 after. Packaged addition, not a gap: the publishing class here is six deep already, carried by PendPost, TimeToPost, BulkPublish, Blog2Social, Chirpie and Maeve Social. Recorded for the pricing shape and the two-tool approval split. |
| Agents Handbook MCP | Productivity | A keyless playbook corpus for agent-run jobs, with evidence grades and human checkpoints as data rather than prose. https://agentshandbook.com/mcp, no sign-in, no key. Probed live and driven, not just handshaked: POST initialize returned HTTP 200 with serverInfo agentshandbook 0.1.0 on protocol 2025-06-18 and server instructions, tools/list returned 11 tools counted rather than claimed, and a real tools/call on search_playbooks for "reconcile receipts against bank transactions" at min_trust: corroborated returned one playbook, match-receipts-to-bank-transactions-for-your-accountant, graded corroborated and high risk, with the four agents it is written for plus page and markdown URLs. Every playbook carries a trust level from reported through corroborated, tested and confirmed, a risk level, and a list of human checkpoints defined as the steps that spend money, send messages, submit forms or delete data. Failures live in a separate get_lessons surface, and report_run writes back a worked, worked-with-changes or failed result into a review queue rather than straight into the record. 54 playbooks across ten categories, markdown twin on every URL, plain search endpoint at /api/v1/search?q=. Reverses a title-level hold: the midday supplement held this as an agent-playbook knowledge surface after reading it at title level. Read to detail, it earns an entry on mechanism rather than topic, because the checkpoint and evidence-grading shape is the closest live analogue to the governance contract this docs set publishes. |
Sources¶
- mcpservers.org
/allpages 1 to 3, newest first, through direct extraction and the r.jina.ai reader proxy, read in full and cross-referenced slug by slug against every sweep record from July 19 to October 21 - mcp.so
/feedpage 1, thirty cards, all thirty matched to prior dispositions; zero new cards at this gap - Vendor documentation read to detail: unipost.ai tool reference, plans and billing, usage and media limits, and
llms.txt; agentshandbook.com landing page,skill.mddescription, and the live MCP tool schemas forsearch_playbooks,get_playbookandreport_run - Live JSON-RPC probes, unauthenticated and recorded with status:
mcp.unipost.ai(401, server message quoted),agentshandbook.com(200 initialize, 11 tools, one realtools/call),lead-gen-system.upapi.io/api/mcp(401missing authorization header),lead-gen-system.upapi.io/settings/mcp(307 redirect to login, confirming the listing URL is a dashboard page rather than the endpoint) - Not run this cycle, and not claimed: the twelve-sitemap canonical refresh, the Glama, Smithery and PulseMCP listing probes, and
tools/liston any server that 401s without a credential
Also Identified (Not Catalogued)¶
Lead Gen Marketing Intelligence, at https://lead-gen-system.upapi.io/api/mcp, held on documentation rather than on product. The directory description promises ten read-only tools for interactions, outcomes, comment liveness, draft pipeline and platform health across nine social platforms, OAuth 2.1 protected, which is the cross-platform measurement surface this catalog has no entry for. The endpoint is live and gated, returning HTTP 401 missing authorization header to an unauthenticated initialize, but /llms.txt serves the sign-in screen, the root serves the sign-in screen, and the listing URL /settings/mcp is a dashboard page that 307s to login. With no public tool names and no published pricing, an entry would be the vendor's own marketing text with a status code behind it. Held until a public reference appears or a credentialed read names the tools.
Polign Recall was read this cycle because it surfaced on /all page 2 as a working card after the midday pass recorded it as a 404 shell. It is Apache 2.0, local or self-hosted, and its distinguishing idea is corrections as records: a new value replaces the old by a stated rule, the answer arrives with what it replaced, and as_of, history and diff answer what the agent believed on any date. Held against the memory class already carried here, which is dense, and noted because the corrections-as-records model is the same shape as the trust contract this catalog argues for.
match.bot is a free agent meeting place with start_here, register and search open without a key, and it is held as novelty infrastructure with no operator data in it. UAIN Research Pilot prepares evidence-backed research work for independent checking and is held as a research marketplace at pilot stage. Waldo drives Mac accessibility controls and asks before money, messages, passwords or deletes, and is held as desktop automation infrastructure on the ShadowCursor and Zeptap precedent. Arquimedes samples a million competition math problems and is held as a study corpus. POEditor is a localisation management surface, held as developer tooling. Twenty-six further /all cards on pages 1 to 3 carried their prior dispositions without re-litigation, including FxSocket, Zeptap, Shopify Speed Check, Next Everything, UXArmy, ReportGem, Rebbel, Underanchor, Geonode, Lium, Hazelchat, Weir, Pear, DomainRef, Open Index, Layrcake, Wine Labs, TwinMind, Visualping, Fettle, ScBee, LLMSlim, PedalBro, smsverbot, BEON PASS and Grocery Navigator.
From the feed, all carried. DolphinRadar, Vectoree, SCDO Agent Ledger, ReportGem, OwAI, botchelor, Volley.sh, paprika3d, 10ML Perfume Finder, FrontRowTickets, AppCreator24, Dreamwork, ARVX, Status Is Down, Veltake Studio, Searcherries, Caimera, NeuralVerge, golfdata.online, tropia.pl, PopNow, Thordata, HousingFeed, QRStuff, Captapi, Cowboy MCP and Kalender Sync each matched an existing sweep record on a slug or name search and kept that disposition. The feed repeated its page 1 in full inside eight hours, which is recorded above as a source behaviour rather than as a content finding.