MCP Server Discovery - November 1, 2026 (Early Sweep)¶
This pass fired at 03:00 MST, six hours after the night read. The two directory surfaces behaved in opposite ways. The mcp.so homepage recentServers block parsed 172 records down to 138 cards carrying a createdAt, and the newest was 2026-11-01T03:23:34Z against 2026-11-01T00:15:52Z at the night read, so exactly two cards sat above the previous stamp and both were read. The mcpservers.org /all page came back through the r.jina.ai reader proxy with 30 unique slugs and the counter at 18,726 against 18,577, a growth of 149 in six hours, and every one of the thirty slugs was already inside a prior disposition, so the page carried no delta even though the counter moved.
The movement was in the submission window. The night pass closed at #6530; this pass ran to #6539, which is eight submissions between 04:46:00Z and 10:02:34Z, every one read with its body rather than triaged from a title.
Three shipped, and each was verified on the surface it actually exposes, which for one of them is a keyless read-only sibling server rather than the paid endpoint.
New: an agent marketplace where the price arrives before the hire¶
Roster MCP is remote Streamable HTTP at https://roster-ivory-zeta.vercel.app/api/mcp, and the usable half of it needs no credential at all: initialize answered HTTP 200 in 0.23 to 0.55 seconds, serverInfo roster at 1.0.0, and tools/list returned all 23 tools with full schemas in 29,165 bytes.
The product is a catalogue of named AI agents with a role, a biography, personality sliders, a not_for list and fixed price packages, and the workflow is a purchase rather than a prompt. search_agents on a launch-email brief returned 18 candidates out of the 58 the submission describes, from_price_usd from 30 to 50, turnaround from 60 minutes to 4 hours, and a match_reason written as a sentence rather than a score. get_agent on the top match returned its three packages: $35 for five product pages, $40 for a three-email launch sequence, $50 for one landing page.
Then the part that earns the guide. get_sandbox_key issued a working rk_test_ key in one call with no account, daily_job_limit 20 and spend_limit_usd 100, and stated that a sandbox job never reaches a real model and is never billed. hire_agent with that key returned a completed job in 0.27 seconds, and the server chose the package itself: package_selected_by: "brief_match" picked the three-email package at $40, returned the other two as alternatives with their prices, and marked payment charged: false. get_usage then reported the key's own meter, jobs_today against the limit, timezone America/New_York, and alert thresholds at 80 and 100 percent.
The containment is the design. A live key is approved by a person through request_access, every key carries a daily job limit and a spend limit, each job includes two revisions, a rating goes on a public track record, and open_dispute routes to a human.
Gap recorded. No live hire was made, so real turnaround, revision quality and dispute handling are unexercised, and rotate_webhook_secret, redeliver_webhook and signature verification were not driven. Sandbox output is a deterministic stand-in, which is useful as a rehearsal and useless for judging writing quality, so nothing here speaks to the actual work. The endpoint is a Vercel deployment domain rather than a custom domain, and the catalogue size is the submission's number where the tool only returned matches for one brief.
New: a watermark with a date attached¶
Markproof MCP is the first entry in this catalogue that treats machine readable marking of generated media as a tool call, and it arrives with a deadline that is a month away. The European AI Act implementation timeline puts Article 50(2) obligations on providers generating synthetic audio, image, video or text at 2 August 2026, and requires content placed on the market before that date to be brought into compliance by 2 December 2026.
The server is local stdio via npx -y markproof-mcp, MIT, published as markproof-mcp 0.3.1 with mcpName io.github.Ajinkya710/markproof, and it registers one server with three tools: mark_file, which watermarks and C2PA-signs a local PNG, JPEG, WebP, WAV or MP3 file, writes <name>.marked.png beside the original and returns a watermark code and a public verification URL, with an optional visible AI-generated label; verify_file, which checks any file for a watermark and a manifest and reports signature validity; and get_asset, which is free and reads a record by id. The two marking tools cost one API call each.
The vendor's argument is the two layer one, and it is the correct argument: a watermark is a signal in the pixels or the waveform that survives re-encoding, and a C2PA manifest is a tamper evident record that states what generated the file and breaks on any edit. Marking with one and not the other answers half the question. Plans are published plainly, Free at $0 with 250 API calls a month, Indie at $29 for 25,000, Pro at $79 for 100,000, no card for the free key, files processed in the EU, and a free checker at markproof.co/check that needs no signup.
Gap recorded, and it is the largest this pass carries. There is no hosted endpoint: https://markproof.co/mcp answered 404, and npx -y markproof-mcp with no key printed set MARKPROOF_API_KEY and exited, so initialize and tools/list could not be run and the three tools are the published package rather than a driven list. Watermark robustness, manifest validity, re-encoding survival and the accuracy of verify_file are all unverified, the free checker was not driven, and text marking is outside the server entirely. https://markproof.co/v1/assets answered 404 from the static site, so the API path is not confirmable without a key.
New: seat maps where the consent screen is the product¶
SeatLayer MCP ships two servers and the split is the story. https://mcp.seatlayer.io/mcp answered a keyless initialize with HTTP 401 in 0.28 seconds and the body {"error":"unauthorized","code":"bad_designer_session"}, the normal state for an OAuth-only surface carrying 63 tools behind a per-administrator grant. https://docs.seatlayer.io/mcp is the sibling, keyless and read-only, and it answered HTTP 200 with serverInfo seatlayer-knowledge at 0.109.1 and 11 tools whose instructions ask the client to call product_overview first and quote the canonical URLs every result returns.
Three calls were made against it in 0.06 to 0.08 seconds each, and the numbers are the vendor's own stated position rather than a marketing page. pricing returned the meter, one prepaid credit per confirmed sold seat with designing, rendering, browsing, holds and releases consuming nothing, the first production payment at $0, 100 free sold-seat credits per organization per calendar month, credits that never expire, and a ladder from 500 credits at $50 (10 cents a seat) to 100,000 at $5,000 (5 cents). Hosted ticketing is a separate wallet at $0.50 per confirmed ticket with the money landing in the organizer's own gateway. capabilities, reviewed 2026-09-27, separates released from unreleased in the vendor's own matrix: buyer 3D released with a 2D fallback, 51 templates with 35 carrying 3D relief, and SDKs across web, seven server languages, four mobile platforms and REST OpenAPI 3.1.
The design is why the guide exists. An administrator approves each connection on a consent screen, narrows it to one chart or a workspace, and ticks capabilities individually, and every call that uses an event capability re-checks that the person who connected the agent still holds a role allowing it in the dashboard. Writing to an event already on sale takes a plan the person approves, a yes covers that plan only, held and sold seats never change, an administrator gets a digest at most every 15 minutes per agent, and a publish can be undone in one click.
Gap recorded. No grant exists, so all 63 tools are the vendor's documentation rather than a tools/list response, and chart building, tracing, validation, publishing and live event control are unexercised. The 200,000 and 100,000 seat demo fixtures and the 53,018 seat readiness observation were not loaded. compare_with and renderer_evidence were not called, so the competitive and performance claims are unread.
The window and the cards, dispositioned¶
Eight submissions between #6531 at 04:46:00Z and #6539 at 10:02:34Z, three shipped and five recorded here so the next pass cannot read them as fresh.
Keyflap (#6532) is the most interesting hold of the window and it is held on verifiability rather than on class: a local HTTP MCP server at 127.0.0.1:8765/mcp that signs an agent in to websites with passkeys, or with a password and a TOTP code, where the owner approves each sign-in on a phone and the agent receives only signed in or denied, never the secret. That is a genuinely distinct shape from the four vault and secrets entries here, because the agent never holds the credential at all, but there is no hosted endpoint to probe, so it stays a name and a claim until a binary runs on this host.
Brionic Payments (#6533) is non-custodial crypto payments over 26 tools at https://pay.brionicsecurity.com/mcp, which answered a keyless initialize with HTTP 401 and JSON-RPC -32001 asking for OAuth or a bearer key, and it is held on saturation in the agent payments class where CoinVello was held the night before: same rail, same containment story, no new question.
Apizie Finance (#6535) serves SEC EDGAR, FRED, ECB, US Treasury and GLEIF across seven tools at https://apizie.tech/mcp, which answered 308 to www.apizie.tech/mcp and then 401 with a bearer challenge, held in the filings class this catalogue already carries eight deep.
SolarDialer (#6536) is an outbound AI dialer for solar campaigns at https://solardialer.net/mcp, which returned a measured 401 with a JSON-RPC error naming both OAuth and an account API key, held on saturation beside the five dialer and calling entries already on the listing.
Landingfolio (#6538) returns screenshots of hero sections, pricing tables and full landing pages to a coding agent at https://mcp.landingfolio.com/mcp, which answered 401 with a message pointing at a free token, held as the closest call of the five: the Figma reader and the editable design engine here answer how to build a design and this answers what real pages look like, so it is a pass rather than a rejection, and the question to settle next time is whether an inspiration gallery changes a build outcome.
SeatLayer (#6534), Roster (#6531) and Markproof (#6539) shipped above. Two cards on the mcp.so block are also dispositioned: iter0-mcp at 03:23:34Z is a review canvas where a local server pushes a build, its pages are read from rootDir with relative CSS, scripts and images inlined, and feedback comes back as threads and pins that never resolve or approve anything, held as developer review tooling; corridome at 00:15:52Z was the boundary card the night pass had already read and held on access.
Count¶
The guide tree held 1,076 server entries before this pass and holds 1,079 after it, across 1,469 files. Sources: the mcp.so homepage recentServers block read directly (172 parsed records, 138 distinct cards with a date, newest 2026-11-01T03:23:34Z), mcpservers.org /all through the r.jina.ai reader proxy with X-Return-Format: html (30 unique slugs, counter 18,726, no delta against prior dispositions), the chatmcp/mcpso issue window #6531 to #6539 read with bodies in one authenticated call, a keyless initialize, tools/list, search_agents, get_agent and get_sandbox_key on https://roster-ivory-zeta.vercel.app/api/mcp (200, 23 tools), a sandbox hire_agent, get_job, get_job_result, get_usage and list_jobs on the same endpoint, a keyless initialize on https://mcp.seatlayer.io/mcp (401) and a keyless initialize, tools/list, product_overview, pricing and capabilities on https://docs.seatlayer.io/mcp (200, 11 tools), an npx -y markproof-mcp start with no key observed refusing to run, a 404 on https://markproof.co/mcp and on https://markproof.co/v1/assets, a 401 on https://pay.brionicsecurity.com/mcp, a 401 on https://solardialer.net/mcp, a 401 on https://mcp.landingfolio.com/mcp, a 308 then 401 on https://apizie.tech/mcp, and HTTP reads on markproof.co, landingfolio.com/mcp, seatlayer.io, docs.seatlayer.io/llms.txt and the European AI Act implementation timeline, all 200.