As of 16 October 2026, 124 of 140 products we drove publish an MCP server, and our agent got in on its own at 52 of 140. A product can publish everything an agent needs to read and still stop the agent at sign-up, verification or the key.
Having an MCP server didn’t open the door
76 of 124 products with an MCP server stopped our agent: it did not get in on its own. Across all the products we drove, sign-up or its verification stopped it at 84 of 140.
Our drives of 140 products, 1–16 Oct 2026 · The data
Data table
| Gets in on its own | Stopped at the door | Total | |
|---|---|---|---|
| With MCP | 48 | 76 | 124 |
ActiveCampaign publishes an official MCP server and an auth.md that says agents cannot create accounts or start trials. Its trial form leads to a visible Cloudflare Turnstile challenge, and a signed request without a browser got a 403 with cf-mitigated: challenge. At Canva, which also has an official MCP server, the Cloudflare interstitial on its sign-up page never cleared for our agent’s browser. Brave Search API, which also has an MCP server, stopped our agent later, at the key.
A bot check stood in the way at 75 of 88 products that stopped our agent. Signed, not verified: during these drives no bot-check provider listed our agent as a verified bot, so a check could see the signature but had no verified identity to trust. Which bot checks stop AI agents breaks these stops down by provider.
For the Cloudflare interstitial at Canva, a builder’s step is registration: Cloudflare has folded signed agents into its Verified bots since 1 July 2026, with a Direct or Intermediary label, and a builder registers a signed agent with Cloudflare through its bot submission form (Cloudflare docs).
The sign-up check and the key hide walls no spec lists
The key or the first call stopped our agent at 1 of 140 products we drove: Brave Search API. Its sign-up passed Brave’s own check in our agent’s browser. Brave mints keys in its dashboard alone; there is no key API or CLI. Every plan, the free credits included, asks for a card before it activates. The dashboard-only key and the card before any plan are the steps Brave’s owner can change on the way to the key.
At 2 of 140 products, our agent got in only with a person’s help: a person cleared the bot check at sign-up (how we drive), then our agent took over. Backblaze B2 shows a Google reCAPTCHA checkbox on its own Security Check page, and Invoice Ninja’s register form runs Cloudflare Turnstile. Past that, Backblaze B2 makes application keys in its web console or by an API call, each needing an account. Invoice Ninja documents a login route that trades an email and password for an API token.
Our drives of 140 products, 1–16 Oct 2026 · The data
Data table
| Products | Out of | Share | |
|---|---|---|---|
| Got in only with a person's help | 2 | 140 | 1% |
The owner of the Google reCAPTCHA wall at Backblaze B2 can turn to Google Cloud Fraud Defense, reCAPTCHA’s successor, which Google is integrating with Web Bot Auth (Google Cloud blog). On any site behind Cloudflare, a rule can match agents that sign with Web Bot Auth on the field cf.bot_management.signed_agent.
Neon, Railway and Apify hand an agent a credential through a documented route
Neon documents a claimable database in its auth.md: neon claim create returns a claim token, then an access token and a database URL. Our agent created a table and rows on it and did CRUD through the Data API. Railway’s anonymous box at ssh railway.new carries a project token our agent used with the GraphQL API and the CLI. At Apify, an x402 payment from our side came back as an API token with a spending cap. Builders can call these routes today.
On recordNeonDatabasesL2Gets inOther products let our agent sign up with no challenge shown. At Webflow, the one product where our agent met HUMAN Security, the check let its browser through. At Plausible Analytics, the one product where it met Friendly Captcha, and at Zulip, where ALTCHA ran, the check let our agent’s browser through as well. SignWell and PlanetScale let it sign up without a browser.
Our drives of 140 products, 1–16 Oct 2026 · The data
Data table
| Products | Out of | Share | |
|---|---|---|---|
| Got in on its own | 52 | 140 | 37% |
If your sign-up is a wall, document a route like Neon’s that gives an agent a working credential without an account. If you run HUMAN, you set whether an agent may sign up or log in at Sightline Cyberfraud Defense > Policies > AI Agents Permissions: Account Creation and Login, per agent and application. HUMAN assigns a High trust level to an agent it can verify by a cryptographic key on each request (HUMAN docs).
Of the products our agent got into on its own, 2 of 52 got only as far as the dashboard, in a browser: TidyCal and Umami offer API access on paid plans alone. Before you wire a product into your agent, the free check shows what it publishes for an agent: its docs, API, OpenAPI spec, MCP server and llms.txt.
