TensorPass

Use cases

An answer engine on a news site, run on the reader's own AI

How a publisher answers readers from its own archive with each reader's own AI, which MCP tools it serves and how subscriptions and clicks stay with it.

Published on

In 2026, 10% of people use AI chatbots for news each week, up from 7% a year earlier, and 16% of under 35s do. Of these users, 42% say they always or often click through to the original sources. Google organic traffic to more than 2,500 news sites fell 33% between November 2024 and November 2025, and 38% in the US. AI Connect, the TensorPass SDK a website installs, puts the answer engine on the publisher's own pages. The reader's model searches the archive through the publisher's tools and answers on the reader's own account through AI Wallet, so the publisher pays no inference.

Answer layers publishers run in 2026

Trust in answers from AI chatbots stands at 20%, against 37% for news overall. Four newsrooms answer readers from their own reporting on their own pages:

PublisherAnswer layerPublished fact
USA TodayTaboola DeeperDiveMore than 25 million questions, about 1 million a week
Financial TimesAsk FTOpen to all FT Professional subscribers since April 2025, answering only from FT content
TimeAI agent built with Scale AIAnswers from about 750,000 archive assets, launched in November 2025
Arc XP publishersAnswer layer embedded on the publisher's siteLaunched in July 2026 on the platform the Washington Post built

Up to 1 in 6 USA Today visitors use DeeperDive. Its adoption rose from under 1% to 15 to 20% once suggested prompts were added, and over 10% of its users click on to another article.

Reader questions an archive answers

Each question below maps to a tool call on the publisher's server:

Archive tools served over MCP

The Model Context Protocol (MCP) is the open standard through which a model calls a site's tools. The AI Connect WordPress plugin serves two of them at /wp-json/aiconnect/v1/mcp: search_posts returns up to 10 matches with their id, title, address, date and excerpt, and get_post returns one post with its text, up to 8,000 characters.

A publisher on its own CMS writes the tools its archive needs:

ToolInputWhat it returns
list_sectionsNoneThe sections, topics and authors the archive accepts as filters
search_articlesWords, section, author, date range, limitThe matching articles, each with its id, headline, date, excerpt and page address
get_articleArticle idThe article's text for a reader without a subscription, with its date and page address

The date on every result lets the model say when a story ran. The MCP guide sets the other rules: every input bounded, every filter value explained, and an empty result that tells the model to relax one filter.

Which interface suits a newsroom?

The SDK ships five ready made interface modules. Three of them suit a news site:

Subscriptions, paywalls and clicks back

Anyone reaches an MCP server without the publisher's page, so the MCP server enforces the paywall. The WordPress plugin returns published posts and pages of the public post types, and leaves out drafts, private posts and posts behind a password. A publisher's own get_article returns the excerpt of a subscriber article. The guards add withholdPatterns to keep content out of an answer, and the server holds back anything a reader must not reach.

The SDK's answer guard lets a link through when it points to the page's own origin, to a host the publisher lists, or to an address a tool returned as a field. A reader who clicks a link in an answer lands on the publisher's own pages or on a host the publisher listed. The subscription page stays on the publisher's site.

Reader comments are third party text. The SDK closes every tool result in an envelope sealed with a random salt drawn for each conversation, and instructions planted in a comment stay inside it as data. This is the guard against prompt injection.

Next steps for a publisher

On 16 November 2026 the SDK packages of version 0.1.0 are on npm, and AI Wallet waits on the Chrome Web Store's review while a visitor can install it by hand from https://tensorpass.app/wallet. A reader without AI Wallet gets the answer engine through a sign in with OpenRouter, or with the TensorPass plan on a site that passes a site key.

Sources

Questions

Who pays for the answers a publisher's answer engine gives?
Each answer runs on the reader's own ChatGPT, Claude or Gemini account through AI Wallet and counts against that account's plan or credits. The publisher pays for its MCP server and its hosting.
Does the reader's AI read articles behind the paywall?
The model reads only what the MCP server returns. The WordPress plugin returns published posts of the public post types without a password, and a publisher's own server returns an excerpt for any article it keeps for subscribers.
Where do the links in an answer lead?
The answer guard lets a link through when it points to the publisher's own origin, to a host the publisher lists, or to an address a tool returned as a field. The links of an answer therefore lead to the publisher's own articles.
What does a reader need to use the answer engine?
Chrome with the AI Wallet extension and one AI account connected in it. On 12 November 2026 the wallet loads unpacked and its Chrome Web Store item is a draft that has not been submitted.
Have it done in 24 hoursRead the quickstart