MCP servers could decentralise AI search by querying publishers instead of crawled indexes

An essay published by Contexte Tech on October 17, 2026, and later surfaced on the Hacker News front page, argues that the Model Context Protocol (MCP) is emerging as a common substrate for AI-assisted search that is decentralised by design. The site, which publishes a catalogue of datasets, prompts and AI tools, frames the protocol as the missing link between assistants and the people who actually publish information.

How MCP works, in brief

According to the article, MCP is an open specification through which a server exposes tools — search, resource reading, template filling — that an AI assistant can call. Because the interface is standardised, one server works across every client that speaks the protocol; Claude, ChatGPT, Cursor and VS Code are named as examples. A publisher therefore builds a single integration rather than one per assistant.

Anthropic introduced the protocol in November 2024, Contexte Tech writes, and mainstream assistants and developer tools adopted it within months. Stewardship passed to the Agentic AI Foundation, which sits under the Linux Foundation, in December 2025, and the specification now advances through dated releases — five between November 2024 and July 2026 — alongside a public registry of available servers.

From index to live query

The essay's central claim concerns search architecture. A conventional engine copies the web into a private index and ranks that copy. Under the MCP model the direction inverts: publishers keep their data in place and expose it through their own servers, assistants query several of those servers at the moment a question is asked, and results arrive with provenance attached — source, licence, author and last-updated date.

Because no single party needs to mirror everything, the author argues, search becomes decentralised by construction. A statistics office, a public administration, a laboratory or a small specialist site can answer assistants directly, serving current figures without passing through an intermediary index.

What users would gain

Contexte Tech lists four practical benefits. Answers would be fresh, read from the publisher at question time rather than recalled from a model's training corpus. Every answer could link back to its original page, making sources verifiable. Where a server declares a licence, that licence travels with the data, clarifying what a user may legitimately do with it. And when a server simply does not hold the requested information it can say so, shrinking the gap a model might otherwise fill with invention — though the author concedes nothing fully prevents fabrication.

Three problems still open

Decentralised search is only worth something if it stays honest, the article continues, naming three gaps that must be closed server by server. Citation: assistants must surface where information came from, down to the author, the original source and the publisher. Trust: whatever a server returns must be treated strictly as data, never as instructions for the assistant — an acknowledgment of prompt-injection risk. Measurement: publishers need a way to see how much traffic assistants send them, without tracking individuals.

A reference implementation

Contexte Tech backs its argument with a server of its own, open to any assistant with no sign-up or API key. It exposes the site's catalogue, built largely on European public sources. The server supports search across nine languages, lets assistants fetch resources by short code, and appends a mandatory source line — resource, author, original data source, publisher and licence — that its instructions require the assistant to repeat. Visits carry a signal identifying which assistant sent them but no personal data, and member-written content is flagged as data and checked before publication. As a worked example, loading one short code returns a dataset that maps a business activity to its French NAF (APE) code, built on INSEE's official classification published on data.gouv.fr, with the citation attached.

Why it matters

The thesis is one publisher's argument, made alongside a plug for its own server, but the mechanism it describes already exists: MCP is standardised, foundation-governed and supported by the major assistants. If adoption follows the pattern the article sketches, AI search shifts from a handful of companies that crawl and rank copies of the web toward a protocol any publisher can implement, with provenance and licensing built in. That would also move the goalposts for publishers: rather than optimising pages for crawlers, the meaningful act becomes operating a server that assistants can query. The hard problems to solve are the ones the article itself names — reliable citation, resistance to injected instructions, and privacy-preserving measurement — and they are the same ones the wider agent ecosystem is currently wrestling with.