Is MCP becoming the common foundation of smart, decentralised search?
An article on contextetech.com asks whether the Model Context Protocol (MCP) is becoming the common foundation of smart, decentralised search. It appeared on the site and was discussed on Hacker News. The piece explains how MCP lets AI assistants fetch data directly from publishers, and why the protocol’s design leaves citation, trust, and measurement to the client side.
The article describes MCP as an open protocol that lets an AI assistant call tools provided by a server: search, read a resource, fill in a text template. It says the same server works with Claude, ChatGPT, Cursor, VS Code and any assistant that speaks MCP. The protocol was launched by Anthropic in November 2024, adopted within months by the main assistants and developer tools, then handed over in December 2025 to the Agentic AI Foundation, a foundation of the Linux Foundation. It evolves through dated versions, and a public registry lists the servers. The article contrasts classic search engines, which copy the web into an index and rank what they copied, with MCP, where each publisher keeps its data at home and exposes it through its own server; the assistant queries several servers on demand, at the moment of the question; the answer comes with the source, the licence, the author, the date of the last update. It says nobody needs to copy everything: search becomes decentralised. A statistics office, a public administration, a laboratory or a small specialised website can answer assistants directly, with up-to-date figures, without going through an intermediate index. The article lists what it changes for users: up-to-date answers, verifiable sources, clear rights, less invention. It says decentralised search is only worth something if it stays honest. Three things still have to be built, server by server: citation, trust, measurement. The article gives the Contexte Tech catalogue as an example server, open to every assistant at https://contextetech.com/mcp, with no account, no key. It describes how search works: the assistant finds resources that match the question, in 9 languages, with their title, licence, author and link. Reading: it loads a resource by its short code. Required citation: every loaded resource ends with a Source line that the server’s instructions require the assistant to repeat: the resource, its author, the original data source, the publisher and the licence, each with its link. Measurement without tracking: the links tell which assistant the visit comes from, with no personal data. Protected content: what members write is flagged to the assistant as data, and resources are checked when they are published. The article invites readers to try it: add https://contextetech.com/mcp as a connector in your assistant, then ask: load n8qhje. The assistant loads a dataset for finding the NAF (APE) code of a business activity, built on the official INSEE classification published on data.gouv.fr, and cites its source. The full guide is on the MCP server page. The article asks what if the MCP protocol became the common foundation of smart, decentralised search? It says MCP does not replace the web: it gives assistants a way in. The more open, sourced and honest servers there are, the more AI search will look like what we expect from it: answers that say where they come from. It concludes: if you publish data or reference content, now is the right time to open your own server.
The system involves an AI assistant with an MCP client library, a public registry of server endpoints, and individual MCP servers run by publishers. When a user asks a question, the assistant’s client library uses the registry to find relevant server URLs. It sends a search request containing the natural-language query to one or more servers. Each server processes the query against its own data and returns a list of matching resources, each with metadata such as title, licence, author, original source, publisher, link, and date of last update. The client library aggregates these results and presents them to the assistant. The assistant selects a resource and issues a read request with the resource’s short code. The server returns the raw content of the resource along with the same metadata fields. The client library formats these fields into a Source line that the assistant includes in its final answer, as required by the server’s instructions. The server does not verify the truthfulness of the content; it only returns data as published. Trust in the content is therefore a function of the client, which must decide whether to trust the server based on reputation or other signals. Measurement of how many times a server is consulted is not built into the protocol; a server can include a referral identifier in returned links to infer which assistant made the request, but it does not receive aggregated usage statistics unless the client voluntarily reports them, a design choice made to avoid collecting personal data.
The design decision to treat servers as stateless data providers and to delegate citation rendering, trust assessment, and usage measurement to the client creates a trade-off. Decentralisation removes the need for a central index, reducing storage and update burdens for publishers. However, it shifts the responsibility for honest attribution and quality control to each client implementation. Likewise, without a standardised trust mechanism, a malicious assistant could present fabricated data without detection at the server level. Without a built-in, privacy-preserving measurement mechanism, publishers lack feedback on how often their servers are used, which may reduce their incentive to maintain high-quality endpoints. The Hacker News comments highlight these exact points: citation is seen as a client-side job because the server only provides metadata; trust is seen as a client function because the server cannot vouch for content; measurement is debated as something that might be better done outside the server to preserve privacy.
The open question is whether the ecosystem will develop shared client-side practices for citation, trust, and measurement, or whether fragmentation in client implementations will erode the protocol’s goal of uniform, verifiable, up-to-date answers from decentralised sources.