Project Envoy

As readers shift from web pages to conversations with AI systems like ChatGPT and Gemini, publishers are at risk of becoming invisible sources of raw material that AI companies repackage without credit or access to the customer. Project Envoy is a prototype that explores new ways for newsrooms to publish their journalism directly into the agentic web in order to meet people where they are increasingly looking for information and answers.

Envoy is a system that lets each publisher run its own agent that speaks as and for the newsroom. It answers questions from readers and other AI systems on the publisher’s terms, and it carries with it the newsroom’s reporting, editorial judgment, and engagement strategies.

It connects in two directions: downward to its underlying data through the Model Context Protocol (MCP), which is how consumer assistants reach it, and sideways to other agents through the Agent-to-Agent protocol (A2A), which lets it join a wider network of newsroom and civic-data agents. A reader’s own AI can discover the agent, get answers, subscribe to the host publication and take actions without ever visiting a website.

Here’s a Quick Tour

The prototype news application, headlined Search NYC's 1,500+ Public Schools, with a search box and a search-by-address form.

The News Application

The working prototype is built on New York City public school data and published under The Data Tribune, a fictional news site made for the experiment. It has a traditional “news application” interactive data front end, which lets users look up a school by name, neighborhood or address. Try it at nycschools.datatribune.io.

The WebMCP Inspector connected to nycschools.datatribune.io, listing declarative tools and showing a find-zoned-schools-by-address call returning a result.

An MCP Server

Envoy also publishes an MCP server, which supports WebMCP, a standard soon to be included in Chrome. The MCP server supports composable data queries as well as predetermined interaction paths.

A LibreChat conversation with Project Envoy. A reader asks which school is zoned for 715 Carroll St., Brooklyn; the agent asks for a free registration, then names P.S. 321 William Penn and explains District 15 middle-school choice. The reader mentions they speak Spanish at home; the agent offers to continue in Spanish, and when the reader asks in Spanish about a tour, it replies in Spanish to book one.

An Agent That Answers Other Agents

Envoy also publishes an agent that can answer questions from other agents using the A2A protocol. Generative AI systems that support A2A can connect to Envoy and ask questions, and even take action, on a user’s behalf.

Try it in a chat app at envoy.datatribune.io, by invitation.

A reader writes: We speak Spanish at home. How is the ESL program? The agent replies in Spanish and in English, asking whether the reader would like it to continue in Spanish.

The Reader’s Language

The agent answers in the reader’s language. It not only supports the twelve languages New York City’s Department of Education translates into, it should work with any language our AI model supports. If a reader mentions they would be more comfortable in one of them, it offers to switch.

Three Agents

The proof-of-concept implementation includes three mock agents: one that uses the Envoy MCP server to query the underlying data, one that is a prototype of an agentic subscriptions interface, and one that imagines an external agent that can book school tours. The recording shows them at work on simulated traffic.

An outside AI agent asked which school 33 Irving Place is zoned for; it authorizes a subscription with the prototype publication and reports the zoning details.

The Future

As the agentic web emerges, Generative UI surfaces that support A2A will be able to connect to Envoy, either directly or via as-yet uninvented discovery mechanisms.

Project Envoy is an agentic news experiment by Scott Klein and Duy Nguyen.