Our research is now directly queryable by AI systems
An open standard, a free server, and a bet on getting cited correctly.
We've learned over the past years how fast models are becoming the interface through which people access information, including in places where the open web is degraded or blocked.
Whoever asks a model about Iran's shutdowns gets an answer shaped by what the model absorbed in training, months or years ago, compressed and unattributed.
We decided to act on our own conclusion. Our research is now directly queryable by AI assistants through an MCP server at mcp.gazzetta.xyz.
MCP (Model Context Protocol) is an open standard that lets AI assistants query external sources live instead of relying on training data.
If you use Claude, you can connect our server in three steps: open Settings, then Connectors, click "Add custom connector" and paste https://mcp.gazzetta.xyz/mcp. It also works in ChatGPT, Claude Code, Cursor and other MCP clients;
Once connected, the assistant can search our publications, read full reports, and retrieve our frameworks in our own words. Ask it how Iranians stayed online during the June 2025 shutdown and it will pull our survey findings with the publication date and a link, rather than paraphrasing from memory.
Why we built it
First, representation. When a model summarizes our helper networks research from training data, we have no say in what survives the compression. The server changes that: it returns our canonical articulations, marked as such, with instructions to cite the source. Whether models follow those instructions is itself an empirical question.
Second, observation. The server logs which tools get called and what queries arrive, nothing about who is asking. Over the coming months this gives us a small dataset on how AI systems actually interrogate a research corpus: what they search for, how they phrase it, which publications they retrieve. We study AI systems as information intermediaries; this is a chance to watch the intermediation happen to our own work.
Third, cost. The honest version of "how hard was this" is minimal time, a Cloudflare free tier, and a few hundred lines of code. That number matters for other research organizations and newsrooms deciding what to do about AI access to their work. If you run one and want to compare notes, write to us: hello@gazzetta.xyz.
The population of people who connect custom MCP servers to their AI assistant is small today. We do not expect meaningful traffic soon, and that is fine; this is an experiment and an instrument, not a strategy. We also do not know yet whether models will handle our material faithfully even with direct access.
The publications on this site remain the canonical versions. The server exists so that when a machine answers on our behalf, it has the chance to get it right.
Connect at mcp.gazzetta.xyz. If your AI is going to talk about us anyway, it might as well ask first.