Just ask "Is Portugal's internet fast?" and you'll know! Cloudflare's new AI tool "Radar Researcher" is here!

Hey, it's Shii! Today I want to tell you about a new tool I found on Cloudflare's Blog that lets you explore Internet data just by chatting with it!

Cloudflare Blog blog.cloudflare.com

What was announced?

Cloudflare's Blog announced a new AI tool added to Cloudflare Radar called "Radar Researcher." Radar is Cloudflare's dashboard where anyone can freely browse Internet data like DNS queries, HTTP traffic, and connection quality, and now it has gained an "AI researcher you can just ask questions in plain language."

The big point is that when you ask a question, you get back the same interactive charts you'd see on Radar's other pages, plus a short explanation, right on the spot. It's rolling out in beta across every page on Radar, and you can open it from the "Researcher" button in the header.

The story so far

Over the past six years, the Radar team has worked on making complex data, like DNS queries (sourced from 1.1.1.1), HTTP traffic across the global network, and connection quality data from Cloudflare Speed Test, clear and easy to visualize. All of that data has also been available through a free API, since Cloudflare cares about keeping this data open.

But until now, even if you had a question, you had to go find the right page and the right chart yourself. Take a journalist writing about an Internet outage, for example: they often don't have time to hunt for the perfect graph. And for engineers and network operators, going back to check the same data manually, over and over, was a real hassle too.

The idea of "you shouldn't need to know a dataset's structure or vocabulary to get a useful answer from it" is what led to Radar Researcher.

What changes

With Radar Researcher, you can just ask what you want to know, and you'll get an answer plus a chart built on real data pulled straight from Radar's API. Here's what it can do:

• Ask a question and get the same interactive charts you see elsewhere on Radar, plus a short explanation
• Choose between a concise, direct answer or a fuller, multi-topic report
• Get useful follow-up questions suggested after each answer
• Conversations are saved in a searchable history you can pin and revisit, and you can share any conversation via a link (shared links expire automatically after 30 days)
• Check how the LLM interpreted your question, which datasets it looked up, and how it worked through the results
• Ask by typing, using voice input, or launching it from Radar's search bar

On top of that, any existing chart on Radar now has an "Explain with AI" action, which opens Radar Researcher with that exact chart already handed over, so you can start chatting about it right away.

The post walks through some concrete examples too. Just ask something like "How good is home Internet in Portugal?" and you get a connection-quality chart back. Follow up casually with "How does that compare to its neighbors?" and it fetches fresh data each time. Radar Researcher even suggests its own next steps, like comparing against Spain or looking at common outage patterns.

There's also an example aimed at more technical readers: investigating the government-directed Internet shutdown in Iran in early 2026, which lasted nearly three months. Since that's an open-ended question, Radar Researcher runs a short investigation of its own: it looks up the recorded outage events and gathers the matching traffic data, then explains the timeline in plain language. In this case, Iran's HTTP traffic index collapsed from around 0.58 on January 7 to effectively zero by January 9, began a partial recovery around January 17, and returned close to pre-shutdown levels by January 27. It presents this as a traffic-index chart with the outage window annotated on the line, plus a table of recorded outages that also flags a second shutdown starting February 28.

Dive Deep

Let's take a closer look at how it's built. Radar Researcher is built entirely on Cloudflare's own developer platform, and it doubles as a showcase of the exact tools any developer can use to build their own AI agent.

Here's the base of the architecture:

• At its core is a Cloudflare Worker running the Cloudflare Agents SDK
• Each conversation is its own stateful Durable Object with its own SQLite database, so chat history, titles, and even streaming responses persist. Generation keeps going server-side even if you leave the page mid-answer, and it picks back up when you return
• The "brain" is Workers AI, running open models like Kimi K2.7
• Instead of betting on a single model, it runs an ordered fallback chain across three different model families, so if one model is momentarily at capacity, the request transparently cascades to the next
• Every call routes through AI Gateway, which handles logging, cost tracking, caching, and safety guardrails

How it reaches the data is one of the more interesting design choices. Radar has hundreds of endpoints, and rather than hand-writing a dedicated tool for each one, Researcher connects to a unified Cloudflare MCP server using Code Mode. The model only gets three tools: search, execute, and docs. It searches the OpenAPI spec to find the right endpoint, then writes and runs a small snippet of code to fetch live data. Since the full API spec lives on the MCP server, nothing about Radar's API is hard-coded into the prompt, so whenever Radar adds new datasets, Researcher can use them with no code changes.

How it turns a Markdown answer into a real chart is another interesting piece. Language models naturally answer in Markdown: text, tables, and bullet points. But Radar Researcher didn't want walls of numbers, and when a model tries to write data directly into its answer, it tends to round, summarize, or truncate it, which is exactly what you don't want from a data tool.

The fix keeps the data out of the model's prose entirely. When Researcher fetches data, the code returns a small envelope pairing the API path with the result. Instead of pasting numbers, the model emits a lightweight chart specification that just references that same path. It looks something like this (the model actually returns it inside a code block tagged "radar-chart"):

{ "type": "speedFlower", "title": "Internet speed quality — Portugal", "dataFrom": "/radar/quality/speed/summary?location=PT" }

Radar's frontend matches that dataFrom to the fetched result and renders it with the exact same visualization components used everywhere else on the site, so the chart stays faithful to the real API data, and Researcher can draw on Radar's whole visual vocabulary: time series, stacked areas, donuts, bar charts, maps, histograms, and more.

There are a few smaller touches too:

• Small, fast models handle side tasks like writing a short title for each new conversation or suggesting follow-up questions, so they never slow down the main reply
• It's given the current date and time, plus the same high-level connection details you'd find on Radar's IP page (like your IP location and network), so a question like "how's traffic here?" gets tailored to where you actually are
• Everything reaches you through Radar's frontend (also a Worker) over a service binding, with per-IP rate limiting, and shared conversations are stored in R2. Every layer, from compute to inference to gateway to storage to data, runs on Cloudflare

One more interesting piece: alongside Radar Researcher itself, there's support for general-purpose AI agents running in your own browser that come to operate Radar on your behalf. Until now, an agent like that had to scrape the page and guess how it works, which is brittle and slow. That's where the new "WebMCP" standard comes in: it lets a page register a predefined set of tools that a browser agent can call directly. Radar implements both flavors of it:

• The imperative API: tools registered in JavaScript that call straight into the same code that powers the UI. This lets an agent filter by country, region, continent, or ASN, change the date range, search pages, sections, and entities like domains, ASNs, or IPs, and jump to any section
• The declarative API: turns existing HTML forms into tools just by adding a couple of attributes, useful for things like scanning a URL with the URL Scanner, looking up a domain's report, or checking a website's post-quantum TLS key exchange support

Both are pure progressive enhancement, since they just drive functionality that's already in the UI. In a browser without WebMCP, the tools simply aren't there, and nothing else changes. Thanks to this, Radar now passes its own "agent-readiness" check too.

Wrap-up

Here's a recap of today's news:

• Cloudflare added a new beta AI tool called "Radar Researcher" to Radar, letting you explore Internet data just by asking questions in plain language
• It answers with the same interactive charts you see elsewhere on Radar, plus an explanation, and supports saving and sharing conversations, follow-up suggestions, and voice input
• Under the hood, it's built entirely on Cloudflare's developer platform: Cloudflare Workers, the Agents SDK, Durable Objects, Workers AI, AI Gateway, an MCP server (with Code Mode), and R2
• It keeps data out of the LLM's prose and has the model output a lightweight chart spec instead, which is a clever way to keep charts accurate
• It also added WebMCP support for general-purpose browser AI agents, aiming to make Radar easy for both people and agents to use

This announcement really shows a genuine desire to make Internet data easier to explore, and honestly, I want to go ask it about the internet situation in a bunch of different countries myself!