A plain-language window into Cloudflare Radar’s data
Cloudflare Radar has published open Internet traffic data at a global scale since 2020, drawing on DNS queries from the 1.1.1.1 resolver, HTTP traffic measurements from across Cloudflare’s network, and connection-quality data collected via Cloudflare’s Speed Test. The audience for that data spans journalists, human-rights advocates, academic researchers, and network operators—many of whom don’t have the time or need to master Radar’s full catalog, its page layouts, or its API documentation to get an answer.
As part of Agents Week, Cloudflare is beta-launching Radar Researcher, a tool that lets you ask questions about Internet conditions in plain language and receive real, interactive Radar charts in response. No hunting through pages, no manually setting filters, no writing API queries by hand.
How Radar Researcher works
Radar Researcher is available from every Radar page, docking as a side panel so you can keep browsing while you chat. At its core, the assistant translates your natural-language question into concrete API calls, fetches live data from Radar’s public API, and responds with a short plain-language explanation alongside the same interactive visualizations used across Radar. The underlying model never invents numbers because it relies on actual API responses rather than estimating from its training data.
A few features distinguish the experience:
- Explain any chart: A new Explain with AI action appears on every Radar visualization. Clicking it opens the assistant already holding three pieces of context: a screenshot of the chart for the vision-capable model, the exact underlying API data, and the current view’s parameters (location, date range, and filters). This means explanations reference the precise view you’re seeing rather than a generic version of the chart.
- Depth control: Choose between a concise answer or a more detailed, multi-topic report.
- Follow-up suggestions: After each response, the tool proposes useful next questions to keep the investigation moving.
- Persistent conversations: Chat history is saved in a searchable list, can be pinned for later reference, and can be shared via a link that expires after 30 days.
- Auditable reasoning: Every answer includes an expandable trace of the model’s tool calls, showing which datasets it queried and how it reached its conclusion.
- Flexible input: Type, use voice input, or launch Researcher directly from Radar’s search bar.
Example: assessing Internet quality in Portugal
To see how it behaves, consider a simple question about home Internet performance in a specific country. Ask “What’s Internet quality like in Portugal?” and Radar Researcher will interpret the question, query Radar’s Internet quality API, and return an answer with the same interactive charts found on Radar’s country-level Internet Quality section—all without any manual API calls or page navigation.
Of course, this type of data work rarely ends with one answer. A natural follow-up might compare Portugal with a neighboring country, at which point Researcher fetches fresh data for the comparison. It also suggests its own next steps, such as how Portugal stacks up against Spain or what the most common outages there have been.
Example: investigating a national Internet shutdown
The tool is also designed for more technical users, particularly those who need to track evolving situations. Cloudflare cites a real early-2026 event in Iran, which experienced a series of government-directed shutdowns, including one lasting nearly three months before traffic partially returned in May. Reporting that story required pulling multiple views and continuously refreshing them as the situation developed.
With Radar Researcher, an open-ended prompt like “investigate Internet disruptions in Iran” triggers a short automated investigation. The assistant looks up outage events recorded by Cloudflare Radar for Iran, gathers the matching traffic data, and explains the timeline. In Cloudflare’s example, 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 near pre-shutdown levels by January 27. The result includes an HTTP traffic index chart with the outage window annotated on the line plus a table of recorded outages flagging a second shutdown starting February 28. From there, the tool suggests next steps, like comparing the traffic pattern with neighboring countries.
Under the hood: an all-Cloudflare architecture
Radar Researcher is built entirely on Cloudflare’s developer platform, using the same building blocks available to any developer constructing an AI agent.
- Core state: A Cloudflare Worker runs the Cloudflare Agents SDK. Each conversation is a stateful Durable Object with its own SQLite database, preserving chat history, titles, and streaming responses—even across page reloads, since generation continues server-side and resumes upon reconnecting.
- Model layer: The “brain” runs on Workers AI via open models like Kimi K2.7. Rather than depending on a single model, the system uses an ordered fallback chain across three model families, automatically cascading requests to the next provider if one is at capacity. Every call routes through AI Gateway for logging, cost tracking, caching, and safety guardrails.
- Data access: Rather than hand-coding a tool for each of Radar’s hundreds of endpoints, the agent connects to the unified Cloudflare MCP server using Code Mode. This gives the model three tools—search, execute, and docs—and lets it write code to query the Radar API directly. It searches the OpenAPI spec to find the right endpoint, then executes a small snippet to fetch live data. Because the full API spec lives on the MCP server, nothing about Radar’s API is hard-coded into the prompt, so new Radar datasets become usable without any code changes.
Keeping charts faithful to the data
One key engineering challenge was bridging Markdown output and interactive visualizations. Language models naturally respond in text, tables, and bullet points, but Radar wanted to avoid walls of numbers. The risk was also that a model writing data directly into an answer would round, summarize, or truncate it—precisely what a data tool shouldn’t do.
The solution completely removes data from the model’s prose. When Researcher fetches data, its code returns a small envelope pairing the API path with the result. The model then emits a lightweight chart specification that references that path by its dataFrom identifier, and Radar’s frontend matches that reference to the fetched result, rendering it with the same visualization components used everywhere else on the site. Charts are always faithful to the API, and the approach supports Radar’s full visual vocabulary, including time series, stacked areas, donuts, bar charts, maps, and histograms.
Small optimizations, real impact
Not every task requires the large model. Small, fast models handle side responsibilities, including writing short titles for new conversations and generating follow-up question suggestions. These run separately so they don’t delay the main answer.
Researcher also receives light context about the present moment—the current date and time plus high-level connection details from Radar’s IP page, such as device location and network. This enables requests like “how’s traffic here?” to produce locally relevant answers. The frontend is itself a Worker communicating over a service binding, with per-IP rate limiting and shareable conversations stored in R2. Every layer, from compute through inference, gateway, storage, and data, runs on Cloudflare’s platform.
Making Radar navigable for AI agents
Radar Researcher is one half of Cloudflare's agent story. The other half is about making Radar itself usable by third-party, general-purpose AI agents running in a user's browser. Historically, an agent interacting with a website had to scrape the DOM and guess how the page worked — a fragile, slow, and error-prone process. WebMCP, an emerging web standard, offers a cleaner path: a page can register a small set of well-defined tools that any browser agent can discover and invoke directly.
Cloudflare has added WebMCP support to Radar, implementing both flavors the standard defines:
- The imperative API registers JavaScript tools that call directly into the code powering the UI. An agent can filter by country, region, continent, or ASN; adjust the date range; search pages, sections, and entities (domains, ASNs, IPs, and more); and jump to any section.
- The declarative API annotates Radar's existing HTML forms with a couple of attributes, turning them into tools. An agent can trigger the URL Scanner, pull up a domain's report, or test a site's post-quantum TLS key exchange support.
These tools only drive functionality already available in the interface. The implementation is pure progressive enhancement: in browsers without WebMCP, the tools simply don't exist and nothing changes. This is also a case of practicing what Cloudflare preaches. Radar's own URL Scanner evaluates how agent-ready a site is, and one of its checks looks for exactly this kind of WebMCP integration. By implementing it, Radar now passes its own agent-readiness check, becoming a site that both humans and agents can navigate with ease.
Beta availability and feedback
Radar Researcher is available in beta today on Cloudflare Radar, via the Researcher button in the header on any page. It runs entirely on Cloudflare's developer platform. The roadmap includes expanding the datasets it understands, sharpening its analysis, and adding new ways to interact with data.
Cloudflare is soliciting feedback on the tool. Users can reach the team on social media at @CloudflareRadar (X), noc.social/@cloudflareradar (Mastodon), and radar.cloudflare.com (Bluesky), or by email at [email protected]. The team wants to hear what works, what doesn't, and what still feels out of reach as they prepare for full launch.
