Cloudflare Radar Researcher turns Internet data into a conversation

Cloudflare has opened the beta of Radar Researcher, an assistant that can query public Cloudflare Radar data from a question written in plain language. It replies with an explanation, but more importantly with the interactive charts and underlying data supporting that answer. The goal is not to ask a model to guess what is happening on the Internet: it must locate the relevant datasets, run queries and expose how it reached the result.

The tool is aimed at journalists documenting an outage, network operators, researchers and developers comparing connectivity across countries. It is available from the header of every Radar page and remains a beta, so its findings should be checked before they inform a decision or publication.

The short answer

A simpler entrance to complex datasets

Cloudflare Radar has published observations from the company's global network since 2020. They include HTTP traffic, queries handled by the public 1.1.1.1 DNS resolver and quality measurements collected through Cloudflare's speed test. Visualizations are available on the Web, while their data can also be queried through a free API.

That breadth creates an access problem. Answering one precise question can require learning the catalog's vocabulary, locating the right page, setting time and geographic filters or understanding API parameters. Radar Researcher adds a conversational layer above those resources without hiding the underlying sources.

A question such as “how does Internet quality in Portugal compare with Spain?” can trigger several lookups, produce charts and suggest follow-up questions. For a network disruption, the assistant can combine an incident timeline with traffic changes instead of forcing the user to assemble multiple views manually.

The crucial design choice: the model does not draw the numbers

A text-only generated response would be a poor fit for this analysis. Models can round a figure, truncate a series or summarize a table too aggressively. Cloudflare says it separated commentary from data: agent code retrieves API results, while the model emits a lightweight specification telling the site which native chart to render.

The browser then renders that chart from the raw result it already received. The prose can remain concise without becoming the only carrier of information. This design does not guarantee that every analysis is relevant, but it reduces the risk that a number is fabricated while the answer is being written.

Users can also expand a trace showing how the question was understood, which sources were queried and which tool calls ran. That is more actionable than a generic instruction to check AI output because it identifies exactly what needs review.

An agent architecture built entirely on Cloudflare

Radar Researcher runs in a Worker with the Cloudflare Agents SDK. Every conversation has a Durable Object and its own SQLite database, allowing history to persist and generation to continue server-side even when a user briefly leaves the page. Conversations can be pinned, searched and shared.

Cloudflare uses Workers AI for inference with an ordered fallback across several open-model families. If one model is unavailable, a request can cascade to another. AI Gateway handles the logging, cost tracking, caching and safety controls described by the company.

Data access relies on a unified MCP server and a Code Mode approach. Instead of exposing hundreds of tools, one for every endpoint, the agent receives a small capability set for searching documentation, executing code and inspecting results. It searches Radar's OpenAPI specification, selects an endpoint and creates a focused query. New datasets can therefore become usable without being added individually to the main prompt.

Why data teams should pay attention

The product demonstrates an architecture for question answering grounded in a structured API. In an enterprise, this is generally easier to control than an assistant fed by an export of mixed documents. Units, time ranges, filters and dimensions remain defined by the data service, while the model acts as an orchestrator and interpreter.

The principle can transfer to an internal catalog of observability metrics, sales data, inventory or security indicators. A team adopting it must add controls that public datasets do not need, particularly user-level authorization, tenant isolation, query logging and limits on how much information can be exported.

It is also important to separate a correct figure from a valid conclusion. A value may faithfully come from the API while being compared over the wrong period or presented without knowledge of a methodology change. The trace improves auditability; it does not replace domain expertise or careful reading of definitions.

Radar is becoming operable by external agents too

Cloudflare also added WebMCP support to Radar. This draft specification lets a website declare structured tools that a browser agent can discover, rather than forcing it to infer the interface from the DOM and simulate brittle clicks.

Radar exposes functions for changing a country or date range, searching for a domain and launching selected analyses. The integration is progressive: browsers that do not understand WebMCP continue to display the regular site. The specification is still a Community Group report, however, not a final W3C standard. A critical workflow should not depend exclusively on it yet.

The combination is significant. Radar Researcher is the built-in agent used by people; WebMCP prepares the same service for external agents. Both can rely on the same functions and data while presenting different interfaces.

How to test it without overreading the output

Start with a question whose answer can be checked easily in Radar's regular pages. State the country, period and metric explicitly. Then open the trace, inspect the endpoints and compare the generated chart with the corresponding native visualization.

Next, ask a comparative question and change only one parameter. If the conclusion shifts sharply, inspect scales, aggregation and data availability. For a published analysis, retain the shared link, record the access date and cite the dataset rather than treating the assistant as the sole source.

Radar Researcher makes exploration considerably more approachable, and its architecture takes numerical traceability seriously. Its real value will depend on whether it can recognize ambiguous questions, disclose dataset limitations and avoid turning visible correlation into a confident explanation. The beta is the right time to test exactly those boundaries.