Radar Researcher: Turn Questions About the Internet into Live Data Charts

Cloudflare’s Natural Language Gateway to Global Internet Data Goes Live

Cloudflare has thrown open a new door to its Radar internet intelligence platform — one that anyone can walk through using plain English, not code. The company’s engineering team published a detailed post on the Cloudflare Blog describing the beta launch of Radar Researcher, an AI-powered interface that turns descriptive questions into live, interactive visualizations of global network data.

The tool, now accessible beta on every Cloudflare Radar page, eliminates the need to memorize API endpoints, filter parameters, or data schemas. Instead, a journalist covering an internet outage in Iran or a policy analyst benchmarking broadband quality in Portugal simply types a question and gets back real-time charts backed by the same APIs Radar’s own dashboards use.

Radar Researcher is not a skin-deep chat overlay. It maintains stateful, resumable conversations. It offers both concise answers and multi‑topic reports. Every response is accompanied by a full audit trail showing how the underlying language model interpreted the query, which datasets it queried, and exactly what tools it called to produce the result.

Under the Hood: How an AI Agent Delivers Live Charts Without Hallucinating Data

The architectural choices that make Radar Researcher trustworthy for data work are as notable as the user interface itself. Rather than placing raw numbers inside the model’s text — where rounding, truncation, or confabulation can creep in — Cloudflare’s team engineered a chart specification language that keeps data entirely outside the language model’s prose.

When the agent fetches an API result, its code returns a small envelope pairing the response payload with the endpoint path. The model then emits a lightweight chart directive that references that same path, and Radar’s frontend renders it using the exact same visualization components deployed across the rest of the site. A single time‑series, donut, map, or histogram is always faithful to the underlying API response.

The agent’s ‘brain’ is a Workers AI instance running open models such as Kimi K2.7, but it does not depend on a single inference provider. An ordered fallback chain cascades across three different model families transparently; if one model is momentarily at capacity, the request silently shifts to the next, insulating the tool from upstream incidents. Every call routes through AI Gateway for logging, cost tracking, caching, and safety guardrails.

Perhaps most consequential is how Researcher reaches Radar’s data. Instead of hand‑coding tools for hundreds of API endpoints, the team connected the agent to Cloudflare’s unified MCP server using a Code Mode. The model receives only three generic tools — search, execute, and docs — then writes small code snippets that query the Radar OpenAPI spec directly. New endpoints become available to Researcher with zero prompt changes or code pushes.

Behind the scenes, each conversation is a stateful Durable Object with its own SQLite database, so chat history, titles, and streaming responses persist even if the user leaves the page mid‑answer. Generation continues server‑side and resumes on reconnect. A pair of small, fast models handle side tasks like titling conversations and suggesting follow‑up questions without adding latency to the primary response.

The WebMCP Play: Why Cloudflare Is Teaching Agents to Navigate the Internet Like Humans

According to the Cloudflare Blog, Radar Researcher’s launch is also a deliberate demonstration of WebMCP, an emerging standard that Cloudflare itself is championing. The traditional model — scraping a page and guessing at its structure — is brittle and error‑prone. WebMCP replaces that with a declarative contract: websites register a set of well‑defined, tool‑callable actions that any browser‑based agent can discover and invoke directly.

Cloudflare’s implementation on Radar uses both flavors the standard offers. The imperative API exposes JavaScript tools that call straight into the user interface’s own logic, letting an agent filter by country, ASN, or date range or search across domains and sections. The declarative API annotates existing HTML forms so an agent can scan a URL, look up a domain’s report, or test post‑quantum TLS key exchange support without any additional scripting.

Every one of these capabilities is a pure progressive enhancement — in a browser that lacks WebMCP support, the tools simply aren’t there and nothing breaks. But for the growing cohort of AI‑capable browser agents, Radar becomes a first‑class data source that can be navigated with the same programmatic certainty a developer expects from an API. It is a case of practicing what Cloudflare preaches: Radar’s own URL Scanner now evaluates whether a site is ‘agent‑ready’ by checking for exactly this kind of WebMCP integration, and by implementing it themselves, the Radar team has made their platform pass its own readiness test.

That alignment between internal tooling and external standard signals a strategic bet. Cloudflare is not simply adding a chat layer to an existing dashboard; it is positioning its entire developer platform — Workers, Durable Objects, AI Gateway, and now MCP — as the substrate on which an agent‑traversable web will run.

Internet Intelligence Becomes Conversational Infrastructure

Radar Researcher transforms Cloudflare’s trove of DNS queries, HTTP traffic patterns, and network quality measurements from a specialized data library into a conversational utility that researchers, journalists, and operators can interrogate in the seconds they need it. The move from manual navigation to plain‑language exploration strips away the final friction between a question about the global internet and a reliable, viz‑rich answer.

For teams building AI‑driven data exploration pipelines, programmatic SEO and AI automation is how Andres SEO Expert approaches it — get in touch.