Cloudflare Radar Researcher: A Plain-Language Agent Over 500 Live API Endpoints
TL;DR
Cloudflare Radar Researcher is a plain-language agent over live internet data, built on MCP code mode and chart specs that keep numbers out of the model.
Last updated: August 7, 2026
Cloudflare's Agents Week closed with a product that quietly demonstrates the week's biggest architectural claims. Radar Researcher, in beta as of August 7 on radar.cloudflare.com, lets you ask questions about global internet traffic in plain language - "what happened to Iran's traffic during the January shutdown?" - and get a written answer plus the same interactive charts Radar renders for hand-built queries. It is not a demo bolted to one endpoint: it is an agent talking to hundreds of live Radar API endpoints it was never individually programmed to understand. That detail is the story.
What shipped#
Radar Researcher is a panel on every Radar page. You ask a question, it investigates, and you get charts with a plain-language explanation, suggested follow-ups, a searchable conversation history, and shareable links that expire after 30 days. Any existing chart has an "Explain with AI" action that hands the assistant the exact visualization plus the raw data behind it.
The underlying data is not generated: every number comes from Radar's public API, the same one any developer can call for free. The tool's job is removing the need to know the API's vocabulary and structure. A journalist covering an outage can ask about it directly instead of reading the docs, picking filters, and assembling charts by hand.
The three moves worth copying#
One tool surface instead of hundreds. Radar's API has hundreds of endpoints, and Cloudflare did not hand-write a function for each. The agent connects to Radar's data through the unified Cloudflare MCP server in "code mode", which exposes three tools: search, execute, and docs. The model searches the OpenAPI spec for the right endpoint, executes a small snippet that fetches live data, and reads docs when stuck. Because the spec lives on the MCP server rather than in the prompt, new Radar datasets become queryable with no code changes. This is the same progressive-disclosure pattern behind skills-over-MCP design discussions: give the agent a small, well-defined surface and let it discover the rest.
Keep numbers out of the model's prose. When an LLM summarizes fetched data, it rounds, truncates, and drifts. Radar Researcher's answer sidesteps this: the executed code returns an envelope pairing the API path with the result, and the model emits a lightweight chart spec that references the path instead of pasting numbers:
{ "type": "speedFlower", "title": "Internet speed quality - Portugal", "dataFrom": "/radar/quality/speed/summary?location=PT" }
The frontend matches dataFrom to the already-fetched result and renders it with the site's real chart components. The model writes Markdown; charts stay faithful to the API by construction. For anyone building an agent that must report data accurately, this "fetch once, reference by path" envelope is the pattern to steal.
Small models for side tasks. The heavy reasoning model is not the only brain. One small model titles each new conversation, another suggests follow-ups, and both run off to the side so they never delay the main answer. The main model runs on Workers AI with an ordered fallback chain across three open-model families - including Kimi K2.7, which Cloudflare has been serving at scale - so a capacity spike at any one provider cascades transparently rather than failing the request. Every call routes through AI Gateway for logging, cost tracking, caching, and guardrails, the same control plane Cloudflare just unified across Workers AI and AI Gateway.
The stateful agent infrastructure#
Radar Researcher is built entirely on Cloudflare's developer platform. Each conversation is a stateful Durable Object with its own SQLite database, so chat history and streamed responses survive page reloads, and generation continues server-side even after you navigate away mid-answer. The frontend is a Worker talking to the agent over a service binding, with per-IP rate limiting and conversations stored in R2. Every answer includes an expandable trace of the model's tool calls, so you can audit exactly which endpoints were hit and how the answer was assembled - a natural fit for the agent development lifecycle Cloudflare has been formalizing this week.
Why it matters#
Two things are worth taking from this launch. First, as a user, it is the most convincing argument yet for the agent-over-live-API pattern: a genuinely useful, honest data tool where the model cannot hallucinate the numbers because it never produces them. Second, as a developer, it is a reference architecture. The three decisions - a small discoverable tool surface over a large API, data envelopes that keep raw numbers out of model output, and cheap side models for peripheral tasks - each solve a real failure mode in agent products, and each is reproducible on any stack, not just Cloudflare's.
There are limits. It is a beta over Radar's own catalog, not a general data assistant, and the analysis depth depends on which models the fallback chain lands on. But as a demonstration of what an agent platform can do when the data layer is designed for it, it lands with the week's bigger claims: the Agentic Internet argues agents will visit every site; Radar Researcher shows what happens when a site is ready for them. The next question for Cloudflare is whether this "ask your data anything" pattern stays inside Radar or becomes a template every Workers developer can ship.
FAQ#
What is Cloudflare Radar Researcher?#
Radar Researcher is a natural-language agent, in beta on radar.cloudflare.com, that answers questions about global internet traffic with a written explanation plus the same interactive charts Radar renders for hand-built queries. Every number comes from Radar's public API; the tool removes the need to know the API's vocabulary and structure.
How does Radar Researcher keep the model from getting the numbers wrong?#
The executed code returns an envelope pairing the API path with the result, and the model emits a lightweight chart spec whose dataFrom field references that path instead of pasting numbers. The frontend matches dataFrom to the already-fetched result and renders it with Radar's real chart components, so charts stay faithful to the API by construction.
What is MCP code mode and why does Radar Researcher use it?#
Code mode on the unified Cloudflare MCP server exposes three tools: search, execute, and docs. The model searches the OpenAPI spec for the right endpoint, executes a small snippet that fetches live data, and reads docs when stuck. Because the spec lives on the MCP server rather than in the prompt, new Radar datasets become queryable with no code changes.
What Cloudflare infrastructure is Radar Researcher built on?#
Each conversation is a stateful Durable Object with its own SQLite database, so history and streamed responses survive page reloads. The frontend is a Worker talking to the agent over a service binding, conversations are stored in R2, the main model runs on Workers AI with a three-family fallback chain, and every call routes through AI Gateway.
Continue Reading#
- Cloudflare Unifies Workers AI and AI Gateway Into One Control Plane
- Cloudflare's Agentic Internet: Readable, Discoverable, Callable, and Payable
- Cloudflare Serves Kimi and GLM at Scale: Smaller, Faster, Safer Models
- Cloudflare's Agent Development Lifecycle: Build, Test, Audit
- Skills Over MCP: Progressive Disclosure for Agent Tool Surfaces
Sources#
Get the next deep dive like this in your inbox
One email a week on News and the rest of the AI dev stack. Free.
Read next on AI coding tools
Cloudflare Folds Workers AI Into AI Gateway: One Control Plane for Every Model Provider
Cloudflare is merging Workers AI and AI Gateway into one control plane with a unified /ai/ API, default gateways and shared credits. What changes for you.
6 min readCloudflare's Agentic Internet: Readable, Discoverable, Callable, and Payable
Cloudflare's Agentic Internet treats agents as a new kind of web visitor with four primitives: readable, discoverable, callable and payable.
7 min readCloudflare Runs Kimi and GLM at Scale: FP8 KV Caches, INT4 Weights, and a Cache Safety Net
Cloudflare runs Kimi K2.6 and GLM 5.2 on Workers AI with FP8 KV caches, INT4 weights and a KV cache integrity check. What each technique buys on GPU fleets.
7 min readNew here? Start with
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.
