AI Agent - Aug 8, 2026

Cloudflare Radar Researcher: Code Mode and MCP Architecture

Quick answer

Radar Researcher is a Cloudflare Worker using the Agents SDK. Each conversation is a Durable Object with SQLite state. Workers AI supplies an ordered multi-family model fallback chain, AI Gateway handles logging and related controls, and the unified Cloudflare MCP server exposes Radar’s API through Code Mode.

The model receives three high-level tools—search, execute, and docs—instead of hundreds of endpoint tools. It discovers the relevant OpenAPI operation, writes a small query, and returns a data reference. The frontend renders exact API results as Radar charts while the user can inspect the tool trace.

Request path

Radar panel
  -> frontend Worker and service binding
  -> stateful conversation Durable Object
  -> Workers AI through AI Gateway
  -> unified Cloudflare MCP Code Mode
  -> Radar API result
  -> chart specification + explanation + trace

Cloudflare also documents per-IP rate limiting and R2 storage for shareable conversations. These are implementation facts from the launch, not a generic blueprint that automatically fits sensitive or regulated data.

Why Code Mode changes the tool surface

A conventional MCP client may load many tool schemas into model context. Code Mode keeps the visible surface small. Researcher searches API metadata for the right endpoint and executes code against a typed interface. As Radar adds API datasets, discovery can change without hard-coding every endpoint into the prompt.

The benefit is compact, composable access. The review obligation remains: code can choose the wrong endpoint, dates, location, unit, or aggregation. Inspect the generated call and result, not just the prose.

Stateful conversations and model fallback

Each conversation’s Durable Object owns streaming state and SQLite history. Cloudflare says generation continues server-side if the user leaves the page and can resume on reconnection.

Workers AI runs an ordered fallback chain across three model families. That improves availability but may introduce behavioral variation. A research record should include the trace and output evidence rather than assuming the same prompt always follows the same reasoning path.

Exact charts without model-rounded data

Researcher does not rely on a model to reproduce full result arrays in Markdown. Cloudflare describes a result envelope pairing the Radar API path with fetched data. The model emits a lightweight radar-chart specification referencing that path, and the frontend renders the result with Radar visualization components.

This design reduces rounding and truncation in chart data. It does not validate whether the selected API path or parameters answer the user’s intended question.

Audit checklist

  1. Read the interpreted question in the trace.
  2. Verify the selected Radar dataset and endpoint.
  3. Check location, ASN, domain, filters, dates, timezone, and units.
  4. Compare displayed chart values with the fetched API result.
  5. Review transformations, baselines, and missing data.
  6. Separate measured observations from causal explanations.
  7. Save evidence outside expiring share links when retention is required.
  8. Re-run time-sensitive queries and record the retrieval time.

Frequently asked questions

How does Radar Researcher access Radar data?

Cloudflare says it connects to the unified Cloudflare MCP server in Code Mode. The model uses search, execute, and docs tools to discover and query Radar API endpoints.

Where is conversation state stored?

Cloudflare describes each conversation as a stateful Durable Object with its own SQLite database, allowing generation and history to persist across page reconnection.

Does the model write chart numbers into prose?

Cloudflare’s design keeps fetched API data separate and has the model emit a chart specification referencing that result, so the frontend renders the exact data with Radar components.

Is an inspectable trace the same as a correctness proof?

No. A trace improves auditability by showing interpretation and tool calls, but reviewers must still check endpoint choice, parameters, data, calculations, and conclusions.

Continue by task

Official sources

Source check: August 8, 2026. Verify current models, tools, APIs, storage, traces, retention, rate limits, and platform behavior before adopting the pattern.