A New Interface for the Cloudflare Stack, Introducing Agent Lee

Agent Lee is an AI assistant built into the Cloudflare dashboard that understands a user's account context, takes questions in natural language, and actually carries out tasks. It's designed to handle things like diagnosing the cause of Worker errors, checking DNS issues, enabling Access, and creating an R2 bucket and connecting it to a Worker, all from a single screen.

It's currently in beta, but it already has over 18,000 daily users, generating around 250,000 tool calls per day. Cloudflare views this not as a simple FAQ responder, but as an operational tool that works against real accounts and real traffic.

The core implementation is Codemode. Instead of exposing MCP tool definitions directly, it turns tools into a TypeScript API and has the model write code that calls that API. The generated code runs as sandboxed execution on the upstream Cloudflare MCP server, with a Durable Object in front acting as a credentialed proxy.

In this structure, the DO classifies the generated code into read or write operations. read requests are proxied immediately, while write requests cannot execute without passing an elicitation gate that requires user approval. The API key never enters the generated code and is kept only inside the DO, injected server-side, so writes are structurally blocked before approval.

Cloudflare built this feature on the same components available to customers rather than as an internal-only tool. It uses the Agents SDK, Workers AI, Durable Objects, and the same MCP infrastructure, and emphasizes that constraints discovered during operation can be fed back into platform improvements.

Another pillar is generative UI. As a conversation progresses, instead of just returning text answers, it dynamically generates and attaches visual blocks like charts, tables, and architecture maps. For example, a question about traffic trends is rendered immediately as a line graph made from real data.

Safety and quality management are measured separately.

• Evals measure conversation success rate and information accuracy
• Collecting user feedback (thumbs up / thumbs down)
• Tracking tool call execution success rate and a hallucination scorer
• Analyzing conversation performance separately by product

In the long term, the plan is to expand Agent Lee beyond the dashboard into a universal interface for the Cloudflare platform. The goal is to reach CLI, mobile, and eventually a proactive agent that remembers a user's behavioral context and detects and alerts on problems before being asked.