Introducing Agent Lee - a new interface to the Cloudflare stack
Key Points
- In-dashboard AI assistant for Cloudflare
- Executes ~250k tool calls/day across services
- Write actions require explicit user approval
Summary
Agent Lee is an in-dashboard AI assistant that understands your Cloudflare account and can inspect, visualize, debug, and apply changes across Cloudflare products (Workers, DNS, R2, SSL/TLS, Cache, Tunnel, etc.). It runs in beta, serves ~18,000 daily users, and executes ~250k tool calls/day. Agent Lee generates TypeScript code to call platform APIs, executes that code via a sandboxed Durable Object proxy, and enforces an explicit approval (elicitation) step for any write operations.
Key Points
- Capabilities: natural-language queries, inline visualizations (charts, tables, architecture maps), debugging, resource creation, and making/config changes when you approve.
- Architecture: uses Codemode to turn tool definitions into a TypeScript API; model generates code that runs on a Cloudflare MCP server via a Durable Object which classifies calls as read or write.
- Security & permissions: API keys are never embedded in generated code; DO injects credentials server-side and blocks writes until the user explicitly approves the elicitation prompt.
- Observability & quality: telemetry includes evals, thumbs up/down feedback, tool execution success, hallucination scoring, and per-product performance metrics.
- Platform primitives: built on public building blocks (Agents SDK, Workers AI, Durable Objects, MCP) so teams can reproduce similar functionality.
- Availability & status: in active beta (Free plan users can enable via Ask AI in the dashboard); expect ongoing improvements and potential edge cases—feedback is requested.
Practical notes for engineers
- Integration model: treat Agent Lee as a scripted tool runner that produces TypeScript calls; investigate generated code patterns to understand multi-step workflows.
- Operational safety: any write/change requires an explicit UI confirmation—Agent Lee cannot bypass the approval gate.
- Extensibility: the same MCP and Workers primitives are available to developers to build custom agents or automate tasks outside the dashboard.
Try it
Log in to the Cloudflare dashboard and click “Ask AI” in the upper-right to try Agent Lee (beta). Provide feedback to help improve behavior and coverage.