Cloudflare Agents SDK Builder
Guides developers in building stateful, real-time, and durable AI agents on Cloudflare Workers.
The material includes OAuth, HMAC secure replies, secret handling, MCP hardening, and human-approval examples, showing partial security awareness. However, for email, browser, webhook, push, MCP, and code-execution capabilities it does not systematically define least privilege, full data flows, sensitive-data handling, rollback, or confirmation boundaries, so points are deducted.
The documents are broadly self-consistent and cover installation, configuration, core APIs, recovery, retries, and some failure messages. However, several references defer to external documentation, with no local test suite, version pinning, key-path reproduction evidence, or unified error-handling guidance. Static calibration caps this at 10, so points are deducted.
The description clearly targets stateful agents, workflows, real-time apps, MCP, chat, voice, and browser automation on Cloudflare Workers, with recognizable use cases. Non-fit boundaries, input/output contracts, and semantic trigger conditions remain incomplete, and core retrieval depends on Cloudflare-hosted documentation without evidence of mainland-China reachability or Chinese-language support, so points are deducted.
The skill has progressive sections, installation commands, configuration examples, reference files, experimental labels, an Apache-2.0 license, and migration cautions. It lacks explicit SDK/documentation versions, a changelog, named maintenance responsibility, an update path, troubleshooting coverage, and a systematic limitations section, so points are deducted.
It provides directly adaptable TypeScript, React, Wrangler, and API examples across many Agents SDK features, giving it clear core utility. But the examples were not executed, some content depends on external documentation, and complex integrations still require user-side assembly, authentication, and testing. Static calibration permits at most 7, so the score is capped.
The topic table links claims to Cloudflare primary documentation and the skill includes auditable code snippets, providing limited traceability. It lacks a committed test suite, CI coverage of key paths, or third-party execution evidence; this review is static only, so the score is capped at 5.
- The skill covers email, browser automation, MCP, webhooks, push notifications, and code execution; review permissions, authentication, data egress, approval flows, and rollback plans before deployment.
- Core guidance relies on retrieving Cloudflare documentation; confirm reachability in the target environment and prepare mirrors, caches, or local copies for mainland-China network constraints.
- Examples were not executed in this review, so API compatibility, migration correctness, and production stability remain unverified.
- No explicit version policy, changelog, or maintenance owner is provided; recheck all external references before upgrading the SDK.
What does this skill do, and when should you use it?
This skill supports developers building AI agents with the Cloudflare Agents SDK. It covers the Agent class, persistent state, WebSockets, callable RPC, scheduling, Workflows, queues, retries, MCP, chat applications, and React client hooks. It prioritizes retrieval from Cloudflare’s Agents documentation because SDK knowledge may become outdated. Voice, browser tools, and Think are identified as experimental features.
It guides users through checking and installing the agents npm package, configuring Wrangler Durable Object bindings, SQLite migrations, and Workers AI bindings, and implementing Agent classes, state validation, callable methods, and routing. It also provides API examples for scheduling, workflows, durable execution, queues, retries, streaming chat, MCP, email, webhooks, push notifications, observability, and React clients, with links to Cloudflare documentation and local reference files.
- A developer is building a stateful chat agent or collaborative application on Cloudflare Workers.
- A team needs WebSocket communication, callable RPC, and React hooks for a real-time client experience.
- An engineer needs durable multi-step background processing, queued work, or retryable workflows.
- A developer needs an agent to connect to MCP services or expose an MCP server on Cloudflare.
- A team needs scheduled tasks, email handling, webhooks, push notifications, or diagnostic events for an agent.
What are this skill's strengths and limitations?
- Covers a broad implementation path from Agent classes and state to workflows, queues, retries, streaming chat, and MCP.
- Includes directly usable examples for Wrangler configuration, routing, RPC, and React hooks.
- Explicitly prioritizes current official documentation for a fast-changing SDK.
- It is primarily suited to the Cloudflare Workers and Agents SDK ecosystem rather than general agent development.
- Voice, browser tools, and Think are still marked experimental.
- The source provides no independent test suite, performance benchmarks, or cost analysis.
How do you install this skill?
This skill is part of the cloudflare/skills collection. Install the collection with npx skills add https://github.com/cloudflare/skills, or copy the relevant skill folder into the directory used by your agent client, such as ~/.codex/skills/ for OpenAI Codex. The repository does not document a separate installation command for only agents-sdk.
How do you use this skill?
In an Agent Skills-compatible client, make a request such as “Build a stateful AI agent with WebSockets on Cloudflare Workers.” Before implementation, follow the skill’s recommendation to consult https://developers.cloudflare.com/agents/. In the project, run npm ls agents; if the package is missing, run npm install agents.