OpenCode, Windsurf, and Cloudflare Skills: AI-Powered Development on the Edge

The Gap

There’s a gap in enterprise sales that nobody talks about. We sell products we’ve never used. We pitch platforms we’ve never built on. We talk about developer experience based on slides, not experience.

I decided to close that gap. Not by learning to code — but by using the same AI tools our customers’ developers are adopting right now to actually build on Cloudflare.

This post is about what happened when I pointed OpenCode and Windsurf Cascade at my own site and let AI do the heavy lifting.

The Tools

Before I get into the builds, let me explain what I’m working with. None of these existed a year ago. All of them are changing how software gets built.

OpenCode is an open-source AI coding agent. You open it, point it at a project, and describe what you want in plain English. It writes the code, configures the infrastructure, and deploys it. Cloudflare uses OpenCode internally — one engineer used it to rebuild 94% of the Next.js API in a single week for $1,100 in AI tokens. That project, called Vinext, has 1,700 unit tests and builds 4.4x faster than the original.

Cloudflare Skills are pre-built knowledge packs that teach any AI coding agent how to build on Cloudflare. Workers, Pages, D1, R2, Durable Objects, AI Gateway, Agents SDK — 60+ products covered. When you install the skills, the AI stops guessing and starts building with production best practices. Auto-loaded based on what you’re working on.

Windsurf Cascade is an AI-powered IDE with Claude built in. Think VS Code, but the AI assistant on the right side can read your entire project, write code, run commands, and manage your infrastructure. It’s the IDE — Cloudflare approved it for enterprise use internally.

Cloudflare MCP Servers are the piece that ties it all together. MCP (Model Context Protocol) lets AI tools talk directly to external services. Cloudflare built five MCP servers that connect your AI IDE to your actual Cloudflare account: documentation search, build status, API management, storage bindings, and observability. Once connected, you manage Cloudflare from a chat window.

Part 1: OpenCode — Deploy a Worker in 60 Seconds

I opened OpenCode, pointed it at my saltwaterbrc-astro repository, and typed one sentence:

“Create a Cloudflare Worker at /api/demo that returns company data from a KV namespace. Use Cloudflare Skills best practices.”

The Cloudflare Skills kicked in immediately. Instead of generating generic code, OpenCode built the Worker with every production best practice baked in:

• wrangler.jsonc with $schema validation (not the old TOML format — the skills know the current recommendation)
• Typed Env interface with satisfies ExportedHandler<Env> for type-safe handler exports
• Coalesced KV keys — a single company:profile JSON key instead of many small keys, which is how Cloudflare recommends structuring KV data
• cacheTtl: 300 on KV reads for edge caching
• CORS headers with preflight handling
• Resilient fallback — if KV returns null, the Worker serves hardcoded defaults so the endpoint never breaks
• Staging environment with a separate KV namespace
• Observability enabled with 100% sampling

Five files. One prompt. Every line following Cloudflare’s own best practices because the skills told the AI exactly how to do it.

I deployed it. It’s live right now: saltwaterbrc.com/api/demo

Hit that URL and you’ll get a JSON response served from Cloudflare KV at the nearest edge location to wherever you are. That’s a production Worker, built by AI, deployed to 300+ data centers worldwide.

The whole thing took minutes. Not hours. Not days. Minutes.

What the Skills Actually Did

This is the part that matters for customer conversations. The Cloudflare Skills aren’t just autocomplete. They’re opinionated guidance that steers AI tools toward production-quality output.

Without the skills, the AI would have generated something that works but cuts corners — maybe a wrangler.toml instead of .jsonc, untyped environment variables, no error handling, no caching. It would deploy, but it wouldn’t be production-ready.

With the skills, the AI builds like a senior Cloudflare engineer. It knows the current API surface. It knows the recommended patterns. It knows the anti-patterns to avoid. This is the same knowledge that Cloudflare’s own engineering teams use.

That’s the pitch for every developer-heavy account: “Your developers’ AI tools already know how to build on Cloudflare. Zero ramp-up time.”

Part 2: Cascade + MCP — Managing Cloudflare from Chat

The second half of this story is about what happens after you build. Because building is only part of the job — you also need to deploy, monitor, debug, and manage.

I connected Windsurf Cascade to my Cloudflare account using five MCP servers:

• cloudflare-docs — searches the full Cloudflare documentation from the chat panel
• cloudflare-builds — shows my Workers, deployments, and build status
• cloudflare-api — manages DNS, WAF rules, page rules, account settings
• cloudflare-bindings — configures KV, D1, R2, and Durable Objects bindings
• cloudflare-observability — pulls logs, analytics, error rates, and performance data

Once connected, Cascade found my three existing Workers immediately: the sandbox, the AI-powered Q&A endpoint, and the visitor counter. It identified my account, my zone, my bindings.

Now I can type “show me the error rate on my ask Worker for the last 24 hours” and get the answer without opening a browser. I can type “add a DNS record for staging.saltwaterbrc.com” and it happens. I can type “search the docs for how Spectrum handles custom hostnames” and get the exact documentation I need — which, by the way, came in handy during a customer call earlier today.

This isn’t a demo environment. This is my real Cloudflare account, managed through AI.

Part 3: D1 — A Serverless Database Built from Chat

The guestbook at saltwaterbrc.com/guestbook runs on Cloudflare D1 — a serverless SQL database built on SQLite that runs at the edge. I built the entire thing through Windsurf Cascade. One conversation. No database provisioning, no connection strings, no servers.

D1 is the product most people overlook, and it shouldn’t be. Here’s why.

Traditional databases force you to pick a region. Your users in Tokyo hit a database sitting in Virginia. D1 distributes read replicas globally — every user gets local-speed reads from the nearest Cloudflare data center. No cold starts on reads. No connection pooling. No pgBouncer. Workers talk to D1 natively.

The business case is straightforward. Think about a global retailer with product catalogs, inventory lookups, and user sessions across every continent. With a traditional database, they’re either running expensive multi-region replicas or accepting latency for users far from the primary. With D1, every product page load is a local read. A SaaS platform can give each tenant their own D1 database — isolated, scalable, zero ops. A financial services company gets low-latency transaction logs and compliance records at the edge.

The guestbook is a simplified version of this pattern, but the architecture is the same: user request hits the nearest edge, Worker processes it, D1 serves the data. The whole thing cost me nothing to build and pennies to run.

Why This Matters for Sales

I’m not a developer. I’m an enterprise account executive. I sell Cloudflare to Global 2000 companies in the Southeast.

But after this exercise, I can walk into any customer meeting and say:

1. “I’ve built on this platform. Here’s my domain running 16 Cloudflare products.”
2. “I deployed a Worker in under a minute using the same AI tools your developers already use.”
3. “I manage my Cloudflare infrastructure through a chat interface — no dashboard required.”

That’s not a demo someone else built. That’s not a slide. That’s lived experience.

And here’s the thing — if a non-developer sales rep can build 16 products in two weeks and deploy a production Worker in minutes, imagine what your engineering team can do with Cloudflare and AI coding tools.

That’s the conversation I want to have with every CTO, VP of Engineering, and platform team lead in my territory.

The Proof Point: Vinext

If my story sounds like a small-scale experiment, here’s the enterprise version.

In February 2026, one Cloudflare engineer used OpenCode and Claude to rebuild 94% of the Next.js API from scratch in one week. The project is called Vinext. It’s open source. It builds 4.4x faster than Next.js and produces bundles 57% smaller. Every line was AI-generated. Every line passes 1,700 unit tests and 380 end-to-end tests.

The total cost: $1,100 in Claude API tokens.

The Cloudflare Skills were the enabler. OpenCode was the tool. Claude was the brain. The result was production-quality software that’s already running government websites.

What’s Next

I’m continuing to build. The next projects on my list:

• A recorded walkthrough of managing Cloudflare entirely through Cascade + MCP — the “manage your infrastructure from chat” demo for customer presentations
• A real-time analytics dashboard on saltwaterbrc.com powered by Workers Analytics Engine — live traffic, request volume, geographic distribution, and cache performance, all queryable from the site itself
• Email Security integration — routing Google Workspace through Cloudflare Email Security to complete the Zero Trust stack, then documenting the setup as a training guide for any customer running M365 or Google Workspace

The site is saltwaterbrc.com. The code is on GitHub. The API is live. And every phase adds another Cloudflare product I can demo and sell from experience.

Try It

Hit these endpoints right now:

• saltwaterbrc.com/api/demo — The KV-backed Worker I deployed with OpenCode
• saltwaterbrc.com/api/status — Site metadata and your nearest Cloudflare data center
• saltwaterbrc.com/api/edge-info — Full request inspector showing your location, TLS, network info
• saltwaterbrc.com/stream — AI-generated video hosted on Cloudflare Stream
• saltwaterbrc.com/ask — Ask This Blog, powered by Workers AI + Vectorize + AI Gateway
• saltwaterbrc.com/playground — Interactive code sandbox running on Cloudflare
• saltwaterbrc.com/guestbook — D1-powered guestbook, deployed with Windsurf Cascade

Every one of these was built by AI, running on Cloudflare, at the edge. That’s the story.