The Cloudflare API has an incredible MCP. It uses Code Mode so you get access to the entire Cloudflare API (~2500 endpoints) in ~1k tokens. In this video we use the MCP to buy a domain, deploy a React app to it, have my agent check errors, fix them... and re-deploy. It's SICK🔥show more

I've LOVED using it to buy domain names!!

So good!! Literally coming up with names, checking availability and price and just buy it. 2026 baby

@mattzcarey Code Mode @mattzcarey

这下代理不只是会写代码，还能直接把域名买了部署完

Video is cleeeeeaaaaaaaaaaaannnnnnnnnnnnnn!

❤️‍🔥

last time I used these cf mcps they had to be constantly reauthorized, one by one, it was pretty annoying

Which coding agent were you using? I’ve only authenticated once /w pi

I think codex? I havent tried on claude yet

MCP is making AI agents much more capable by giving them direct access to real tools and APIs. The ability to build, deploy, debug, and redeploy through one workflow is where agents get really interesting. @Jarvixdotlive

this is my priority today to-do, ty! 🤗🤍❤️🤍🥰 ive been meaning to do it for too long 🫣😶‍🌫️

2500 endpoints in ~1k tokens is the part worth stealing. most MCPs dump the whole schema into context, then act surprised when the agent forgets the task. buying domains from a chat window is either the future or how I end up owning nine of them.

amazing.

Let’s discuss a possible marketing partnership

Is this wrangler?

Wrangler deploys, Cloudflare MCP does the domain buying. Agent uses both in the video!

This has a bright future. DM me and let’s discuss a collaboration 👀

Your project caught my attention 🚀 Let’s discuss how we can work together, send me a DM

Code Mode solves schema cost beautifully. The production question is authority: buying a domain and deploying in one trace combines billing, DNS and code execution. Did you scope credentials per resource, or was this using an account-wide token? Broad discovery works best with narrow, reversible permissions.

Looking great so far. Free assistance anytime, drop me a DM.

The development looks promising, feel free to reach out 🔥

"agent checks errors, fixes them, re-deploys" is the loop I live in. the one habit that keeps it honest: I open the deployed thing myself instead of trusting the deploy output. done is a hypothesis until it's clicked.

Great video. Do you like it better than vercel?

I Really like what you’re creating. Want some help? No charge, just reach out.

2500 endpoints stuffed as tool schemas eats the window before work starts. Exposing them as code the agent writes against is the only shape that survives long runs.

Cloudflare 这套 MCP 确实省事，一个 token 打通 2500 个端点，代理自查自修的体验很惊艳。只是权限粒度太粗，一个全量 token 泄露基本等于整个账号沦陷，生产环境还是得按域拆分范围并配审计日志，视频里这点没提。

code mode is the fix for tool bloat. 2500 endpoints as 1k tokens instead of 300 tool defs eating your context. we cut ours down to like 6 tools for the same reason

The part I keep staring at is the agent buying a domain. Code Mode compresses the surface, but spend is the one call where a retry isn't free. An agent that reads an ambiguous error and tries again has just bought two.

Can you clarify? A retry for the same domain name would not result in a double charge.

Fair, you're right there. The name itself is the idempotency key, so attempt two comes back "already exists" rather than charged. I meant the spend calls with nothing unique behind them: a renew, a top-up. Nothing in those separates a retry from a genuine second buy.

Let’s talk⚡️dm us

Code Mode's win is the model writing TS against bindings that run in an isolate, so those huge API responses never touch the context.

When will the grok bot plugin use this?

Hey 👋 can you send me Dm.

Never used it, is it a remote server?

Yes and it’s fucking good

How do u avoid llm knowing ur access tokens.. it can read env

That is a very compelling token trade. Generated API calls still need an authorization surface a human can inspect.

Yeah most coding agents will have better rules. I am using pi with kind of a "yolo" mode in the video, so it doesn't ask me for permission unless I tell it to.

Code Mode trades an enumerated surface for a searchable one. Works when the agent already knows it wants Cloudflare. The tool list was a terrible context budget, but it was the only enumeration the model got. Once you import what you need, discovery is a step nobody owns.

2500 endpoints in ~1k tokens is genuinely wild. does it hold up when you start chaining those calls or does context blow up fast?

2500 endpoints in 1k tokens is just moving the context problem into the error messages

One of the best, @Cloudflare is amazing