Giving a Chat App Operational Access to My Cloudflare Account with MCP

#agents #llm #mcp

I recently built a small demo that changed how I think about agent tooling and platform integrations.

In short, I gave a chat-based application operational access to my entire Cloudflare account by connecting Cloudflare’s managed MCP servers to a LangChain agent powered by Claude.

No custom APIs.
No giant tool schemas.
No hand-written adapters.

Just MCP.

The idea

Most agent setups today treat tools as static. You define them upfront, wire them into prompts, and hope the model uses them correctly.

That works for small demos. It starts breaking down once you point an agent at a real platform like Cloudflare, which exposes hundreds of possible operations across many services.

MCP changes that model.

Instead of shipping a massive list of tools, the agent connects to MCP servers that describe available capabilities. The model can then discover tools dynamically at runtime and load only what it actually needs.

What I built

In the demo, I connect a LangChainJS agent to multiple Cloudflare MCP servers, including:

the Browser MCP to fetch and convert live web pages
the Docs MCP for up-to-date Cloudflare documentation
Cloudflare’s GraphQL MCP for analytics queries

Using Anthropic’s native MCP toolset and tool search, Claude explores the available tools, decides what to use, and executes them on demand.

From the outside, it just looks like a chat interface. Under the hood, that chat app has real operational access to my Cloudflare account.

Why this felt different

What stood out to me while building this is how little glue code was needed.

I didn’t have to:

model every API endpoint as a tool
maintain schemas as the platform evolves
preload hundreds of tools into the prompt

Instead, the agent discovers capabilities as it goes. Tools feel less like rigid contracts and more like something the agent can explore.

That’s a big shift if you’re thinking about agents interacting with real infrastructure, not just answering questions.

Why MCP matters

MCP is interesting not because it’s another protocol, but because it enables a different mental model:

platforms expose capabilities, not static tools
agents discover and load functionality dynamically
providers like Cloudflare can safely expose large surfaces without overwhelming the model

This feels much closer to how we’ll want agents to interact with production systems going forward.