MCP on Code Mode (Interview)

The Changelog: Software Development, Open Source1h 54mMay 15, 2026

Get the full intelligence

Search transcripts, export clips, track mentions, and explore all topics from “MCP on Code Mode (Interview)” inside PodZeus.

AI-Generated Summary

AI agents don’t need more tools—they need the ability to write and run code on demand, and Cloudflare’s Model Context Protocol (MCP) has cracked the code by turning agents into full-stack developers through server-side TypeScript execution. Matt Carey, Cloudflare’s MCP engineer, reveals how his team bypassed the crippling limitations of traditional tool calling by shifting to a dynamic code execution model: agents now write code that runs securely in isolated Cloudflare Workers, exposing all 2,500+ APIs with just 1,000 tokens of context—no more context bloat, no more rigid tool catalogs. This breakthrough, born from a Tuesday epiphany, enables agents to deploy a full Next.js app in a single prompt, with zero local code storage. But the real game-changer isn’t just the tech—it’s the mindset. Carey argues that true AI mastery comes not from chasing cloud APIs, but from building and testing in a home lab: Raspberry Pis, Proxmox servers, and self-hosted Linux environments. He envisions a self-learning DNS server, 'DNS Hole,' that uses a tiny local AI model to detect threats in real time—responding in under 10 seconds by analyzing domain patterns and escalating to deeper analysis, all without relying on outdated blocklists. This shift from reactive to proactive, agent-native infrastructure is the future of secure, scalable development. To unlock this future, developers must redesign their tools from the ground up.

Key Takeaways
1

Use server-side code execution via TypeScript in isolated Cloudflare Workers to expose 2,500+ APIs with just 1,000 tokens of context.

2

Replace rigid tool calling with dynamic code generation: agents write code that runs securely on the server, eliminating context bloat.

3

Implement a three-step planning loop—plan, review for blind spots, then fix—to dramatically improve agent reliability and reduce sabotage.

4

Build memory systems as provider-agnostic APIs that support search, compaction, and continual learning across storage backends like Durable Objects or DuckDB.

5

Use DuckDB + Parquet for lightning-fast local data analysis—ideal for querying telemetry, book content, or agent logs without complex setup.

…and 3 more takeaways available in PodZeus

Chapters
0:00
10 min

Welcome to the Pod: Cloudflare's Agent Revolution

Adam introduces Matt Carey from Cloudflare, setting the stage for a deep dive into Code Mode and MCP. The episode kicks off with sponsor reads for Coder.com and Tailscale, highlighting secure cloud development environments and mesh networking.

10:00
10 min

The Problem with Local Machines and Tool Calling

Matt explains why local developer machines are a security and consistency nightmare—leading to 'it works on my machine' issues, supply chain risks like the Shai Halud virus, and long onboarding times. He contrasts this with cloud development environments that enable instant recovery and enforced security.

20:00
20 min

Demystifying MCP: From Local Tools to Server-Side Code

Matt breaks down MCP (Model Context Protocol) as a standardized way for agents to access tools, prompts, and resources. He critiques the traditional one-to-one tool mapping that limits scalability and introduces Code Mode as a paradigm shift: letting models write code to interact with APIs dynamically.

40:00
20 min

The Breakthrough: Server-Side Code Mode at Cloudflare

We have that massively reduced context window allowance and we can use it to enable our MCP server to access the whole of the Cloudflare platform.

Highlight
1:00:00
20 min

The Workflow: From Idea to Deployment in Minutes

Matt describes his personal workflow: using Claude Opus 4.6 with a custom Git wrapper for safety, avoiding one-on-one prompting to prevent sabotage, and relying on a structured planning loop to generate robust, reviewable plans before execution.

High-Impact Quotes
I kind of wanted to work at Cloudflare about midway through this podcast. You know, I feel like I can make a dent there.
The Changelog Host114:01
Viral: 88.0
I can unleash my agents. I can build something, deploy it to that, and battle test it in almost real time on my own hardware.
Matt Carey91:00
Viral: 88.0
drink the Cloudflare Kool -Aid too long, man. Get your own VM, get your own Linux, play with your own keys, with your own rules, with your own sudo and feel the metal, man. Feel the metal.
Matt Carey93:58
Viral: 88.0
Speakers

Hosts

AdamThe Changelog Host

Guest

Matt Carey
Topics Discussed
code mode95%home labbing95%mcp apis95%dns security92%model context protocol90%agent-native cli90%server side code execution88%local ai models88%duckdb and parquet85%non-interactive cli85%agent memory85%cloudflare api80%open source show notes75%
People & Brands

Cloudflare

organization

39xPositive

Matt Carey

person

29xNeutral

Claude

product

18xPositive

Proxmox

product

8xPositive

DuckDB

product

8xPositive

agent

other

7xPositive

CLI

other

6xNeutral

Raspberry Pi

product

6xPositive

Pi.dev

product

5xPositive

Opus 4.6

product

5xPositive

Get the full intelligence

Search transcripts, export clips, track mentions, and explore all topics from “MCP on Code Mode (Interview)” inside PodZeus.

Start discovering podcast insights today

Start with a 7-day trial and explore a growing catalog of popular podcasts. No credit card required.

No credit card required • 7-day trial • Cancel anytime