Uploaded May 2026 | Updated September 2026, 1 day ago
Asking an LLM to do tool calling is like asking Shakespeare to take a month-long crash course in Mandarin and then write a play in it. A year into MCP, that mismatch is the bottleneck — and Cloudflare's fix is to let the model do what it's actually good at: write code.
This talk covers MCP (Model Context Protocol), Code Mode, Cloudflare Workers, the agents SDK, remote MCP servers, LLM tool calling, AI agents, sandboxes, and Cloudflare Containers. One year in, the first wave of production remote MCP servers — Block, Figma, WorkOS, PayPal — all shipped on Cloudflare. The next phase is reframing how agents use tools: instead of forcing LLMs to call tools in a format they barely saw in training, Code Mode has the model emit code that calls the tools, executed inside Workers' lightweight V8 isolates that spin up instantly. The argument: open protocols thrive when they aren't owned, and Workers are the right primitive for the billions of agents about to need secure, instant compute.
Recorded at MCP Night, timed with the Agentic AI Foundation (AAIF) announcement that Cloudflare is part of. Rita Kozlov, VP of Product at Cloudflare leading the AI and developer product suites, gives a candid year-one retrospective on MCP — what worked, what's still rough, and why her team built Code Mode. She makes the case that LLMs trained on every open-source repo ever written are vastly better at writing code than at tool calling, a paradigm they've barely seen, so the right move is to lean into code generation as the tool-invocation layer.
What's covered:
- Why the first wave of remote MCP servers (Block, Figma, WorkOS, PayPal) all landed on Cloudflare
- The PayPal origin story: payments as a remote MCP server, day one
- The "31 minutes between meetings" anecdote — a CTO standing up an MCP server in the gap between two calls
- Kenton Varda's Shakespeare-learning-Mandarin analogy for why LLM tool calling is fundamentally constrained
- What Code Mode actually does: LLMs write code, code calls the tools, much better outputs than raw tool calls
- Why Workers' V8 isolates (sub-second cold start) are the right substrate for agent code execution at scale
- When to reach for Sandboxes plus Cloudflare Containers instead — e.g., full coding agents that need to install packages
- The case for open standards and why MCP's longevity depends on no single company owning it
Chapters:
0:00 Code Mode at MCP Night
0:18 The AAIF announcement
0:30 One year of MCP, in retrospect
1:18 Why open standards matter for protocols
2:09 First-wave remote MCP servers on Cloudflare
2:43 PayPal as the first payments MCP server
3:07 Agents SDK and the 31-minutes-between-meetings story
4:00 Lessons from building real agents
4:31 The Shakespeare-Mandarin analogy for tool calling
5:25 Code Mode: LLMs write code to call tools
5:51 Why Workers isolates fit agents
6:35 Agents SDK, Code Mode beta, Sandboxes
7:45 Where to learn more
Learn more:
agents.cloudflare.com
Search "Cloudflare Code Mode" for docs and the beta. Code Mode is currently in beta; Rita asks viewers to "go ham" when it hits GA.
If you're building AI agents, MCP servers, or anything on Workers, subscribe for more talks from the people designing this stack.
#MCP #ModelContextProtocol #Cloudflare #CloudflareWorkers #CodeMode #AIAgents #AIAgent #LLM #LLMs #AgentsSDK #Anthropic #OpenAI #RemoteMCP #MCPServer #Sandboxes #CloudflareContainers #AIInfrastructure #DeveloperTools #ToolCalling #RitaKozlov #MCPNight #AAIF #AgenticAI #V8Isolates #PayPal
Asking an LLM to do tool calling is like asking Shakespeare to take a month-long crash course in Mandarin and then write a play in it. A year into MCP, that mismatch is the bottleneck — and Cloudflare's fix is to let the model do what it's actually good at: write code.
This talk covers MCP (Model Context Protocol), Code Mode, Cloudflare Workers, the agents SDK, remote MCP servers, LLM tool calling, AI agents, sandboxes, and Cloudflare Containers. One year in, the first wave of production remote MCP servers — Block, Figma, WorkOS, PayPal — all shipped on Cloudflare. The next phase is reframing how agents use tools: instead of forcing LLMs to call tools in a format they barely saw in training, Code Mode has the model emit code that calls the tools, executed inside Workers' lightweight V8 isolates that spin up instantly. The argument: open protocols thrive when they aren't owned, and Workers are the right primitive for the billions of agents about to need secure, instant compute.
Recorded at MCP Night, timed with the Agentic AI Foundation (AAIF) announcement that Cloudflare is part of. Rita Kozlov, VP of Product at Cloudflare leading the AI and developer product suites, gives a candid year-one retrospective on MCP — what worked, what's still rough, and why her team built Code Mode. She makes the case that LLMs trained on every open-source repo ever written are vastly better at writing code than at tool calling, a paradigm they've barely seen, so the right move is to lean into code generation as the tool-invocation layer.
What's covered:
- Why the first wave of remote MCP servers (Block, Figma, WorkOS, PayPal) all landed on Cloudflare
- The PayPal origin story: payments as a remote MCP server, day one
- The "31 minutes between meetings" anecdote — a CTO standing up an MCP server in the gap between two calls
- Kenton Varda's Shakespeare-learning-Mandarin analogy for why LLM tool calling is fundamentally constrained
- What Code Mode actually does: LLMs write code, code calls the tools, much better outputs than raw tool calls
- Why Workers' V8 isolates (sub-second cold start) are the right substrate for agent code execution at scale
- When to reach for Sandboxes plus Cloudflare Containers instead — e.g., full coding agents that need to install packages
- The case for open standards and why MCP's longevity depends on no single company owning it
Chapters:
0:00 Code Mode at MCP Night
0:18 The AAIF announcement
0:30 One year of MCP, in retrospect
1:18 Why open standards matter for protocols
2:09 First-wave remote MCP servers on Cloudflare
2:43 PayPal as the first payments MCP server
3:07 Agents SDK and the 31-minutes-between-meetings story
4:00 Lessons from building real agents
4:31 The Shakespeare-Mandarin analogy for tool calling
5:25 Code Mode: LLMs write code to call tools
5:51 Why Workers isolates fit agents
6:35 Agents SDK, Code Mode beta, Sandboxes
7:45 Where to learn more
Learn more:
agents.cloudflare.com
Search "Cloudflare Code Mode" for docs and the beta. Code Mode is currently in beta; Rita asks viewers to "go ham" when it hits GA.
If you're building AI agents, MCP servers, or anything on Workers, subscribe for more talks from the people designing this stack.
#MCP #ModelContextProtocol #Cloudflare #CloudflareWorkers #CodeMode #AIAgents #AIAgent #LLM #LLMs #AgentsSDK #Anthropic #OpenAI #RemoteMCP #MCPServer #Sandboxes #CloudflareContainers #AIInfrastructure #DeveloperTools #ToolCalling #RitaKozlov #MCPNight #AAIF #AgenticAI #V8Isolates #PayPal
