Matt Carey is a senior systems engineer on Cloudflare’s agents team and a maintainer of the Model Context Protocol TypeScript SDK. He builds infrastructure that lets AI agents discover APIs, generate useful code, and execute it without exposing credentials or overwhelming their context windows.
Before entering software, Carey competed internationally as a windsurfer for Malta and coached athletes. He studied at the University of Bristol, where his projects included reinforcement learning for autonomous hydrofoil control. From 2022 to 2024, he worked on AWS serverless systems and cloud migrations at aleios, became an AWS Community Builder, and helped revive Serverless London. He subsequently joined StackOne as its founding AI engineer, developing API integrations and internal agents, before joining Cloudflare in 2025.
Progressive tool discovery:Cloudflare’s API includes more than 2,500 endpoints, making it impractical to load every operation into an agent’s context. Carey’s Code Mode architecture replaces sprawling tool inventories with two operations, search and execute, compressing access to approximately 1,000 tokens. Agents discover relevant capabilities as needed and generate code against typed interfaces derived from API specifications.
Capability-restricted execution: Generated code runs inside isolated Dynamic Workers without automatic access to secrets, unrestricted networking, or a full shell. Explicit permissions determine which services it can reach. Carey also emphasizes API rate limiting: agents can multiply requests across loops and execution environments faster than human operators.
Cloudflare Computer: Developed with Aron Carroll, Cloudflare Computer gives agents access to files, commands, tools, and executable programs while moving between lightweight isolates and Linux containers according to each task’s requirements.
Git-native agent infrastructure: Carey helped develop Artifacts, Git-compatible versioned storage for agents and automations. His open-source Shippie automates code review and quality assurance, while zagi provides agent-oriented Git commands, compact output, worktree support, attribution, and safeguards against destructive operations.
A lightweight Model Context Protocol: His MCP engineering work advances a stateless core, simpler Workers deployments, and SDK migration. He envisions application frameworks exposing MCP alongside existing APIs, with clients discovering capabilities progressively and preserving generated scripts for repeated tasks.
Making every API available to agents requires more than publishing tools: capabilities need progressive discovery, and generated programs need controlled execution.
A broken serverless hit counter leads to a broader design: persistent agents coordinate work, while isolated Workers execute generated code with explicitly granted capabilities.
