Vercel just shipped Eve — a filesystem-first AI agent framework where the entire agent is just a folder. No graphs, no orchestrators, no config files fighting you. Instructions, tools, channels, and subagents are all just files on disk. In 100 seconds, here's how it works,...
Vercel just shipped Eve — a filesystem-first AI agent framework where the entire agent is just a folder. No graphs, no orchestrators, no config files fighting you. Instructions, tools, channels, and subagents are all just files on disk. In 100 seconds, here's how it works, how to build one, and whether it's actually ready for production.
Eve: https://github.com/vercel/eve
Most agent frameworks make you learn a new abstraction before you can ship anything — a graph, a state machine, a registry pattern. Eve drops all of that and goes filesystem-first instead: the directory you author IS the running agent. Write a markdown file for instructions, drop a TypeScript file in a tools folder, and the agent compiles it into a real backend behind a durable HTTP API. It's built by Vercel, ships on top of their AI Gateway, and defaults to Claude Sonnet out of the box.
This video covers the actual mechanics, not just the pitch: how the scaffolding CLI works, what the two required files do, how the filename-to-tool-name convention works under the hood, what the session/streaming API looks like, and how channels (like Slack) get bolted on. It also covers what's NOT ready yet, because this is a v0.9 beta released days ago and the API surface will move.
#llm #aiagents #vercel #programming #softwareengineering
