Eve: Filesystem-First Framework for Durable Agents

Vercel's new open-source framework, eve, takes a stance that feels almost old-fashioned in the age of black-box agent platforms: put everything in the filesystem. No proprietary dashboard, no hidden state living in some vendor's database. Just a folder called agent/ with an instructions.md file for the system prompt, a tools/ directory for typed functions, skills/ for on-demand procedures, channels/ for Slack or Discord integrations, and schedules/ for cron jobs.

Getting started is a single command: npx eve@latest init my-agent scaffolds the project, installs dependencies, sets up Git, and drops you into an interactive terminal UI. You can also bolt eve onto an existing app by running the init command with a path instead of a fresh project name. The documentation ships inside the npm package itself, sitting in node_modules/eve/docs, which is a small but telling detail — it means coding agents and IDE assistants can read the docs locally without a network call.

The example Vercel walks through is intentionally trivial: a weather assistant that returns mocked data. You write a one-line instructions.md telling the model to be concise and admit the data is fake, define a get_weather tool using Zod for input validation, and pick a model — Claude Sonnet 5 is the default suggestion — in agent.ts. Restart, and you've got a working agent that can call a tool, all defined in files you could open in any text editor and diff in Git.

That last point is really the pitch. Because everything is a conventional file, you get inspectability and version control for free, plus the ability to layer in human-in-the-loop approval steps, subagents, and scheduled jobs without learning a new abstraction each time. Vercel is running eve as a beta under its usual beta terms, so the APIs and even the folder layout could shift before general availability, and the project leans on GitHub Discussions for community support rather than a dedicated forum.