Vercel Open Sources Eve — Agents are just folders!

Today we’re releasing Eve, an Apache-2.0 licensed agent framework that lets you write AI agents as folders of files. We released it as an npm package on June 17th (in public preview), under the Apache-2.0 license. It’s not sexy — there’s no magic here. What Eve does is letting you define your agent’s structure as folders on disk instead of registering your tools and capabilities via code.

In practice, an Eve agent is a folder containing an agent.ts (model + config), an instructions.md (system prompt) and folders corresponding to capabilities (tools/, skills/, connections/, channels/, schedules/, subagents/, sandbox/ and lib/). To scaffold a new agent, simply run npx eve@latest init my-agent and deploy it with vercel deploy, keeping the same environment as in development. By adding files to the appropriate folder, you can give your agent new capabilities without any additional wiring.

1The hard parts come built in

The key insight here is that Eve provides much more than just folder structure. To support durable conversations with checkpoints, Eve leverages Vercel’s open Workflow SDK to allow sessions to be interrupted, restarted after a crash or redeployment, and resumed. For sandboxing untrusted code, Eve uses Vercel Sandbox (in production) and Docker (locally). For requesting human input, Eve lets you block an agent until an action is approved by a human. For tracing, Eve uses OpenTelemetry spans which can be sent to your favorite backends such as Datadog, Honeycomb, or Jaeger. Finally, evals can be executed locally or in CI.

2Why this matters for you

If you’ve ever worked on deploying an agent, you know that 10% of the work is building the demo. 90% of the work is building the plumbing. You have to deal with retries when steps fail, sandboxing untrusted code, blocking until the user approves a risky action, recording what your agent did, and avoiding vendor lock-in. Eve’s vision is that these concerns should be solved by infrastructure, not by each team implementing them independently. At Vercel, we’re already running over 100 agents (a data analyst answering more than 30,000 questions per month, a sales agent we estimate to be worth 32x our investment, a support agent closing 92% of tickets) using this framework. Of course, take these figures with a grain of salt (they are just another vendor’s marketing material), but at least you can verify that the framework is open source and can be inspected on disk.