Durable execution
Workflows survive crashes and restarts. Every step is checkpointed. Agents park when waiting and resume on the next message.
Vercel Agent Framework
The framework for agents.
Like Next.js for web apps, but for agents. Define an AI agent as a directory — instructions and skills in Markdown, tools in TypeScript, durable workflows by default.
npx eve@latest init my-agent
- Open source
- Durable by default
- TypeScript tools
- Deploys anywhere
agent running
~/my-agent
Anatomy
Your agent/ is a directory.
An instructions.md file is all you need to run an agent. Skills, tools,
channels, and the rest are optional building blocks you add as it grows. Pick an entry — see exactly what
the file looks like.
Composition
Leverages every Vercel AI primitive.
One folder, wired into the whole platform. Your agent.ts talks to models through
AI Gateway, steps survive on Vercel Workflows, code executes in Vercel Sandbox, services connect with Vercel
Connect, and every channel ships through the Chat SDK. Select a node to see how it plugs in.
Production
Everything you need for production agents.
Durability, sandboxing, human-in-the-loop, and evals are built into the framework. Focus on building your agent. Every demo below is live — poke at it.
Sandboxed compute
Agents run code in isolated sandboxes. File system access, bash execution — fully isolated, snapshot in, snapshot out.
Multi-channel delivery
One agent codebase, surfaced everywhere. Channel files route the same agent to Slack, Discord, Teams, web chat, WhatsApp, API, Twilio, cron, and Linear.
Human-in-the-loop
Tools that need confirmation trigger approval gates. Sessions park until resolved, then resume seamlessly. Decide below.
Subagents
Delegate specialized work to child agents with their own prompts, tools, and sandbox. Results stream back and merge.
Evaluations
Define test suites with scoring rubrics. Run evals on every deployment and on a schedule, and catch regressions as the agent evolves.
Integration
Works natively with React and Next.js.
Wrap your config with withEve() and the agent mounts into your existing app.
Same dev server, same deploy. useEveAgent() finds its routes on its own — no CORS to configure,
no URL env vars to keep in sync.
import { withEve } from "eve/next";
const nextConfig = {};
// Agent + app: one dev server, one deploy.
export default withEve(nextConfig);
"use client";
import { useEveAgent } from "eve/react";
export function Chat() {
// Same-origin routes, found automatically.
const agent = useEveAgent();
// agent.messages, agent.sendMessage, ...
}
Ownership
Open SDKs, yours to self-host.
Swap any backend and self-host the whole runtime, with zero managed-infrastructure dependencies. Postgres-backed durability, Docker sandbox, one deploy — toggle the column to compare.
| Capability | Managed |
|---|---|
| Runtime durable execution, state, streaming | Vercel Agent Runtime |
| Durable workflow checkpointed steps, park & resume | Vercel Workflows |
| Model calls unified access, fallback, spend caps | AI Gateway |
| Sandbox isolated code execution | Vercel Sandbox |
| Connections MCP / HTTP endpoints | Vercel Connect |
| Channels where the agent is surfaced | Chat SDK |
You can deploy anywhere. Postgres-backed durability, Docker sandbox, Ansible deploy, zero managed services.
Build your first agent today.
One command, one folder, one deploy. Your agent is durable before you write a line of orchestration code.
npx eve@latest init my-agent