Vercel Ships eve, an Open-Source Framework for Production Agents

On June 17, 2026, Vercel open-sourced eve, a TypeScript agent framework the company uses internally for v0 and its own agent fleet. The pitch mirrors what Next.js did for the web: stop hand-rolling the same production plumbing for every agent and ship a framework where durable execution, sandboxes, approvals, and evals are defaults, not weekend projects.

An Agent Is a Directory

eve is filesystem-first. A complete agent is a folder tree you can read at a glance:

agent/
  agent.ts              # model config
  instructions.md       # system prompt
  tools/                # typed tool definitions
  skills/               # domain knowledge files
  subagents/            # delegated agents
  channels/             # Slack, web chat, etc.
  schedules/            # cron-style autonomous runs

agent.ts configures the model (with AI Gateway fallbacks). instructions.md is the system prompt eve prepends to every call. Tools and skills are individual files eve wires automatically. No boilerplate runner, no custom state machine per project.

Production Built In

What separates eve from a thin LLM wrapper:

• Durable execution. Long-running agent work survives restarts and retries without you building checkpoint logic.
• Sandboxed compute. Tool execution runs isolated from the host process.
• Human-in-the-loop approvals. Gate high-risk actions before they commit.
• Subagents. Delegate subtrees of work to specialized agents with their own instructions and tools.
• Evals. Built-in evaluation hooks so you can regression-test agent behavior, not just unit-test functions.

Vercel's internal experience: once coding agents made building agents trivial, the company shipped hundreds of internal agents and hit the same wall every team hits. Reliability, observability, and deployment plumbing did not carry over project to project. eve is the extraction of what they rebuilt repeatedly.

Getting Started

Scaffold a new agent:

npx eve@latest init my-agent

Add --channel-web-nextjs if you want a Web Chat UI. The init command installs dependencies, initializes Git, and starts the dev server. Add tools with defineTool from eve/tools and Zod schemas for inputs.

The HTTP API exposes session creation (POST /eve/v1/session), streaming (GET /eve/v1/session/:id/stream), and follow-ups via continuation tokens. Model routing goes through Vercel AI Gateway with provider fallbacks configured in agent.ts.

Where eve Fits Vercel's Agent Strategy

eve complements rather than replaces the Vercel plugin for coding agents. The plugin gives Cursor, Claude Code, and Codex platform-specific knowledge for deploying on Vercel. eve is for building custom agents that run on Vercel infrastructure: internal ops bots, customer-facing assistants, scheduled report generators, and anything that needs subagents plus durable state.

Against LangGraph, Mastra, or raw Agent SDK usage, eve bets on convention over configuration: one directory layout, one dev server, one deployment path. The trade-off is less flexibility for exotic orchestration patterns in exchange for faster time-to-production for standard agent shapes.

Why Web Developers Should Care

If you already ship Next.js on Vercel, eve is the lowest-friction path from "agent demo in a notebook" to "agent with approvals, schedules, and a Slack channel in production." The filesystem layout also plays nicely with coding agents: Claude Code or Cursor can add a tool by creating a file in agent/tools/ without touching framework internals.

eve is open source on npm (npm install eve@latest). Docs ship inside the package at node_modules/eve/docs and on Vercel's site. Watch for deeper integration announcements at Vercel Ship on June 30.