Vercel Releases Eve: An Open Framework for AI Agents Where Each Agent Is a List of Force-Mapped Files

Vercel is released the nightan open source framework for building, running, and scaling agents. The project is published as an npm package evelicensed under Apache-2.0.

Building an agent should mean explaining what it does. It should not mean assembling all the plumbing that the agent needs to work on in production.

eve is the framework Vercel builds and uses for its agents. According to the Vercel post, it employs more than 100 agents in production today.

What is eve?

eve is the first file system framework for long-lived agents. You create the agent as a directory on disk. A directory is a contract.

Each file defines one component of the agent. At a glance, the tree shows what the agent is and what it does. It also shows where he lives and where he makes himself.

The smallest agent that works with two files. One sets the model. One sets the instructions.

// agent/agent.ts
import { defineAgent } from "eve";

export default defineAgent({
  model: "anthropic/claude-opus-4.8",
});

The model is single-line, and provider regression is supported through AI Gateway. I instructions.md file becomes a system command that eve places before every model call.

An agent is a directory

Vercel’s main idea is that agents have shape. Each team keeps rebuilding the same structure to meet the same needs. eve made that situation into a frame.

The directory structure shows each skill in a folder. Here is the contract:

You can add a tool, skill, channel, or schedule by adding a file. Eve takes them during construction and installs the cables. There is no subscription plan.

The tool is a single TypeScript file with the Zod input schema. The name of the file and its location in the tree become its description.

// agent/tools/run_sql.ts
import { defineTool } from "eve/tools";
import { z } from "zod";

export default defineTool({
  description: "Run a read-only SQL query.",
  inputSchema: z.object({ sql: z.string() }),
  needsApproval: ({ toolInput }) => estimateScanGb(toolInput.sql) > 50,
  async execute({ sql }) { /* ... */ },
});

What ships in the box

Vercel describes the early date as ‘batteries included.’ Six production capabilities come with the framework:

• Making it last longer: Every conversation is a long-term workflow, each step has a test. A session can rest, survive a crash or use, and resume where it stopped. This is built on the open source Workflow SDK.
• A sandboxed computer: Agent-generated code is considered untrusted. Each agent gets its own sandbox of shell commands, scripts, and file read and write. The backend is an adapter, which works in Vercel Sandbox when it is also used in Docker, microsandbox, or just bash locally.
• Human permissions inside the loop: Any action can be set to require approval. The agent stops there and waits, indefinitely if necessary, without using the computer. Once approved, eve continues where it left off.
• Secure the connection: A connection is a file that points to an MCP server or an OpenAPI-compatible API. eve brokers the auth, and the model never sees the URL or the details. At launch, agents can connect to Slack, GitHub, Snowflake, Salesforce, Notion, and Linear.
• Channels: The same agent works everywhere. The HTTP API is enabled by default, including Slack, Discord, Teams, Telegram, Twilio, GitHub, and Linear. One channel can transmit to another.
• Follow up and evals: All runs generate traces using standard OpenTelemetry parameters. They ship to Braintrust, Honeycomb, Datadog, or Jaeger. Evals are test suites that you get that you run locally or connect to CI.

Use cases, with real examples

Vercel published six internal agents for the evening:

• d0, data analyst: Its most used internal tool, handling more than 30,000 queries per month. Every query is accessed from the requester’s permissions.
• Lead Agent, independent SDR: It processes all new leads and tracks itself. Vercel says it costs about $5,000 a year and returns 32 times, maintained part-time by one engineer.
• Athena, commercial cockpit: RevOps builds in six weeks without developers. Answers pipeline questions from Snowflake and Salesforce in simple language.
• Vertex, support engineer: Manages help center tickets, documents, and Slack. Vercel reports that it resolves 92% of tickets on its own and escalates the rest.
• draft0, content agent: Uses a review queue that catches obvious problems before a human editor sees the piece.
• V, routing agent: Tasks go to V in Slack first. V each routes to an agent you can answer.

Interactive Simulation