# Eve

> Eve is Vercel's open-source agent framework that lets developers build, run, and scale production AI agents using a file-based convention—no infrastructure assembly required.

Eve is an open-source agent framework from Vercel, designed to let developers build, run, and scale AI agents in production without manually assembling infrastructure. It draws an explicit parallel to Next.js: just as a Next.js folder becomes a route by convention, an Eve file becomes an agent capability by convention, with no boilerplate or manual wiring required. The public preview is open now, and the CLI wizard gets a developer from nothing to a running local dev server in under a minute.

## What It Is

Eve is a file-first agent framework where an agent is defined as a directory of files, each describing one component of the agent's behavior. `agent.ts` sets the model and options; `instructions.md` is the system prompt prepended to every model call; files under `tools/` are typed TypeScript functions using `defineTool` with Zod schemas; files under `skills/` are markdown documents loaded contextually by the model; and `subagents/`, `channels/`, and `schedules/` directories extend the agent further. Because everything is a file, the entire agent lives in Git with diffs, reviews, and history like any other codebase.

## Production Infrastructure Built In

Eve ships with durable execution, a sandboxed compute environment, human-in-the-loop approvals, multi-channel support, tracing, and evals—all included by default.

- **Durability** is built on Vercel's open-source Workflow SDK: every conversation is checkpointed step by step, so a session can survive a crash or redeploy and resume exactly where it stopped.
- **Sandboxing** isolates agent-generated code from the application runtime entirely; locally it uses Docker, microsandbox, or just-bash, and in production it switches to Vercel Sandbox automatically.
- **Approvals** are a single field on a tool definition (`needsApproval`), which can accept a function to evaluate dynamically; the agent pauses indefinitely without consuming compute until a person responds.
- **Tracing** produces a standard OpenTelemetry span tree that exports to Braintrust, Honeycomb, Datadog, Jaeger, or any other tracing service, and surfaces in Vercel's Observability tab under "Agent Runs" when deployed.
- **Evals** let developers write scored test suites alongside the agent, checking tool calls, rule adherence, and expected reply content; these can run locally or be wired into CI as a deploy gate.

## Connections, Channels, and Scheduling

Agents connect to external services through connection files that point at MCP servers or any OpenAPI-compatible API, with Eve brokering auth so the model never sees credentials or URLs. At launch, Vercel states supported integrations include Slack, GitHub, Snowflake, Salesforce, Notion, and Linear, plus any OAuth or API-key-protected service.

Channels work by convention: running `eve channel add slack` writes a single `channels/slack.ts` file, and the deployed agent immediately answers in Slack with platform-native UI elements like approval buttons and select menus. Schedules are one additional file with a cron expression, which on Vercel deploys automatically as a Vercel Cron Job.

## How Vercel Uses Eve Internally

Vercel states it runs over a hundred agents in production on Eve today. According to Vercel's own published claims:

- **d0**, the most-used internal tool, handles more than 30,000 questions a month by letting any employee ask warehouse data questions in Slack, scoped to their own permissions.
- **"Lead Agent"**, an autonomous SDR, works every inbound lead immediately and follows up independently.
- **"Vertex"**, the support agent, resolves 92% of support tickets on its own across the help center, docs, and Slack, escalating the remaining 8% to the human team.
- **"V"**, a routing agent, receives all Slack tasks and dispatches them to whichever specialist agent in the fleet is best suited.

## Current Status: Public Preview

Eve is currently in public preview. Vercel states that a year ago agents triggered less than 3% of deployments on Vercel, and that figure has risen to around 29%, with agents expected to account for half of all deployments in the near future. Source code and discussions are public at `github.com/vercel/eve`, and documentation is available at `eve.dev/docs`, with the docs also bundled inside the npm package at `node_modules/eve/docs` once installed. Eve works with any model and any MCP server, and support for deployment targets beyond Vercel is described as forthcoming.

## Features
- File-based agent definition (agent.ts, instructions.md, tools/, skills/)
- Durable execution with step-by-step checkpointing via Vercel Workflow SDK
- Sandboxed compute environment (Docker, microsandbox, or Vercel Sandbox in production)
- Human-in-the-loop approvals via needsApproval field on tool definitions
- Multi-channel support (Slack, and more) via convention-based channel files
- Cron-based scheduling via schedule files deployed as Vercel Cron Jobs
- MCP server and OpenAPI-compatible API connections with brokered auth
- OpenTelemetry-based tracing exportable to Braintrust, Honeycomb, Datadog, Jaeger
- Scored eval test suites runnable locally or in CI
- Preview deployments per commit with channel support
- Local dev server with terminal UI showing real-time step traces
- TypeScript tool definitions with Zod schema validation
- Skill files (markdown) loaded contextually by the model
- Subagent support via subagents/ directory
- Works with any LLM model and any MCP server

## Integrations
Slack, GitHub, Snowflake, Salesforce, Notion, Linear, Braintrust, Honeycomb, Datadog, Jaeger, Docker, MCP servers, OpenAPI-compatible APIs, Vercel Sandbox, Vercel Cron Jobs, Vercel Observability

## Platforms
WINDOWS, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Open Source, Free tier available

## Links
- Website: https://vercel.com/eve
- Documentation: https://eve.dev/docs
- Repository: https://github.com/vercel/eve
- EveryDev.ai: https://www.everydev.ai/tools/eve