Vercel Eve
What is Vercel Eve
Vercel Eve is an open-source, filesystem-first TypeScript framework for building durable backend AI agents. Announced at Vercel Ship 2026 (June 17), it's described as "Next.js for agents" — a directory of files compiles into a production-ready agent with durable execution, sandboxed compute, and multi-channel delivery. Built on AI SDK 7 and the broader Vercel infrastructure.
- Status: Public Beta (v0.17.0)
- License: Apache 2.0
- Downloads: ~257K weekly
- GitHub stars: 2.9K
- Contributors: 25+
Why it's interesting
- Filesystem-first authoring — an agent is a directory; each file name becomes a tool name. Zero registration boilerplate.
- Durable execution — built on Vercel Workflow SDK; sessions survive crashes, deploys, and cold starts.
- Sandboxed compute — Vercel Sandbox in production, Docker locally. Code execution is isolated by default.
- Human-in-the-loop — any tool can require approval; the agent pauses without consuming compute until a human responds.
- Multi-channel delivery — one codebase deploys to HTTP, Slack, Discord, Microsoft Teams, Telegram, Twilio, and more. Adding a channel = adding one file.
- Subagents & Skills — delegate work to child agents with their own prompts and tools; markdown "skills" (playbooks) are loaded on demand to keep context windows lean.
- Vercel ecosystem integration — AI SDK 7, Vercel Connect (OAuth), AI Gateway (model routing), Observability (OpenTelemetry), Workflows. Deploy with
vercel deploy. - MCP client support — agents can consume MCP servers as tool sources, bridging the MCP ecosystem.
- Cron schedules — agents can run on a schedule without external triggers, deploying as Vercel Cron Jobs for autonomous background tasks.
When to use
- Building durable, long-running AI agents that must survive deploys and crashes
- Teams already on Vercel wanting an opinionated agent framework with zero-config deployment
- Multi-channel agent delivery (Slack bots, Discord bots, API endpoints) from a single codebase
- Projects needing human-in-the-loop approval workflows for agent actions
- Scheduled/cron-based autonomous agent tasks
When NOT to use
- Frontend AI features — Eve is a backend agent framework. For AI-powered UI (chat, streaming text), use Vercel AI SDK or TanStack AI instead.
- Non-TypeScript teams — Eve is TypeScript-only. Python teams should look at LangGraph, CrewAI, or similar.
- Vercel lock-in concerns — heavily coupled to Vercel infrastructure (Workflows, Sandbox, Connect). Self-hosting is not a first-class path today.
- Simple, stateless tool calls — if your agent is a single prompt + tool call with no durability needs, Eve's machinery is overkill. AI SDK alone suffices.
- Early-stage risk aversion — Public Beta with rapid iteration (82 npm versions in ~2 weeks). The API surface is still moving.
Comparison to similar tools
| Vercel Eve | AI SDK (manual agents) | LangGraph | |
|---|---|---|---|
| Language | TypeScript | TypeScript | Python |
| Authoring model | Filesystem-first (directory = agent) | Code-first (imperative) | Graph-first (nodes + edges) |
| Durability | Built-in (Workflow SDK) | DIY | Built-in (checkpointing) |
| Sandboxing | Vercel Sandbox / Docker | DIY | DIY |
| Multi-channel | Built-in (Slack, Discord, Teams…) | DIY | DIY |
| Human-in-the-loop | Built-in | DIY | Built-in |
| Deployment | vercel deploy | Any | LangSmith / self-host |
| Maturity | Public Beta | Stable (v6/v7) | Stable |
Eve is opinionated and batteries-included for Vercel users. If you need more control or aren't on Vercel, building on AI SDK directly or using LangGraph gives more flexibility at the cost of more setup.