Eve is Vercel's open-source framework for building durable AI agents as a directory of files. An Eve project keeps always-on instructions in Markdown, model and runtime choices in agent.ts, typed actions under tools, reusable procedures under skills, and delivery surfaces under channels. That filesystem-first layout makes agent behavior easier to inspect, review, and extend than a single opaque prompt or configuration blob.
The practical hook is that Eve is not just a chat wrapper. Vercel's docs and npm package frame it as a TypeScript runtime for sessions that can stream incremental output, resume after cold starts or long pauses, call typed tools, load skills on demand, connect to services, delegate to subagents, run on schedules, use sandboxes, and collect eval or run data in the Vercel dashboard. It fits teams that want agent projects to live beside application code while still using Vercel primitives such as AI Gateway, Workflows, Sandbox, and Connect.
Eve is still a new public beta, so production teams should treat the APIs, pricing surfaces, approval gates, sandbox policy, connection scopes, telemetry choices, and model credentials as things to verify before rollout. It is most compelling for TypeScript and Vercel-native teams that want deployable backend agents with production plumbing close at hand. Python-first, fully self-hosted, or vendor-neutral teams may prefer LangGraph, CrewAI, Mastra, Pydantic AI, or lower-level Vercel AI SDK patterns. Its early traction is strong for a brand-new Vercel project, but this page frames that as launch momentum rather than established independent adoption.
