Vercel Launches eve: The "Next.js for Agents" Is Here, and It's Open Source
By AgentRiot Desk
Vercel's new filesystem-first framework treats every AI agent as a directory of files, bundling durable execution, sandboxed compute, and multi-channel deployment into a single open-source package.
Vercel Launches eve: The "Next.js for Agents" Is Here, and It's Open Source
Vercel's new filesystem-first framework treats every AI agent as a directory of files, bundling durable execution, sandboxed compute, and multi-channel deployment into a single open-source package.
On June 17, 2026, at Vercel Ship in London, the company that redefined frontend deployment did something unexpected: it open-sourced the exact framework it uses internally to build and run production AI agents. The project is called eve, and Vercel describes it with a single, audacious tagline: "Like Next.js for web apps, but for agents."
Five days later, the GitHub repository sits at 2.2k stars, 161 forks, and 47 pull requests. Not bad for a project that launched in public preview with a simple npx eve@latest init my-agent command.
What eve Actually Is
eve is a filesystem-first, TypeScript-native framework where an AI agent is literally a directory on disk. The directory is the contract. You write instructions in Markdown, define tools in TypeScript, and the framework compiles the whole thing into a durable production service.
Here's the structure Vercel expects:
| File/Directory | Purpose |
|---|---|
instructions.md | System prompt and agent identity, written in Markdown |
agent.ts | Model configuration and runtime settings |
skills/ | Reusable Markdown playbooks loaded contextually |
tools/ | TypeScript files that become callable tools (filename = tool name) |
sandbox/ | Isolated execution environment customization |
channels/ | Slack, Discord, Teams, web chat, WhatsApp, and more |
connections/ | Authenticated service integrations (GitHub, Stripe, Linear) |
subagents/ | Specialized child agents the main agent can delegate to |
schedules/ | Cron-based durable workflows |
No registration boilerplate. No YAML ceremony. Add a TypeScript file to tools/ and the model can call it. That's the whole API.
The Infrastructure You Don't Have to Build
Where eve gets interesting is what ships built in, not as plugins or separate services, but as first-party primitives:
- Durable execution via Vercel Workflows: agents checkpoint their state between steps, park when idle, and resume on the next message without losing context.
- Sandboxed compute via Vercel Sandbox: every agent runs in an isolated environment with file tools.
- AI Gateway handles model routing, provider fallbacks, and streaming automatically.
- Vercel Connect provides MCP and HTTP endpoint connections with managed authentication.
- Multi-channel deployment surfaces the same agent in Slack, Discord, Microsoft Teams, Google Chat, WhatsApp, and web chat via the Chat SDK.
- Approvals and evals provide built-in governance and observability for production deployments.
Vercel CTO Malte Ubl has been vocal about the framework on social media, noting that he "spun [an agent] up pretty fast using @vercel Eve framework" and highlighting that Vercel handles complex Slack API scopes internally so developers don't have to.
The "Next.js for Agents" Claim, Examined
The comparison is deliberate. Next.js abstracted away routing, SSR, and deployment configuration into a filesystem convention. eve does the same for agent infrastructure: durable state, sandboxed execution, channel wiring, and tool registration disappear into the directory structure.
But there's a critical difference in business model. Next.js is tightly coupled to Vercel's hosting platform. eve, by contrast, is Apache 2.0 licensed and designed to run anywhere TypeScript runs. You only pay Vercel if you deploy on their infrastructure (Functions, sandboxes, AI Gateway). Self-hosting is explicitly supported.
This is a notable strategic choice. In a market where competitors are increasingly bundling frameworks with proprietary cloud services, Vercel is betting that open-source ubiquity will drive enterprise adoption of its paid infrastructure. That playbook worked extraordinarily well for Next.js.
The Shadow AI Problem Vercel Is Solving
Alongside eve, Vercel introduced Passport and Vercel Connect, enterprise products designed to put employee-built AI agents behind corporate identity providers. The Register's coverage of the Ship event highlighted this angle: Vercel is trying to fix "shadow AI," the proliferation of unauthorized AI tools employees build and deploy without IT oversight.
Passport gives enterprises governance. eve gives developers velocity. The combination is Vercel's pitch for how agentic development should happen inside regulated organizations.
What Developers Are Actually Building
Vercel claims it already runs 100+ agents on eve internally, including data pipeline automation, content moderation, and customer support routing. The company also released a content agent template that writers can @mention in Slack to load per-surface style skills and generate draft content.
Early community adoption shows similar patterns: Slack-based assistants, scheduled reporting agents, and multi-step workflow automation with human-in-the-loop approvals. The framework's subagent delegation, where a main agent parcels out tasks to specialized child agents and synthesizes results, is particularly useful for complex, multi-domain operations.
How It Stacks Up
eve enters a crowded field. LangGraph, CrewAI, AutoGen, and the OpenAI Agents SDK all offer agent orchestration. Vercel's differentiators are:
- Filesystem-first design: No Python-class inheritance or graph DSL, just files in a directory.
- Production infrastructure bundled: Durable execution and sandboxing are framework primitives, not external services to wire up.
- TypeScript-native: The entire stack is typed, from tool definitions to channel configurations.
- Multi-channel by default: One agent, many surfaces, not a web API you have to wrap.
The tradeoff is ecosystem maturity. LangGraph has years of community tooling. CrewAI has extensive role-based patterns. eve has five days of public availability and 77 commits on main. For teams already invested in Vercel's platform, the integration advantage is obvious. For everyone else, the self-hosting promise will need to prove itself.
Getting Started
The onboarding is deliberately minimal:
npx eve@latest init my-agent
This scaffolds a directory with instructions.md, a basic tool file, and channel stubs. Add a model in agent.ts, write your first tool in tools/, and deploy with vercel deploy, or self-host if you prefer.
The framework is in public preview. Vercel has not announced a general availability timeline, but the commit velocity (daily merges, active issue triage) suggests rapid iteration.
The Bottom Line
Vercel is making a bet that the future of software isn't just AI-assisted. It's agent-native. eve is their attempt to own the application layer for that transition, just as Next.js owns the frontend layer for modern web development. The open-source licensing and self-hosting support lower the barrier to experimentation, while the tight integration with Vercel's infrastructure stack creates a clear upgrade path to production.
Whether it becomes the Next.js of agents or merely a compelling option in a crowded field depends on whether developers find the filesystem-first model as intuitive for agents as it was for web pages. Five days in, 2.2k GitHub stars suggests the proposition is resonating.
License: Apache 2.0
GitHub: github.com/vercel/eve
npm: eve
Launch date: June 17, 2026 (Vercel Ship, London)
Status: Public preview
Source notes: Vercel official blog and changelog, GitHub repository (vercel/eve), The New Stack, The Register, MarkTechPost, TechTimes, and social posts from Vercel CTO Malte Ubl. GitHub metrics current as of June 22, 2026.