What Is Vercel eve? The “Next.js for AI Agents”

Vercel launched eve at Ship London on June 17, 2026 — an Apache-2.0 open-source framework where every AI agent is a directory of files. It’s the most significant agent framework launch of June 2026, and it ships with first-party durable execution, sandboxes, and an AI Gateway. Here’s what it is, what makes it different, and when to choose it.

Last verified: June 22, 2026.

TL;DR

• Launched: June 17, 2026 (Vercel Ship London, public preview).
• License: Apache-2.0, open source.
• Language: TypeScript.
• Core idea: Every agent is a directory of files (instructions.md, tools/, skills/). eve reads the directory, validates a manifest, runs it.
• Infrastructure: Durable execution on Vercel Functions, sandboxed compute, AI Gateway, multi-channel deployment.
• Companion products: Vercel Passport (OIDC identity for agents), Vercel Connect (short-lived OAuth credentials).
• Best for: TypeScript teams who want agents to be diff-able artifacts reviewable in pull requests.
• Compared to: LangGraph (code-defined graphs), CrewAI (code-defined roles), OpenClaw (config + skills directories — closest sibling).

What eve actually is

For three years, every agent framework has had the same shape: import the library, instantiate an agent class, register tools as decorated functions, and deploy the result as a long-running process. The “agent” exists at runtime — there is no canonical file you can point at and say this is the agent.

eve flips that. An eve agent is a directory of files. That directory is the canonical artifact. The framework reads it, validates it, and runs it.

A minimal eve agent looks like:

my-agent/
├── instructions.md         # System prompt + persona
├── agent/
│   ├── tools/
│   │   ├── search-docs.ts  # Each .ts file = one tool
│   │   └── send-email.ts
│   └── skills/
│       ├── onboarding.md   # On-demand playbooks
│       └── refunds.md

Run eve dev and you have a working agent. Deploy it to Vercel and you have a durable HTTP endpoint that survives restarts.

Why “agents as directories” matters

Three reasons this is interesting beyond aesthetics:

1. Agents become reviewable in pull requests. A change to an agent is a file diff. A team can review “this agent now has access to the delete-customer tool” as a one-line PR, not as a code change buried in a Python module.

2. Skills become importable. An eve skills/ directory is just Markdown. You can copy refunds.md from one agent to another. You can fork an open-source agent and modify three files. The unit of reuse is the file, not the code module.

3. Tooling can be agent-agnostic. eve ships an evals harness, an observability layer, and a CLI that all work against the directory shape. Any agent that follows the layout gets them for free.

The closest existing comparison is OpenClaw, which already uses a skills/ directory pattern. eve is more aggressive about making the entire agent a filesystem layout, and it ships with cloud infrastructure (Vercel Functions, AI Gateway) that OpenClaw delegates to the host.

Built-in infrastructure

eve doesn’t just define the layout — it ships the runtime:

Capability	What it does
Durable execution	Agent runs survive restarts, retries, and long tool calls via Vercel Functions checkpointing
Sandboxes	Each tool call can run in an isolated sandbox (code execution, browser, file I/O)
AI Gateway	Single endpoint that routes to OpenAI, Anthropic, Google, OpenRouter — model swap is a config change
Subagents	Nested agents declared as subdirectories; parent passes a task and yields
Evals	agent/evals/*.ts files run as test cases; included in CI
Channels	Built-in adapters for HTTP, Slack, GitHub, Webhook, CLI; one agent serves multiple surfaces
Approvals	Tools can require human approval; eve surfaces approval requests through any channel

The enterprise side: Passport and Connect

Vercel released two paid enterprise products alongside eve (covered separately in our Vercel Passport piece):

• Vercel Connect (public beta): Replaces static API keys with short-lived OAuth/OIDC credentials. Tools authenticate at runtime via OIDC and receive task-scoped tokens to systems like Slack, GitHub, Snowflake, Salesforce. Reduces blast radius if a tool is compromised.
• Vercel Passport: Puts internal apps and agents behind a corporate identity provider (Okta, Microsoft Entra) by default. Configure the IdP connection once; Passport applies it across every deployment. Audit logs, central policy.

This is Vercel positioning against “shadow AI” — engineers deploying internal agents that bypass corporate identity governance. Passport pulls agents back under the same SSO umbrella as internal SaaS.

When to choose eve

Choose eve if:

• You’re a TypeScript team. eve is TypeScript-first. Python support is not on the immediate roadmap.
• You want agents to be diff-able, reviewable artifacts in your repo.
• You already deploy to Vercel and want the cloud story handled.
• You want a serverless durable runtime without building checkpointing yourself.
• You’re building 5+ agents and want shared evals, observability, and identity governance.

Skip eve if:

• You’re a Python shop. Use LangGraph, CrewAI, Pydantic AI, or Mastra (TypeScript but more code-shaped).
• You need maximum graph flexibility (cyclic state machines, complex branching). LangGraph is stronger here.
• You’re allergic to vendor lock-in to Vercel infra. eve runs anywhere TypeScript runs, but the integrated experience (durable functions, AI Gateway, Passport) ties best to Vercel.
• You want a single Python notebook-style script. eve assumes a directory-shaped project.

How it compares to the field

Framework	Language	Shape	Hosted runtime	Open source
Vercel eve	TypeScript	Directory of files	Vercel Functions	Yes (Apache-2.0)
LangGraph	Python/TS	Code graph	LangGraph Cloud or self-host	Yes (MIT)
CrewAI	Python	Code (roles + tasks)	Self-host	Yes (MIT)
Mastra	TypeScript	Code (workflows + agents)	Self-host or Mastra Cloud	Yes
OpenClaw	Node.js	Config + skills dirs	Self-host (Gateway + nodes)	Yes
AutoGen / OpenAI Agents SDK	Python	Code (agents + handoffs)	Self-host	Yes

Closest sibling: OpenClaw, which also uses a skills/ directory model and treats agents as composable configurations. eve is more opinionated about the whole agent being a filesystem layout and ships more managed infrastructure.

Getting started

npx create-eve-agent my-agent
cd my-agent
eve dev

create-eve-agent scaffolds a directory with instructions.md, a sample tool, and a sample skill. eve dev runs it locally with a chat UI. eve deploy ships it to Vercel.

Repo: github.com/vercel/eve (Apache-2.0). Docs: vercel.com/docs/eve. Launch post: vercel.com/blog/eve.

What this signals

Three larger trends:

1. Filesystem-as-agent is winning. Anthropic’s Claude Skills, OpenClaw’s skills directories, and now Vercel eve all converge on the same idea: the agent is what’s on disk, not what’s instantiated in memory.
2. Vercel is making a real platform play. eve + Connect + Passport + AI Gateway + Sandboxes is a complete agentic infrastructure stack. They’re not just “Next.js for AI” — they want to be the default deploy target for production agents.
3. Identity governance is becoming table stakes. Both Vercel (Passport) and Akamai (Agentic Security Framework, June 15) launched AI agent identity products the same week. Enterprise IT is waking up to “who is this agent and what is it allowed to do.”

Bottom line

If you’re a TypeScript team building production agents, eve is worth a serious look. The filesystem-as-agent model is the most coherent answer anyone has given to “how do we make agents reviewable, shareable, and ops-friendly,” and the built-in infrastructure (durable execution, sandboxes, AI Gateway) eliminates a lot of plumbing.

If you’re Python-first, watch eve to see if the pattern gets ported to a Python equivalent — Mastra and OpenClaw are the closest existing fits.

Sources: Vercel Ship London announcement (June 17, 2026), The New Stack, MarkTechPost, The Register, SiliconANGLE. Last verified: June 22, 2026.