Eve + Eve Studio

Vercel’s open-source agent framework — and the native Mac app that turns the folder into a control center

← Part of the AI Agents knowledge base  ·  AI  ·  Tech

Why This Page

Vercel introduced Eve as an open-source framework for building, running, and scaling production agents — the “Next.js for agents” idea: stop re-hand-rolling durable sessions, sandboxes, channels, and evals for every project.

My friend Ian Borders (@OpenMotus) built Eve Studio, a macOS desktop app for that filesystem-first model: run, chat, wire tools/skills/channels, and deploy to Vercel without living in the terminal. He recommends checking it out — and this page is our working guide for when and how we might use Eve + Eve Studio.

What Is Eve?

Eve is a filesystem-first TypeScript agent framework from Vercel. An agent is a directory of files — not a giant opaque config object. At a glance the tree tells you what the agent is, what it can do, where it lives, and when it acts on its own.

agent/
  agent.ts              # model + defineAgent
  instructions.md       # who it is (system prompt)
  tools/                # what it can do (e.g. run_sql.ts)
  skills/               # what it knows (markdown)
  subagents/            # who it delegates to
  channels/             # where it lives (Slack, web, …)
  schedules/            # when it acts on its own (cron)
  connections/          # MCP / OpenAPI services

File name + place in the tree are the definition. Eve owns the agent loop (like Next.js owns routing), so you write tools and skills as plain files without re-plumbing registration every time.

Vercel runs its own fleet on Eve (data analyst, SDR, support, content, routing agents, etc.). The public pitch: agents are where the web was before frameworks — Eve is meant to be that shared production shape.

Batteries Included (Why Eve vs DIY)

Durable sessions

Conversations as durable workflows (Workflow SDK). Pause, survive deploys/crashes, resume from checkpoints — hours or days long is normal.

Sandbox per agent

Agent-generated code runs isolated (Vercel Sandbox in prod; Docker / microsandbox / just-bash locally). Untrusted code stays out of your app runtime.

Human-in-the-loop

Any tool can require approval. Agent parks (no compute burn) until a human okays expensive or destructive actions.

Connections

MCP / OpenAPI as files. Auth brokered so the model never sees raw credentials. Vercel Connect for OAuth services (Slack, GitHub, Linear, Notion, etc.).

Multi-channel

Same agent on Slack, Discord, Teams, Telegram, web, GitHub, Linear, cron, webhooks… One channel file per surface; sessions can hand off between channels.

Traces & evals

OpenTelemetry spans per turn/tool call. eve eval for scored suites in CI. Observability on Vercel Agent Runs tab.

Also: tools as typed TypeScript, skills as markdown, subagents as nested directories with their own sandbox, agent can write and run code in the sandbox when no tool exists, schedules as cron files → Vercel Cron Jobs in prod.

How to Use Eve (CLI Path)

Scaffold

npx eve@latest init my-agent
# or: npm install eve@latest in an existing app

Minimal agent

// agent/agent.ts
import { defineAgent } from "eve";
export default defineAgent({
  model: "anthropic/claude-opus-4.8",
});

Plus agent/instructions.md for role/personality, and tools under agent/tools/ with defineTool + Zod schemas.

Dev loop

eve dev
# talk in the TUI; every tool call is a visible checkpointed step

eve eval
# run scored evals locally or against a deployed app

vercel deploy
# same directory in production; sandbox adapter swaps automatically

Add a channel

eve channels add slack
# writes channels/slack.ts — approvals become Slack buttons, etc.

Full docs: eve.dev/docs · source: github.com/vercel/eve

Coding agents can bootstrap Eve from a setup prompt (Vercel published a long “Set up an Eve agent…” instruction in the blog post) — same idea as giving Claude/Codex a skill brief for a new framework.

Eve Studio (Ian Borders) — Operate Without the Terminal

Eve Studio is a free, open-source, native macOS (Apple silicon) control center for Eve agents. Ian’s framing: Eve is filesystem-first; Eve Studio is the UI for that — run, chat, wire tools/skills/channels, deploy to Vercel.

Source tweet: @OpenMotus · Built a desktop app for @evedev_

What Eve Studio is good for

  • Chat with local or deployed agents — stream turns, tool calls, subagent handoffs, approvals in one window
  • Edit capabilities in-app — tools, skills, subagents, hooks, instructions, model, connections, channels, schedules
  • One-click Vercel deploy — link project, pull AI Gateway token, env/secrets/logs, team picker
  • Terminal-free onboarding — provisions Node, installs Eve, Vercel sign-in (non-technical operators never open a shell)
  • Memory wiring — durable Eve sessions or external brain connections
  • Integrations — MCP/OpenAPI connections; channels via Vercel Connect (Slack, Linear, GitHub, …); schedules and evals as tabs

First-time path (Studio)

  1. Download the signed/notarized Mac app from releases
  2. Create an agent — Studio scaffolds Eve + Node for you
  3. Connect Vercel — AI Gateway token pulled for you
  4. Chat, edit capabilities, ship to production from the same app

When to use Studio vs CLI: Studio for operators, product people, and day-to-day chat/deploy. CLI for CI, monorepo automation, and deep scriptable workflows. Same Eve agent directory either way.

Phone vs Mac: How Do You Access Eve?

Short version: Eve Studio is Mac-only. Eve agents (the thing you build and deploy) can reach you on your phone through normal chat apps and web — not through a mobile Eve Studio app.

Piece Where it runs Phone access?
Eve Studio Native macOS / Apple silicon desktop app No — no iOS/Android Studio app today
Eve (framework + agent) Local Mac + Vercel when deployed Yes — via channels (Slack, Telegram, web chat, etc.)
Eve CLI (eve dev, deploy) Any machine with terminal + Node Not practical on phone — use Mac/laptop

What you do on the Mac

  • Install Eve Studio or use the CLI
  • Create the agent folder, edit tools/skills/instructions
  • Connect Vercel, set secrets, wire MCP connections
  • Add channels, run evals, deploy to production
  • Full “control center” lifecycle (chat with local or production agent in Studio)

What you do on the phone

Once the agent is deployed and a channel is connected, you talk to that same agent from your phone in apps you already use. Eve is multi-channel by design — one agent, many surfaces.

  • Slack / Discord / Microsoft Teams / Google Chat — open the phone app, message the bot
  • Telegram / WhatsApp / Twilio — very phone-native messaging
  • Web chat — browser on the phone (if you enable the web channel)
  • GitHub / Linear / etc. — more for tickets and queues than casual chat
  • API / webhooks / cron — agent works without you chatting at all

Sessions can move across surfaces depending on setup (e.g. start something in Slack on your phone, continue later on web or desktop). Approvals can show as native channel UI (e.g. Slack buttons) — useful when you need to green-light an action from your phone.

Recommended path for “I want it on my phone”

  1. Build and deploy on Mac — Eve Studio (easiest) or CLI
  2. Add a phone-friendly channel — Slack or Telegram are the usual first picks
  3. Message the agent from the phone app — same agent, production deployment
  4. Keep Studio/CLI for edits, new tools, secrets, and redeploys when you are back at a Mac

What you cannot do on the phone (today)

  • Open Eve Studio and edit tools, skills, or schedules in the desktop control center
  • One-click Vercel deploy from a native Eve mobile app
  • Full onboarding / Node provisioning that Studio does for non-terminal users

Think of it like this: Mac = cockpit (build, wire, ship). Phone = radio (talk to the deployed agent while you are out).

How We Might Use It (Ideas for This Stack)

CBS / GTM operator agent

Eve agent with Onboard + Deepline connections, scheduled weekly lead digests, Slack channel for the team. Aligns with CBS GTM plan.

Support / ops agents

Same pattern as Vercel’s Vertex/d0: tools + skills + Slack channel + approvals on destructive actions.

Multi-agent fleet on one shape

Monorepo of agents (routing parent + specialists) — matches agentic setup and company-brain style fleets.

Human-safe prod tools

needsApproval on DB writes / prod mutations — same spirit as Ryan’s OnePassword one-time secrets on Devon workflows.

Non-dev operators

Eve Studio so someone can chat and ship without learning eve dev first — good for content/ops teammates.

Evals as deploy gate

Prompt/tool changes as PRs with eve eval in CI — pairs with verification loops.

Quick Start Checklist

Related Guides

ChatGPT + Codex Super App

Codex computer use and building-block stacks (Convex, sandboxes) that pair with agent frameworks.

Read part 2 →

Agentic Setup Checklist

Harness maturity: evals, worksheets, night shift, multi-model review.

Open the checklist →

Claude Skills

Skill packaging patterns — Eve skills are the same idea as markdown in a skills/ folder.

Read the deep dive →

Loops & Verification

Autonomous loops and checks — map cleanly to schedules + eve eval.

Read the guide →

OpenClaw

Another agent runtime path in our stack (e.g. Lisa for CBS outreach).

Read the deep dive →

CBS 2026 GTM Plan

Concrete agent use case: Onboard + Deepline + outreach operator.

Review the plan →

Eve framework by Vercel — announcement · eve.dev. Eve Studio by Ian Bordersevestudio.dev · tweet.

Comments

Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login

View comments archive