Question 1

What is Eve?

Accepted Answer

Eve is Vercel's filesystem-first framework for durable backend agents — described by its own docs as being like Next.js for agents. An agent is one folder: instructions.md is the system prompt, agent.ts is the runtime configuration, and tools, skills, subagents, schedules and connections are each a directory whose files Eve discovers. It runs locally with npm run dev and deploys to Vercel with eve deploy.

Question 2

How do you define an Eve agent?

Accepted Answer

By writing files, not by registering anything. Identity comes from the path: agent/tools/get_weather.ts is the tool the model calls as get_weather, agent/skills/summarize.md is the skill summarize, and agent/subagents/researcher/ is the subagent researcher. There is no manifest listing them and no name field inside the file, which is why a folder generated with the wrong filename is a bug you only notice at runtime.

Question 3

What goes in an Eve tool file?

Accepted Answer

A description written for the model, an inputSchema (Zod, Standard Schema or JSON Schema — z.object({}) when the tool takes no input) and an execute(input, ctx) function. Sensitive calls add approval: always() or once() from eve/tools/approval, which gates the call on a human before it runs. The filename supplies the name, so it is the one field you cannot set from inside the file.

Question 4

Do subagents inherit the root agent's model and tools?

Accepted Answer

No. A declared subagent under agent/subagents/ inherits nothing from the root — it receives only what is authored in its own directory, plus framework defaults. That is why each subagent row here has its own tools and skills: they are emitted to agent/subagents/<id>/tools/ and .../skills/ by the same generators as the root's, and a subagent with neither can do nothing but write prose back to its caller. It also requires a description, and Eve's compiler rejects an agent.ts that leaves it out. And a subagent never sees the parent's conversation: it is called with one message, so every fact it needs has to be restated in that message.

Question 5

Which parts of an Eve agent does this generate?

Accepted Answer

Ten of Eve's authored slots: instructions.md, agent.ts, tools, skills, subagents, schedules, connections, hooks, the sandbox definition with its network policy, a defineState handle in agent/lib/, and evals with their config. Channels are the deliberate exception — eve add channel/slack scaffolds the channel file, installs the dependency and writes its own credential lines, so picking a channel here emits that command into the README rather than a hand-written guess at the file. Extensions, instrumentation and remote agents are out of scope: they are integration-level, not part of defining one agent.

Question 6

Why does the sandbox section only offer one backend?

Accepted Answer

Because eve/sandbox/vercel is the backend import path Eve's docs spell out. The alternative is not a guess at another path but the absence of a backend field, which is exactly how you ask for the framework's own default. The network policy is the part worth deciding either way: an agent running model-authored code with allow-all egress can reach anything its credentials can, and deny-all blocks everything including DNS.

Question 7

Does this generator send my agent definition anywhere?

Accepted Answer

No. The page runs entirely in your browser, needs no account, and never asks for an API key — connection credentials are emitted as environment variable names in a .env.example, so there is no secret on this page to leak. The download is a .zip built client-side.

Question 8

Is this an official Vercel tool?

Accepted Answer

No. It is a free tool built by Nodlume against Eve's public documentation. The framework itself, and the authoritative reference for anything here, lives at eve.dev.

Eve agent generator

The path is the identity

Validated, never auto-fixed

No secrets on the page

Eve agents, answered

Design the app the agent works for