eve from Vercel: an agent framework where the project folder is the agent
Today's Extensions digest covers vercel/eve, a beta, Apache-2.0 framework where instructions, tools, skills, channels and schedules live in plain files. I read the README and metadata and ran nothing.
Facts checked October 15, 2026
vercel/eve is a filesystem-first framework for durable AI agents. An agent is a folder: an instructions file, typed tool functions, on-demand skills, channels and cron schedules. The LICENSE file is Apache-2.0 and the README says it is in beta. I read the README and repository metadata and ran nothing. Repository
What it is
The README calls eve a filesystem-first framework for durable AI agents, with core capabilities in conventional locations so projects are easier to inspect, extend and operate. A project has an agent folder with instructions.md (required, the always-on system prompt), and optional agent.ts, tools, skills, channels (HTTP, Slack, Discord) and schedules.
The README does not mention MCP. Skills here are procedures loaded on demand, and tools are typed functions the model can call. vercel/eve README
Setup, as documented
The quick start is npx eve@latest init my-agent, which creates the folder, installs dependencies, initializes Git and starts a terminal UI. The generated agent uses openai/gpt-6-luna-fast with high reasoning, and the README shows how to pass another AI Gateway model ID. Web Chat needs a linked Vercel project and sign-in with Vercel. I did not run any of this, so these are the vendor's steps.
License and maintenance
The LICENSE file in vercel/eve is the Apache License 2.0, and the npm metadata for eve also says Apache-2.0. The README says eve is in beta, subject to the Vercel beta terms, and that the framework, APIs, documentation and behavior may change before general availability.
On October 15, 2026 the GitHub API showed 5,566 stars, 612 forks, 740 open issues, a creation date of June 16, 2026 and a last push earlier that day. The latest release is eve@0.79.1, published October 14. Stars are a dated signal, not a quality measure. LICENSE npm metadata
Privacy and security limits
Security reports go to responsible.disclosure@vercel.com, not public issues, per the README and SECURITY.md. The README does not describe what data leaves your machine when an agent runs. By default the model is reached through AI Gateway, so prompts go to whichever provider you pick. A tool is code the model can call, so it runs with the permissions you give it. SECURITY.md
Analysis: my inference
Putting the agent in files means the prompt, tools and skills can go through code review and Git history like any other code. That is the part I find useful. The risk moves to the tool files: review each one as you would a script that anyone who can reach the agent can trigger.
Version 0.79 and a beta notice say the interface will move. I would pin a version and try it on a throwaway project first.
What I could not confirm
I did not install or run eve. I read the README, LICENSE, SECURITY.md and npm and GitHub metadata, not the documentation site or the source. I do not know the hosting cost, how permissions work in channels, or whether any telemetry is sent.
Layout from the README, not a tested config
my-agent/
└── agent/
├── agent.ts # optional: model and runtime config
├── instructions.md # required: the always-on system prompt
├── tools/ # optional: typed functions the model can call
├── skills/ # optional: procedures loaded on demand
├── channels/ # optional: HTTP, Slack, Discord
└── schedules/ # optional: recurring cron jobs
# README quick start (not run by me):
npx eve@latest init my-agent
# my review checklist before use:
[ ] pin the eve version (beta, 0.x)
[ ] read every file in tools/ like a script anyone can trigger
[ ] pick the model provider knowing prompts go to it