vercel/eve

What it solves

It simplifies the creation and operation of durable AI agents by using the filesystem as the primary authoring interface. This approach makes agent configurations, prompts, and tools easier to inspect, extend, and manage compared to traditional code-heavy configurations.

How it works

Eve treats the project directory structure as the agent's configuration. Developers define the agent's behavior through a set of organized folders:

• Instructions: A markdown file (instructions.md) serves as the always-on system prompt.
• Tools: Typed functions in the tools/ directory that the model can call.
• Skills: Procedures loaded on demand in the skills/ directory.
• Channels: Integration points for communication (e.g., HTTP, Slack, Discord) in the channels/ directory.
• Schedules: Recurring cron jobs in the schedules/ directory.

Who it’s for

Developers who want to build AI agents with a clear, organized project layout and a filesystem-first approach to agent authoring.

Highlights

• Filesystem-first design: Agent capabilities are stored in conventional locations for better visibility.
• Durable agents: Support for long-running agents with schedules and multiple communication channels.
• Integrated documentation: The eve package includes local documentation for coding agents to read.
• Quick setup: Simple initialization via npx eve@latest init.