Vercel Introduces eve: A Filesystem-First Framework for AI Agents
Vercel has launched eve, a beta open-source framework that applies a Next.js-like filesystem architecture to build, test, and deploy durable AI agents.
Vercel's new open-source agent framework, eve, uses a directory-first approach.
- Vercel's "eve" is a new open-source agent framework using a directory-based configuration, similar to Next.js.
- Agents require only an "instructions.md" file to run, with optional directories for tools, skills, and delivery channels.
- It features durable execution natively, automatically parking and checkpointing agent states during human-in-the-loop delays or schedule waits.
- The framework can run on Vercel's managed infrastructure or be entirely self-hosted using open-source alternatives like PostgreSQL and Docker.
Vercel has introduced eve, an open-source framework designed for building artificial intelligence agents using a filesystem-first architecture. Drawing structural inspiration from Next.js, eve organizes agent logic, system prompts, and tool configurations into a standardized directory layout. By defaulting to file-based configuration, the framework is designed to reduce boilerplate code and standardizes how developers assemble durable, stateful AI agents. Currently in beta, eve can integrate with existing Next.js web applications or operate independently, offering deployments through Vercel's managed infrastructure or as fully self-hosted, independent runtimes.
Filesystem-First Architecture
At the core of the framework is the principle that an agent is defined entirely by its directory structure. The only strict requirement for initializing an agent is a single file named instructions.md. This document functions as the agent's persistent system prompt, establishing its core identity, goals, and operational boundaries. Beyond this foundational file, developers can expand functionality by creating specific subdirectories.
For custom tool integration, adding a TypeScript file into the tools/ directory automatically registers a function that the underlying language model can call. The system maps the file's name directly to the tool's invocation name, bypassing the need for manual registration registries. Procedural knowledge is managed similarly in a skills/ directory, which holds Markdown files detailing specific playbooks. The agent dynamically loads these skills only when they become relevant to the user's request, a mechanism intended to prevent bloat in the model's context window.
Configuration of the language model itself takes place within an optional agent.ts file, where developers can specify an underlying provider such as OpenAI or Anthropic. By default, Vercel routes these API calls through its AI Gateway.
Durability and State Management
A primary technical focus of eve is execution durability. AI agent workflows frequently involve long-running operations or required pauses—such as waiting for human-in-the-loop approvals before executing sensitive tools. Instead of requiring developers to manually build state persistence layers or manage message queues, eve integrates durable execution natively.
By leveraging checkpointing systems, eve ensures that an agent's state is preserved after every operational step. In a managed context, this is handled by Vercel Workflows. In a self-hosted configuration, it defaults to PostgreSQL. If an agent must wait for a user response, its underlying process parks securely to free up compute resources, automatically resuming from its saved state upon message delivery. This human-in-the-loop feature allows administrators to easily create approval gates for actions like database modifications or large API requests.
Sandboxing and Delivery Channels
Safe code execution is enforced through secure environments established within the sandbox/ directory. Here, agents operate within isolated virtual machines capable of executing bash commands and manipulating files safely away from the host system. While Vercel provides isolated VMs natively, developers opting for a self-hosted implementation can direct eve to utilize isolated Docker containers instead.
Connectivity to external APIs is structured within channels/ and connections/ directories. To integrate third-party data seamlessly without storing raw access tokens in tool logic, eve natively supports the Model Context Protocol (MCP). By defining an MCP client connection, standard data sources like GitHub or Linear are securely exposed to the agent. Meanwhile, the channels module enables a single agent codebase to simultaneously operate across multiple delivery platforms—such as Slack, Microsoft Teams, web chat interfaces, and automated scheduled routines via cron syntax in the schedules/ directory.
Native Integration and Independence
For engineering teams working in the React ecosystem, eve features native Next.js compatibility. A Next.js configuration can be wrapped with withEve(), allowing the agent to mount directly into the existing application bundle. Frontend integration is handled via the useEveAgent() React hook, which automatically detects same-origin routes without requiring custom CORS configuration or manual environment variable synchronization between backend and frontend servers.
Crucially, despite being developed by Vercel and defaulting to its proprietary systems—like AI Gateway and Vercel Sandbox—the framework is built strictly on open-source SDKs. Developers retain the capability to deploy agents anywhere. By migrating the persistent state backend to PostgreSQL and sandbox operations to standard Docker environments, teams can run enterprise-grade agents without relying on external managed-infrastructure components.
Enjoyed this?
Get more posts like this delivered to your inbox.
