What the Vercel eve framework is and why directories matter
The Vercel eve framework is an open source system for building AI agents that treats each agent as a directory of files, turning prompts, tools and workflows into a structured project rather than a single opaque configuration. Instead of a monolithic agent definition, eve organizes behavior into TypeScript and Markdown files that sit together in one folder, with infrastructure for running and deploying agents bundled in. Vercel describes eve as “Next.js for agents,” because the directory defines not only what the agent does, but also how it talks to models, tools and users. This focus on clear file layout aims to make AI agent architecture more understandable and easier to maintain, especially for web developers already used to framework conventions. Eve’s open source Apache 2.0 license also positions it as part of a broader ecosystem of open source agent tools and frameworks.
Inside eve’s directory-first AI agent architecture
Eve’s core idea is that an agent is a directory that the framework compiles into a running workflow. One file selects the LLM model, typically via Vercel’s AI Gateway for provider fallback, while another Markdown file holds the system prompt. Tools live as individual TypeScript files whose filenames become tool names, removing the need for separate registration. Agents can connect to external capabilities through MCP servers, skill.md files and OpenAPI documents, and they expose channels for Slack, Discord, Microsoft Teams, Telegram, Twilio, GitHub and Linear. Every conversation runs as a durable workflow using Vercel’s Workflow SDK, with checkpoints so sessions can pause, survive crashes and resume. Each agent also receives its own sandboxed VM for the code it writes, and tools can require human approval before execution. This layout turns AI agent architecture into something closer to a traditional codebase than a black-box assistant.
Developer experience: from local testing to vercel deploy
Eve is designed to feel familiar to developers who ship web apps on Vercel. An agent can run locally using a simple npx eve dev command and be exercised through a built-in testing tool that evaluates responses. When it is time to ship, developers use the same vercel deploy command as for Next.js apps, and sessions in progress finish on the version they started on thanks to immutable deployments. According to Vercel, agents now trigger around 29 percent of deployments on its platform, up from less than 3 percent a year earlier. Each agent run produces an OpenTelemetry trace that appears in a new Agent Runs view in Vercel’s observability dashboard and can be exported to tools like Datadog or Honeycomb. Providers for LLMs and sandboxes are configurable, and Vercel says it is committed to making eve work beyond its own hosting platform.
Passport and the shadow AI governance problem
While eve focuses on AI agent architecture, Vercel’s new Passport feature aims to address shadow AI by bringing identity and access control to agents and apps. Passport uses OpenID Connect to place all team applications and AI agents behind an identity provider such as Okta or Microsoft Entra. This tackles a common pattern where employees build AI-assisted applications that default to Next.js and Vercel hosting, but sit outside formal IT policy. Vercel positions this as a response to enterprise demand for better oversight of “vibe-coded” tools that still touch company data. Passport sits alongside Vercel Connect for short-lived tokens, Enterprise Managed Users for directory sync, and a Bring Your Own Cloud option that lets organizations run Vercel’s platform on their own AWS accounts. Together with eve, these features make open source agent tools more compatible with enterprise governance, rather than existing as unmanaged side projects.
Modular agents, compute efficiency and agent framework comparison
Treating agents as directories pushes a modular, composable view of AI systems, where prompts, tools and channels are concrete files that teams can swap, version and reuse. Eve’s durable workflows and sandbox isolation make it easier to chain agents together, create subagents and connect to external services without building bespoke orchestration each time. Vercel claims that its efficient use of AWS resources mitigates the typical premium associated with running on a platform layer instead of directly on a cloud provider, which may help reduce compute costs for low-scale agents compared to hand-rolled infrastructure. Internally, Vercel says it runs more than 100 agents on eve, including a routing agent and a heavily used Slack data-analysis agent, showing one way this architecture scales. In the broader agent framework comparison landscape, eve joins tools like Mastra and LangGraph but stands out through its TypeScript-first, directory-based design and tight integration with existing web workflows.
