Vercel Releases Eve 0.22, a Filesystem-Based Framework for Building AI Agents

The release, published July 12 under the Apache-2.0 license, is the newest point release in Eve's 0.22 line. Developers scaffold a new project by running npx eve@latest init my-agent, which generates a conventional folder structure: an instructions.md file holding the agent's system prompt, a tools folder for typed, Zod-validated functions the model can call, a skills folder for on-demand procedures, a channels folder for integrations like HTTP endpoints, Slack, and Discord, and a schedules folder for cron-triggered tasks.

Eve supports multiple LLM backends through configuration, so a project isn't locked to a single model provider. The framework also includes built-in human-in-the-loop prompts, letting an agent pause and ask a person to make a decision rather than guessing, and subagent support, allowing one agent to delegate work to another and incorporate the result. A terminal-based development interface lets developers watch an agent execute a task in real time.

The project has drawn 3,500 GitHub stars and 300 forks since launch, with 120 open issues against 132 pull requests. Vercel's documentation for the framework is hosted separately at eve.dev, and generated project scaffolds ship with local documentation under node_modules/eve/docs.

Eve is one of several recent releases betting that removing configuration overhead, rather than adding features, is the more useful direction for agent tooling, alongside IBM's CUGA framework (IBM's open-source CUGA agent harness skips the…) and the broader push toward file-based, inspectable agent instructions seen in projects like Matt Pocock's Claude Code skills repository (Inside the GitHub Repo Teaching AI Agents to Code Like…).