Vercel Eve

Vercel Eve takes an opinionated approach to agent architecture: instead of a class hierarchy or a JSON config, an agent is just a folder. A tool is a single TypeScript file that exports a function. A skill is a Markdown file. You can inspect, version, fork (copy and modify for a new use case), and deploy the whole thing with the same Git workflow you use for any other Vercel project.

The filesystem-first design makes agents easy to reason about without a visual builder or a proprietary SDK. It also fits naturally into the patterns builders already use: code review via pull request works on agent tools the same way it works on application code. Vercel ships integrations for connecting Eve agents to external memory layers (like Hindsight), voice stacks, and MCP servers, so the framework composes with the broader agent tooling ecosystem rather than trying to own the whole stack.

Eve is distinct from the Vercel AI SDK (which is a lower-level library for calling models and streaming responses) and from Vercel's hosted infrastructure (the deployment target). It sits in between as an agent-level abstraction: the SDK handles model calls, Eve handles agent structure, and Vercel deploys the result. For builders already on Vercel who want to ship an agent without adopting a heavy framework like LangGraph or CrewAI, Eve is the path-of-least-resistance option.