Vercel launches eve, an open-source framework for building agents
By Steven Van ·
The framework bundles durable execution, sandboxed compute, human approvals, and multi-channel deployment into one agent directory structure.
Vercel has released eve, an open-source framework for building, running, and scaling agents in production. Instead of assembling the plumbing an agent needs by hand, eve ships it built in: durable execution, sandboxed compute, human-in-the-loop approvals, subagents, and evals.
An eve agent is defined as a directory of files: agent.ts sets the model (with provider fallback through AI Gateway), instructions.md serves as the system prompt, and separate files define its tools, skills, subagents, channels, and schedules. Each conversation runs as a durable, checkpointed workflow that can pause for hours or days and resume after a crash or deploy, built on the open-source Workflow SDK. Agent-generated code runs in an isolated sandbox rather than the application runtime, backed by Vercel Sandbox when deployed or by Docker, microsandbox, or just-bash locally. Any action can be configured to require human approval before the agent continues.
Connections to tools and data are defined as files pointing at an MCP server or an OpenAPI-compatible API, with Vercel Connect handling OAuth. At launch, eve agents can connect to Slack, GitHub, Snowflake, Salesforce, Notion, and Linear, among other services reachable by API key, OAuth, or MCP. The same agent can run across multiple channels (HTTP API, Slack, Discord, Teams, Telegram, Twilio, GitHub, Linear, or a custom channel), and one channel can hand off to another. Every run produces an OpenTelemetry trace showing each model and tool call in order, exportable to services like Braintrust, Arize, Honeycomb, Datadog, or Jaeger, and surfaced in an Agent Runs tab under Observability on Vercel.
Read the full announcement for setup details and code examples.
