Vercel Open-Sources Eve: Every Agent Is Just a Directory

Forum topic · QianXun · 2026-06-18

Summary

On June 17, 2026, Vercel open-sourced Eve, its in-house agent framework, on GitHub under the Apache-2.0 license and released it as an npm package. Eve's core idea is "filesystem-first": each agent is simply a directory on disk. agent.ts defines the model, instructions.md holds the system prompt, and folders for tools, skills, connections, sandbox, subagents, channels, and schedules map one-to-one to capabilities—adding a tool means adding a file, with no registration code or boilerplate. Eve ships six production-grade features: durable execution with per-step checkpoints, sandboxed code execution (Vercel Sandbox in production, Docker locally), human-in-the-loop approvals, proxied MCP/OpenAPI connections that keep tokens out of model context, multi-channel deployment (Slack, Discord, Teams, Telegram, and more), and OpenTelemetry tracing with eval suites. Vercel reports 100+ internal agents run on Eve, including d0 (30,000 data queries monthly), Lead Agent ($5,000 annual cost for a 32x return), and Vertex (92% autonomous ticket resolution). Deployment is npx eve@latest init my-agent followed by vercel deploy. Noted trade-offs include Vercel ecosystem lock-in and configuration complexity at scale.

On June 17, 2026, Vercel released its in-house agent framework Eve on GitHub under the Apache-2.0 license, published as an npm package. Vercel says 100+ internal agents already run on Eve, including three flagship cases: data analyst d0 (30,000 queries per month), sales agent Lead Agent ($5,000/year cost, 32x return), and support agent Vertex (92% autonomous resolution).

Core Idea: Filesystem-First Agents

Every agent is a directory on disk—the directory structure itself is the contract:

agent.ts — the model
instructions.md — system prompt
tools/ — tools (filename = tool name)
skills/ — on-demand loaded knowledge
connections/ — MCP/OpenAPI bridges
sandbox/ — sandbox overrides
subagents/ — child agents
channels/ — deployment channels
schedules/ — timers/cron
lib/ — shared code

A minimal runnable agent is 2 files (one model, one instruction). Adding a tool means adding a TypeScript file—no registration code, no boilerplate; Eve auto-discovers everything at build time.

Six Built-in Production Capabilities

1. Durable execution: per-step checkpoints; recoverable after crashes or deploys—built on the open-source Workflow SDK. 2. Sandboxed compute: agent-generated code is treated as untrusted; shell/scripts/file I/O run in a sandbox. Vercel Sandbox in production, Docker / microsandbox / bash locally. 3. Human approvals: any action can be marked needsApproval. The agent pauses without burning compute, waits for a human, then resumes from the original step. 4. Secure connections: MCP / OpenAPI auth is proxied by Eve—URLs and tokens never enter model context. Launch support: Slack, GitHub, Snowflake, Salesforce, Notion, Linear. 5. Multi-channel: HTTP API on by default; Slack / Discord / Teams / Telegram / Twilio / GitHub / Linear supported, with cross-channel handoff. 6. Tracing & evals: every run emits OpenTelemetry traces, exportable to Braintrust / Honeycomb / Datadog / Jaeger. Evals run as test suites locally or in CI.

Six Internal Agents Open-Sourced

d0 — data analyst
Lead Agent — sales SDR, maintained half-time by one engineer, $5,000/year cost for $160,000 in returns
Athena — sales dashboard, built in 6 weeks by RevOps without writing code
Vertex — support engineer, autonomously resolves 92% of tickets
draft0 — content moderation
V — routing agent that dispatches tasks in Slack

Deployment is one command: npx eve@latest init my-agent → vercel deploy. Local code and production are identical; the sandbox backend switches automatically.

Why It Matters

1. "Directory-as-agent" is a reaction to the code bloat of LangChain / AutoGen, which require assembling chains, routers, and tool registries. Eve pushes that back into the filesystem—one file per capability; add capability by adding files. 2. Vercel backs the framework with 100+ production agents: d0's 30k monthly queries and Vertex's 92% resolution rate are production workloads, not demos. 3. A closed loop with AI Gateway (provider routing + fallback) and Vercel Sandbox (millisecond cold starts), unified by a single deploy.

Risks and Open Questions

Ecosystem lock-in: vercel deploy is the first-class path; migrating to AWS or Cloudflare may require rewriting channels.
Engineering-oriented: directories are git-friendly, but a barrier for non-engineering teams (PMs, ops).
Config complexity: the six capabilities are "built-in but opt-in," and configuration may still grow with project size.
Scale unknowns: no public data on multi-tenant performance versus the single-tenant figures cited.
Orchestration: V's Slack routing is elegant, but whether the open-source version handles enterprise-grade scheduling remains to be seen.

Source: https://www.marktechpost.com/2026/06/17/vercel-releases-eve