Vercel Open-Sources Eve: Every Agent Is Just a Directory
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 claims 100+ internal agents already run on Eve in production. Notable examples:
- d0 — data analyst, processing ~30,000 queries per month
- Lead Agent — sales agent, $5,000/year cost with 32x return
- Vertex — support engineer, autonomously resolving 92% of tickets
agent.ts— defines the modelinstructions.md— the system prompttools/— tools (file name = tool name)skills/— knowledge loaded on demandconnections/— MCP/OpenAPI bridgessandbox/— sandbox overridessubagents/— child agentschannels/— deployment channelsschedules/— timers/cronlib/— shared code- d0 — data analyst
- Lead Agent — sales SDR, maintained half-time by one engineer; $5,000/year cost for $160,000 in return
- Athena — sales dashboard, built by RevOps in 6 weeks without writing code
- Vertex — support engineer, autonomously resolves 92% of tickets
- draft0 — content moderation
- V — router agent that dispatches tasks in Slack
- Vercel ecosystem lock-in —
vercel deployis the first-class path; migrating to AWS or Cloudflare requires rewriting channels. - "Agent as directory" is git-friendly (diffs are clear), but a barrier for non-engineering teams—PMs and ops staff can't easily modify agents.
- The six capabilities are "built-in but opt-in"; configuration complexity may still grow with project size.
- Single-tenant scale (100 agents, 30k queries) vs. multi-tenant performance data hasn't been published.
- The V routing/orchestration design is elegant internally, but whether the open-source version holds up for enterprise-grade scheduling remains to be seen.
Core Idea: Filesystem-First Agents
Every agent is a directory on disk—the directory structure itself is the contract:
The minimal runnable agent is just 2 files (a model and instructions). Adding a tool means adding one TypeScript file—no registration code, no boilerplate; Eve auto-discovers and wires everything at build time.
Six Built-in Production Capabilities
1. Durable execution — every step is checkpointed; agents recover after crashes or deploys (built on an open-source Workflow SDK).
2. Sandboxed compute — agent-generated code is treated as untrusted; shell, scripts, and file I/O run in isolated sandboxes (Vercel Sandbox in production; Docker / microsandbox / bash locally).
3. Human approval — any action can be marked needsApproval; the agent pauses without burning compute and resumes from the original step after approval.
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, and Linear supported, with hand-off between channels.
6. Tracing & evals — every run emits OpenTelemetry traces exportable to Braintrust, Honeycomb, Datadog, or Jaeger; evals run as test suites locally or in CI.
Six Internal Agents Open-Sourced
Deployment
One command: npx eve@latest init my-agent, then vercel deploy. Local code and production are identical, with sandbox backends switching automatically.
Why It Matters
1. "Directory as agent" is a reaction to code bloat in LangChain / AutoGen, which require assembling chains, routers, and tool registries. Eve pushes this back to the filesystem—one file per capability; adding capability means adding a file. 2. 100+ internal agents as proof, with production-scale numbers (d0: 30k queries/month; Vertex: 92% ticket resolution)—not demos. 3. AI Gateway + Vercel Sandbox loop closed: models route via AI Gateway (provider routing + fallback), code execution via Vercel Sandbox (millisecond cold starts), all tied together in one deploy.