@vercel/dream

Write a spec. Get a deployed app.

Dream is a CLI that turns markdown specifications into fully built applications on Vercel. It runs an AI agent in a loop, reading your specs, writing code, and producing a deployable build — no scaffolding, no boilerplate, no manual steps.

pnpm add @vercel/dream

How it works

1. You write a specification in specs/
2. dream reads it, plans the work, and starts building
3. The agent loops — reading, coding, verifying — until every requirement is met
4. Output lands in .vercel/output/ ready to deploy

▲ dream · my-app

  dir         /Users/you/code/my-app
  model       vercel/anthropic/claude-opus-4.5
  timeout     60.0m
  max         100 iterations

  ● OpenCode ready
  ● Provider vercel connected

  [1] Running session...
       ▸ read specs/app.md
       ▸ write PROGRESS.md

       Starting with the HTML structure and core game logic.

       ▸ write .vercel/output/config.json
       ▸ write .vercel/output/static/index.html
       ✎ .vercel/output/static/index.html
       ▸ write .vercel/output/static/styles.css
       ✎ .vercel/output/static/styles.css
       ▸ write .vercel/output/static/app.js
       ✎ .vercel/output/static/app.js

       All tasks complete. Verifying output structure.

       ▸ glob .vercel/output/**/*
       12 tools · 48.2k→3.1k · $0.12
  [1] ✓ Done (34.2s)

  ✓ Completed in 1 iteration(s) (34.2s)

Quick start

# Scaffold a new project
dream init my-app
cd my-app

# Write your spec
cat > specs/app.md << 'EOF'
# Landing Page

A minimal dark-themed landing page with:
- Hero section with animated gradient title
- Feature grid (3 columns)
- Email signup form
EOF

# Build it
pnpm install
dream

Specs

Specs are markdown files in the specs/ directory. Write what you want — the more detail, the better the output. The agent reads every .md file in the directory.

specs/
├── app.md           # Main application spec
├── design.md        # Visual design requirements
└── accessibility.md # A11y requirements

Output

Dream produces Vercel Build Output API v3 — static files in .vercel/output/static/ with a config.json. Deploy to Vercel or serve anywhere.

Commands

| Command | Description | |---------|-------------| | dream | Build the project from specs | | dream init | Scaffold a new dream project | | dream models | List available models and auth status | | dream config | Show project configuration |

Options

| Flag | Description | Default | |------|-------------|---------| | -m, --model | Model in provider/model format | vercel/anthropic/claude-opus-4.5 | | -t, --timeout | Timeout in milliseconds | 3600000 (60m) | | -i, --max-iterations | Maximum agent loops | 100 | | -v, --verbose | Show all events | false | | -d, --dir | Working directory | . |

Authentication

Dream uses the Vercel AI Gateway. Authenticate with either:

OIDC token (Vercel deployments & local dev):

vercel env pull  # writes .env.local with VERCEL_OIDC_TOKEN
source .env.local
dream

API key:

export VERCEL_API_KEY=your_key
dream

Requirements

• OpenCode installed and available on PATH
• Node.js 18+
• A Vercel account with AI Gateway access

License

MIT