The easiest way to run Ralph Wiggum autonomous coding loops — pull specs from GitHub, Linear, Notion, Figma, and more, then let AI build it.

Integrations • Quick Start • Features • Docs

Most AI coding tools work in isolation. You describe a task, AI builds it, done.

ralph-starter brings the Ralph Wiggum technique to production. It connects to your existing workflow — pulling specs from GitHub issues, Linear tickets, Notion docs, or any URL — then runs autonomous Ralph Wiggum loops until the task is complete. One command to go from spec to shipped code.

# Build from a GitHub issue
ralph-starter run --from github --project myorg/myrepo --label "ready"

# Build from a Linear ticket
ralph-starter run --from linear --project "Mobile App" --label "sprint-1"

# Build from a Notion spec
ralph-starter run --from notion --project "https://notion.so/Product-Spec-abc123"

# Or just describe what you want
ralph-starter run "build a todo app with React" --commit

ralph-starter integrates with your favorite tools out of the box:

# Check available integrations
ralph-starter integrations list

# Test connectivity
ralph-starter integrations test github
ralph-starter integrations test linear

# Preview data before running
ralph-starter integrations fetch github owner/repo

Want more integrations? PRs welcome! See CONTRIBUTING.md to get started.

• Integrations
• Quick Start
• Features
• Commands
• Configuration
• Contributing

# Simple task
ralph-starter run "build a todo app" --commit --validate

# With preset
ralph-starter run --preset tdd-red-green "add user authentication"

# With safety controls
ralph-starter run --rate-limit 50 --circuit-breaker-failures 3 "build X"

# Interactive wizard
ralph-starter

The Ralph Wiggum technique is an autonomous AI coding pattern where you run a coding agent (like Claude Code, Codex, or Cursor) in a loop — feeding errors back into the agent until the task is complete. Originally described by Geoffrey Huntley, the Ralph Wiggum loop has become one of the most popular approaches for autonomous AI development.

ralph-starter is the production-grade Ralph Wiggum implementation — it adds integrations, safety controls (circuit breakers, rate limiting, cost tracking), git automation, validation backpressure, and multi-agent support on top of the core Ralph Wiggum loop pattern. Instead of a bare bash script, you get a full orchestration engine.

Why ralph-starter over a raw Ralph Wiggum script?

• Spec-driven: Pull tasks from GitHub, Linear, Notion, Figma — not just manual prompts
• Safe: Circuit breakers stop stuck loops, rate limiters control costs
• Multi-agent: Works with 8+ coding agents, not just one
• Observable: Cost tracking, progress logs, iteration history
• Automated: Auto-commit, push, and PR creation built in

npm install -g ralph-starter
# or
npx ralph-starter

After installing, run the setup wizard and verify your environment:

ralph-starter setup    # Configure API keys and preferences
ralph-starter check    # Verify system requirements and connectivity

Just run ralph-starter with no arguments to launch the interactive wizard:

ralph-starter

The wizard will:

1. Ask if you have a project idea (or help you brainstorm one)
2. Refine your idea with AI
3. Let you customize the tech stack
4. Build your project automatically

ralph-starter ideas

This launches Idea Mode - a brainstorming session to help you discover project ideas:

• Brainstorm with AI - Get creative suggestions
• See trending ideas - Based on 2025-2026 tech trends
• Based on my skills - Personalized to technologies you know
• Solve a problem - Help fix something that frustrates you

# Run a single task
ralph-starter run "build a todo app with React"

# With git automation
ralph-starter run "add user authentication" --commit --pr

# With validation (backpressure)
ralph-starter run "refactor auth" --commit --validate

# Fetch specs from external sources
ralph-starter run --from https://example.com/spec.md
ralph-starter run --from github --project myorg/myrepo --label "ready"
ralph-starter run --from linear --project "Mobile App"

# Fetch a specific GitHub issue
ralph-starter run --from github --project owner/repo --issue 123

# Specify output directory (skips "where to run?" prompt)
ralph-starter run --from github --project owner/repo --issue 42 --output-dir ~/projects/new-app

ralph-starter automatically detects existing projects when you run the wizard:

Ralph Playbook Project (has AGENTS.md, IMPLEMENTATION_PLAN.md, etc.):

cd my-ralph-project
ralph-starter

The wizard will detect the Ralph Playbook files and let you:

• Continue working (run the build loop)
• Regenerate the implementation plan
• Add new specs

Language Project (has package.json, pyproject.toml, Cargo.toml, go.mod):

cd my-existing-app
ralph-starter

The wizard will detect the project type and let you:

• Add features to the existing project
• Create a new project in a subfolder

Launch with ralph-starter (no args) for a guided experience:

• Describe your idea in plain English
• AI refines and suggests features
• Choose your tech stack
• Auto-runs init → plan → build

For users who don't know what to build yet:

ralph-starter ideas

Use ralph-starter from Claude Desktop or any MCP client:

ralph-starter mcp

Add to Claude Desktop config:

{
  "mcpServers": {
    "ralph-starter": {
      "command": "ralph-starter",
      "args": ["mcp"]
    }
  }
}

Available MCP Tools:

• ralph_init - Initialize Ralph Playbook
• ralph_plan - Create implementation plan
• ralph_run - Execute coding loop
• ralph_status - Check progress
• ralph_validate - Run tests/lint/build

Works with your favorite coding agents:

• Claude Code (recommended)
• Cursor
• OpenCode
• OpenAI Codex
• GitHub Copilot
• Gemini CLI
• Amp
• Openclaw

ralph-starter supports multiple LLM providers for internal features:

These keys are for ralph-starter's internal LLM calls. Coding agents handle their own authentication.

ralph-starter run "your task" --commit      # Auto-commit after tasks
ralph-starter run "your task" --push        # Push to remote
ralph-starter run "your task" --pr          # Create PR when done

ralph-starter run "your task" --validate    # Run tests/lint/build after each iteration

The --validate flag runs test, lint, and build commands (from AGENTS.md or package.json) after each iteration. If validation fails, the agent gets feedback to fix the issues.

Pre-configured settings for common development scenarios:

# List all 16+ presets
ralph-starter presets

# Use a preset
ralph-starter run --preset feature "build login"
ralph-starter run --preset tdd-red-green "add tests"
ralph-starter run --preset debug "fix the bug"
ralph-starter run --preset refactor "clean up auth module"
ralph-starter run --preset pr-review "review changes"

Available Presets:

Automatically stops loops that are stuck:

# Stop after 3 consecutive failures (default)
ralph-starter run "build X" --validate

# Custom thresholds
ralph-starter run "build X" --circuit-breaker-failures 2 --circuit-breaker-errors 3

The circuit breaker monitors:

• Consecutive failures: Stops after N validation failures in a row
• Same error count: Stops if the same error repeats N times

Writes iteration logs to activity.md:

# Enabled by default
ralph-starter run "build X"

# Disable if not needed
ralph-starter run "build X" --no-track-progress

Each iteration records:

• Timestamp and duration
• Status (completed, failed, blocked)
• Validation results
• Commit info

The loop automatically checks for completion signals:

• RALPH_COMPLETE file in project root
• .ralph-done marker file
• All tasks marked [x] in IMPLEMENTATION_PLAN.md

Control API call frequency to manage costs:

# Limit to 50 calls per hour
ralph-starter run --rate-limit 50 "build X"

When rate limits are reached, ralph-starter displays detailed stats:

⚠ Claude rate limit reached

Rate Limit Stats:
  • Session usage: 100% (50K / 50K tokens)
  • Requests made: 127 this hour
  • Time until reset: ~47 minutes (resets at 04:30 UTC)

Session Progress:
  • Tasks completed: 3/5
  • Current task: "Add swarm mode CLI flags"
  • Branch: auto/github-54
  • Iterations completed: 12

To resume when limit resets:
  ralph-starter run

Tip: Check your limits at https://claude.ai/settings

This helps you:

• Know exactly when you can resume
• Track progress on your current session
• Understand your usage patterns

Track estimated token usage and costs during loops:

# Cost tracking is enabled by default
ralph-starter run "build X"

# Disable cost tracking
ralph-starter run "build X" --no-track-cost

Cost tracking provides:

• Per-iteration cost displayed during the loop
• Running total of tokens and cost
• Cost summary at the end of the loop
• Cost logged in activity.md for each iteration
• Projected cost for remaining iterations (after 3+ iterations)

Supported models for cost estimation:

• Claude 3 Opus ($15/$75 per 1M tokens)
• Claude 3.5 Sonnet ($3/$15 per 1M tokens)
• Claude 3.5 Haiku ($0.25/$1.25 per 1M tokens)
• GPT-4 ($30/$60 per 1M tokens)
• GPT-4 Turbo ($10/$30 per 1M tokens)

ralph-starter follows the Ralph Playbook methodology:

# 1. Initialize Ralph Playbook files
ralph-starter init

# 2. Write specs in specs/ folder

# 3. Create implementation plan
ralph-starter plan

# 4. Execute the plan
ralph-starter run --commit --validate

This creates:

• AGENTS.md - Agent instructions and validation commands
• PROMPT_plan.md - Planning prompt template
• PROMPT_build.md - Building prompt template
• IMPLEMENTATION_PLAN.md - Prioritized task list
• specs/ - Specification files

Use RALPH_DEBUG=1 to see detailed output during execution:

# See detailed agent output, timing, and prompts
RALPH_DEBUG=1 ralph-starter run "build a todo app"

# Debug with GitHub issue
RALPH_DEBUG=1 ralph-starter run --from github --issue 42

Debug mode shows:

• Exact commands being run
• Agent output in real-time
• Timing information
• Error details

# List all available presets
ralph-starter presets

# Use a preset
ralph-starter run --preset feature "build login page"
ralph-starter run --preset tdd-red-green "add user validation"
ralph-starter run --preset debug "fix the auth bug"

# Stop when agent outputs "FEATURE_DONE"
ralph-starter run --completion-promise "FEATURE_DONE" "build X"

# Require explicit exit signal
ralph-starter run --require-exit-signal "build Y"

# Limit to 50 API calls per hour
ralph-starter run --rate-limit 50 "build X"

# Stop after 2 consecutive failures
ralph-starter run --circuit-breaker-failures 2 "build Y"

# Fix build/lint errors automatically
ralph-starter fix

# Fix a specific design/visual issue
ralph-starter fix "fix the paddings and make the colors brighter"

# Full scan with auto-commit
ralph-starter fix --scan --commit

For design-related tasks (CSS, colors, spacing, etc.), the fix command automatically:

• Detects and applies installed design skills
• Instructs the agent to visually verify changes via browser screenshots

# Set credentials
ralph-starter config set linear.apiKey <key>
ralph-starter config set notion.token <token>
ralph-starter config set github.token <token>

# View config
ralph-starter config list
ralph-starter config get linear.apiKey

# Remove
ralph-starter config delete linear.apiKey

mkdir my-saas && cd my-saas
git init

ralph-starter run "Create a SaaS dashboard with:
- User authentication (email/password)
- Stripe subscription billing
- Dashboard with usage metrics
- Dark mode support" --commit --pr --validate

# Watch the magic happen...
# Loop 1: Setting up Next.js project...
# Validation passed
# Committed: chore: initialize Next.js with TypeScript
# Loop 2: Adding authentication...
# ✓ Validation passed
# ✓ Committed: feat(auth): add NextAuth with email provider
# ...
# ✓ Created PR #1: "Build SaaS dashboard"

You can test ralph-starter with public URLs - no API keys required:

# Test with a public GitHub gist or raw markdown
ralph-starter run --from https://raw.githubusercontent.com/multivmlabs/ralph-starter/main/README.md

# Test with GitHub issues (requires gh CLI login)
gh auth login
ralph-starter run --from github --project multivmlabs/ralph-starter --label "enhancement"

# Launch the interactive wizard
ralph-starter

# Or test idea mode
ralph-starter ideas

# Create a simple spec file
echo "Build a simple counter app with React" > my-spec.md

# Run with local file
ralph-starter run --from ./my-spec.md

Before using an integration, verify it's working:

# Check what integrations are available
ralph-starter integrations list

# Test each integration
ralph-starter integrations test github
ralph-starter integrations test linear
ralph-starter integrations test notion

# Preview items (dry run)
ralph-starter integrations fetch linear "My Project" --limit 3

Set environment variables in your shell profile or .env file:

# Add to ~/.bashrc, ~/.zshrc, or .env file
export LINEAR_API_KEY=lin_api_xxxxx
export NOTION_API_KEY=secret_xxxxx
export GITHUB_TOKEN=ghp_xxxxx

Environment variables take precedence over the config file.

Use the CLI to store credentials:

ralph-starter config set linear.apiKey lin_api_xxxxx
ralph-starter config set notion.token secret_xxxxx
ralph-starter config set github.token ghp_xxxxx

Credentials are stored in ~/.ralph-starter/sources.json.

• Node.js 18+
• At least one coding agent installed (Claude Code, Cursor, etc.)
• Git (for automation features)
• GitHub CLI gh (for PR creation and GitHub source)

Full documentation available at: https://ralphstarter.ai

Contributions welcome! See CONTRIBUTING.md for guidelines.

• Feature requests & ideas: ralph-ideas
• Project templates: ralph-templates

For creating custom integrations, agents, or using the programmatic API, see the Developer Extension Guide.

Using ralph-starter? Add the badge to your README:

[![built with ralph-starter](https://ralphstarter.ai/img/badge-built-with@2x.png)](https://github.com/multivmlabs/ralph-starter)

More styles available at ralphstarter.ai/badge.

MIT