Ralphy

Join our Discord - Questions? Want to contribute? Join the community!

Autonomous AI coding loop. Runs AI agents on tasks until done.

Install

Option A: npm (recommended)

npm install -g ralphy-cli

# Then use anywhere
ralphy "add login button"
ralphy --prd PRD.md

Option B: Clone

git clone https://github.com/michaelshimeles/ralphy.git
cd ralphy && chmod +x ralphy.sh

./ralphy.sh "add login button"
./ralphy.sh --prd PRD.md

Both versions have identical features. Examples below use ralphy (npm) - substitute ./ralphy.sh if using the bash script.

Two Modes

Single task - just tell it what to do:

ralphy "add dark mode"
ralphy "fix the auth bug"

Task list - work through a PRD:

ralphy              # uses PRD.md
ralphy --prd tasks.md

Project Config

Optional. Stores rules the AI must follow.

ralphy --init              # auto-detects project settings
ralphy --config            # view config
ralphy --add-rule "use TypeScript strict mode"

Creates .ralphy/config.yaml:

project:
  name: "my-app"
  language: "TypeScript"
  framework: "Next.js"

commands:
  test: "npm test"
  lint: "npm run lint"
  build: "npm run build"

rules:
  - "use server actions not API routes"
  - "follow error pattern in src/utils/errors.ts"

boundaries:
  never_touch:
    - "src/legacy/**"
    - "*.lock"

Rules apply to all tasks (single or PRD).

AI Engines

ralphy              # Claude Code (default)
ralphy --opencode   # OpenCode
ralphy --cursor     # Cursor
ralphy --codex      # Codex
ralphy --qwen       # Qwen-Code
ralphy --droid      # Factory Droid
ralphy --copilot    # GitHub Copilot
ralphy --gemini     # Gemini CLI

Model Override

Override the default model for any engine:

ralphy --model sonnet "add feature"                    # use sonnet with Claude
ralphy --sonnet "add feature"                          # shortcut for above
ralphy --opencode --model opencode/glm-4.7-free "task" # custom OpenCode model
ralphy --qwen --model qwen-max "build api"             # custom Qwen model

Engine-Specific Arguments

Pass additional arguments to the underlying engine CLI using -- separator:

# Pass copilot-specific arguments
ralphy --copilot --model "claude-opus-4.5" --prd PRD.md -- --allow-all-tools --allow-all-urls --stream on

# Pass claude-specific arguments  
ralphy --claude "add feature" -- --no-permissions-prompt

# Works with any engine
ralphy --cursor "fix bug" -- --custom-arg value

Everything after -- is passed directly to the engine CLI without interpretation.

Task Sources

Markdown file (default):

ralphy --prd PRD.md

## Tasks
- [ ] create auth
- [ ] add dashboard
- [x] done task (skipped)

Markdown folder (for large projects):

ralphy --prd ./prd/

When pointing to a folder, Ralphy reads all .md files and aggregates tasks:

prd/
  backend.md      # - [ ] create user API
  frontend.md     # - [ ] add login page
  infra.md        # - [ ] setup CI/CD

Tasks are tracked per-file so completion updates the correct file.

YAML:

ralphy --yaml tasks.yaml

tasks:
  - title: create auth
    completed: false
  - title: add dashboard
    completed: false

JSON:

ralphy --json PRD.json

{
  "tasks": [
    {
      "title": "create auth",
      "completed": false,
      "parallel_group": 1,
      "description": "Optional details"
    }
  ]
}

Titles must be unique.

GitHub Issues:

ralphy --github owner/repo
ralphy --github owner/repo --github-label "ready"

Parallel Execution

ralphy --parallel                  # 3 agents default
ralphy --parallel --max-parallel 5 # 5 agents

Each agent gets isolated worktree + branch:

Agent 1 → /tmp/xxx/agent-1 → ralphy/agent-1-create-auth
Agent 2 → /tmp/xxx/agent-2 → ralphy/agent-2-add-dashboard
Agent 3 → /tmp/xxx/agent-3 → ralphy/agent-3-build-api

Without --create-pr: auto-merges back to base branch, AI resolves conflicts. With --create-pr: keeps branches, creates PRs. With --no-merge: keeps branches without merging or creating PRs.

YAML parallel groups - control execution order:

tasks:
  - title: Create User model
    parallel_group: 1
  - title: Create Post model
    parallel_group: 1  # same group = runs together
  - title: Add relationships
    parallel_group: 2  # runs after group 1

Branch Workflow

ralphy --branch-per-task                # branch per task
ralphy --branch-per-task --create-pr    # + create PRs
ralphy --branch-per-task --draft-pr     # + draft PRs
ralphy --base-branch main               # branch from main

Branch naming: ralphy/<task-slug>

Browser Automation

Ralphy can use agent-browser to automate browser interactions during tasks.

ralphy "test the login flow" --browser    # force enable
ralphy "add checkout" --no-browser        # force disable
ralphy "build feature"                    # auto-detect (default)

When enabled, the AI gets browser commands:

• agent-browser open <url> - navigate to URL
• agent-browser snapshot - get element refs (@e1, @e2)
• agent-browser click @e1 - click element
• agent-browser type @e1 "text" - type into input
• agent-browser screenshot <file> - capture screenshot

Use cases:

• Testing UI after implementing features
• Verifying deployments
• Form filling and workflow testing

Config (.ralphy/config.yaml):

capabilities:
  browser: "auto"  # "auto", "true", or "false"

Webhook Notifications

Get notified when sessions complete via Discord, Slack, or custom webhooks.

Config (.ralphy/config.yaml):

notifications:
  discord_webhook: "https://discord.com/api/webhooks/..."
  slack_webhook: "https://hooks.slack.com/services/..."
  custom_webhook: "https://your-api.com/webhook"

Notifications include task completion counts and status (completed/failed).

Sandbox Mode

For large repos with big dependency directories, sandbox mode is faster than git worktrees:

ralphy --parallel --sandbox

How it works:

• Symlinks read-only dependencies (node_modules, .git, vendor, .venv, .pnpm-store, .yarn, .cache)
• Copies source files that agents might modify (src/, app/, lib/, config files, etc.)

Why use it:

• Avoids duplicating gigabytes of node_modules across worktrees
• Much faster sandbox creation for large monorepos
• Changes sync back to original directory after each task

When to use worktrees instead (default):

• Need full git history access in each sandbox
• Running git commands that require a real repo
• Smaller repos where worktree overhead is minimal

Parallel execution reliability:

• If worktree operations fail (e.g., nested worktree repos), ralphy falls back to sandbox mode automatically
• Retryable rate-limit or quota errors are detected and deferred for later retry
• Local changes are stashed before the merge phase and restored after
• Agents should not modify PRD files, .ralphy/progress.txt, .ralphy-worktrees, or .ralphy-sandboxes

Options

Flag	What it does
--prd PATH	task file or folder (auto-detected, default: PRD.md)
--yaml FILE	YAML task file
--json FILE	JSON task file
--github REPO	use GitHub issues
--github-label TAG	filter issues by label
--sync-issue N	sync PRD progress to GitHub issue #N
--model NAME	override model for any engine
--sonnet	shortcut for --claude --model sonnet
--parallel	run parallel
--max-parallel N	max agents (default: 3)
--sandbox	use lightweight sandboxes instead of git worktrees
--no-merge	skip auto-merge in parallel mode
--branch-per-task	branch per task
--base-branch NAME	base branch
--create-pr	create PRs
--draft-pr	draft PRs
--no-tests	skip tests
--no-lint	skip lint
--fast	skip tests + lint
--no-commit	don't auto-commit
--max-iterations N	stop after N tasks
--max-retries N	retries per task (default: 3)
--retry-delay N	seconds between retries
--dry-run	preview only
--browser	enable browser automation
--no-browser	disable browser automation
-v, --verbose	debug output
--init	setup .ralphy/ config
--config	show config
--add-rule "rule"	add rule to config

Requirements

Required:

• AI CLI: Claude Code, OpenCode, Cursor, Codex, Qwen-Code, Factory Droid, GitHub Copilot, or Gemini CLI

npm version (ralphy-cli):

• Node.js 18+ or Bun

Bash version (ralphy.sh):

• jq
• yq (optional, for YAML tasks)
• bc (optional, for cost calc)

Both versions:

• gh (optional, for GitHub issues / --create-pr)
• agent-browser (optional, for --browser)

Engine Details

Engine	CLI	Permissions	Output
Claude	claude	--dangerously-skip-permissions	tokens + cost
OpenCode	opencode	full-auto	tokens + cost
Codex	codex	N/A	tokens
Cursor	agent	--force	duration
Qwen	qwen	--approval-mode yolo	tokens
Droid	droid exec	--auto medium	duration
Copilot	copilot	--yolo	tokens
Gemini	gemini	--yolo	tokens + cost

When an engine exits non-zero, ralphy includes the last lines of CLI output in the error message to make debugging easier.

---

Changelog

v4.7.2

• Improved auth error detection: simplified extractAuthenticationError function with better edge case handling (e.g., JSON dumps during login)
• Added project standards: CLAUDE.md, .cursorrules, CONTRIBUTING.md for consistent AI-assisted development
• Enhanced default prompts: enforce concise, focused code changes

v4.7.1

• Copilot engine improvements: non-interactive mode (--yolo), proper error detection for auth/rate-limit/network errors, token usage parsing, temp file-based prompts for markdown preservation
• Fixed infinite retry loop: tasks now properly abort on fatal configuration/authentication errors
• Project standards: added .editorconfig and .gitattributes for consistent coding styles

v4.7.0

• JSON PRD support: new --json flag to use JSON files as task sources with support for parallel groups and task descriptions

v4.6.0

• Gemini CLI support: new --gemini engine option for Google Gemini CLI
• GitHub issue sync: --sync-issue <number> syncs PRD progress to a GitHub issue after each task
• performance improvements: reduced redundant file reads, exponential backoff for retries, non-blocking logging, operation timing visibility
• version fix: CLI version now reads dynamically from package.json

v4.5.3

• parallel reliability: fallback to sandbox mode on worktree errors
• error output: include CLI output snippet for failed engine commands
• retry handling: detect rate-limit/quota errors and stop early
• merge safety: stash local changes before merge phase and restore after
• prompts: explicitly avoid PRD and .ralphy progress/sandbox/worktree edits

v4.5.0

• sandbox mode: lightweight isolation using symlinks for dependencies (faster than worktrees)
• performance improvements: task caching, parallel merge analysis, smart branch ordering
• webhook notifications: Discord, Slack, and custom webhooks for session completion (configure in .ralphy/config.yaml)
• engine-specific arguments: pass arguments to underlying CLI via -- separator
• Windows improvements: better error handling for .cmd wrappers

v4.4.1

• Windows line ending handling fixes
• Windows Bun command resolution fixes

v4.4.0

• GitHub Copilot CLI support (--copilot)

v4.3.0

• model override: --model <name> flag to override model for any engine
• --sonnet shortcut for --claude --model sonnet
• --no-merge flag to skip auto-merge in parallel mode
• AI-assisted merge conflict resolution during parallel auto-merge
• root user detection: error for Claude/Cursor, warning for other engines
• improved OpenCode error handling and model override support

v4.2.0

• browser automation: --browser / --no-browser with agent-browser
• auto-detects agent-browser when available
• config option: capabilities.browser in .ralphy/config.yaml

v4.1.0

• TypeScript CLI: npm install -g ralphy-cli
• cross-platform binaries (macOS, Linux, Windows)
• no dependencies on jq/yq/bc for npm version

v4.0.0

• single-task mode: ralphy "task" without PRD
• project config: --init creates .ralphy/ with rules + auto-detection
• new: --config, --add-rule, --no-commit

v3.3.0

• Factory Droid support (--droid)

v3.2.0

• Qwen-Code support (--qwen)

v3.1.0

• Cursor support (--cursor)
• better task verification

v3.0.0

• parallel execution with worktrees
• branch-per-task + auto-PR
• YAML + GitHub Issues sources
• parallel groups

v2.0.0

• OpenCode support
• retry logic
• --max-iterations, --dry-run

v1.0.0

• initial release

Contributing

See CONTRIBUTING.md for development guidelines.

Key principles:

• Keep changes small and focused - one logical change per commit
• Break large tasks into micro-tasks
• Quality over speed
• Don't leave dead code
• Fight entropy - leave the codebase better than you found it

AI coding assistants can reference:

• CLAUDE.md - Claude Code instructions
• .cursorrules - Cursor IDE rules

Community

• Discord

License

MIT