Claude Code Autopilot Kit (Claude + OpenClaw Engineer Workflow Kit)

A portable .claude/ bundle for Claude Code with a staged agent workflow, safety hooks, logging, and a strict verification loop.

It also includes OpenClaw and CrewAI bootstrap scripts for remote control, browser automation, multi-agent routing, and cross-session memory.

This repo is built to support a full local engineer workflow:

• fix
• build
• run local stack (yarn dev, make up, docker compose up)
• test
• confirm
• commit and report

The goal is not code-only output. The goal is a repeatable engineering loop.

Once installed, Claude can auto-route substantive tasks into the autopilot pipeline and use OpenClaw when you enable it.

Quick Install

curl -fsSL https://raw.githubusercontent.com/NorkzYT/claude-code-autopilot/main/install.sh \
  | bash -s -- --repo NorkzYT/claude-code-autopilot --ref main --force --bootstrap-linux

After install, you'll see your ntfy.sh subscription URL — subscribe to get notified when Claude needs your attention.

Then restart Claude Code to load the new configuration.

Note: Run Claude Code as a non-root user. The kit's hooks, logs, and VS Code integration are designed for regular user accounts. If using VS Code Remote (SSH/Tailscale), ensure you attach as the same user that runs Claude.

Install With OpenClaw (Discord/browser/multi-agent)

curl -fsSL https://raw.githubusercontent.com/NorkzYT/claude-code-autopilot/main/install.sh \
  | bash -s -- --repo NorkzYT/claude-code-autopilot --ref main --force --bootstrap-linux --with-openclaw

This Docker/OpenClaw install defaults to /opt/openclaw-home when --dest is omitted, regardless of the directory where you ran the command.

OpenClaw Docker Quickstart

For a new user who wants Docker-only OpenClaw with access to repos under /opt/repos:

1. Install with --with-openclaw
2. Copy the env template:

cp /opt/openclaw-home/.env.example /opt/openclaw-home/.env

3. Edit .env and set:
   • HOST_REPOS_DIR=/opt/repos
   • GIT_AUTHOR_NAME
   • GIT_AUTHOR_EMAIL
   • GIT_COMMITTER_NAME
   • GIT_COMMITTER_EMAIL

Recommended defaults:

• OPENCLAW_MODEL_PRIMARY=anthropic/claude-sonnet-4-6
• OPENCLAW_THINKING_DEFAULT=high

Optional auth envs:

• ANTHROPIC_API_KEY
• OPENAI_API_KEY
• OPENCLAW_ANTHROPIC_SETUP_TOKEN

4. Put the repos you want OpenClaw to access under /opt/repos
5. Start the Docker stack after .env is ready:

cd /opt/openclaw-home
openclaw up

6. Authenticate providers inside the container wrapper:

Anthropic subscription:

claude setup-token
openclaw models auth paste-token --provider anthropic

OpenAI subscription:

openclaw models auth login --provider openai-codex

7. Verify the stack:

openclaw status
openclaw logs

8. Open the browser viewer when manual login or 2FA is needed:

openclaw viewer-url

9. Register mounted repos as agents when needed:

openclaw agents add my-app --workspace /opt/repos/my-app --non-interactive

What this means:

• OpenClaw runs in Docker, not on the host
• the host openclaw command is a wrapper into the container
• the default control directory is /opt/openclaw-home unless you override it with --dest
• /opt/repos is mounted read/write into the gateway container
• ~/.openclaw on the host is bind-mounted into the container and reused for state
• if you ever need a different host state path, set OPENCLAW_HOST_STATE_DIR explicitly

Install With CrewAI (marketing/research/ops crews)

curl -fsSL https://raw.githubusercontent.com/NorkzYT/claude-code-autopilot/main/install.sh \
  | bash -s -- --repo NorkzYT/claude-code-autopilot --ref main --force --bootstrap-linux --with-crewai

Install With OpenClaw + CrewAI Together

curl -fsSL https://raw.githubusercontent.com/NorkzYT/claude-code-autopilot/main/install.sh \
  | bash -s -- --repo NorkzYT/claude-code-autopilot --ref main --force --bootstrap-linux --with-openclaw --with-crewai

Start Here

Pick one path:

1. Claude Code only (local terminal workflow): use the Quick Install command above
2. Claude Code + OpenClaw (Discord, browser, remote control): add --with-openclaw
3. Claude Code + CrewAI (crew-based domain teams): add --with-crewai
4. Claude Code + OpenClaw + CrewAI: use both flags in one command
5. Refresh an existing repo: re-run the install command with --force

What this kit covers:

• Claude Code hooks and guardrails (.claude/hooks/*)
• staged agent workflow (autopilot, triage, fixer, closer)
• OpenClaw gateway, Discord, browser, and agent bootstrap scripts
• CrewAI bootstrap with a marketing crew scaffold under .crewai/
• Codex compatibility layer (AGENTS.md, .agents/skills, .codex/rules)

Then read:

• .claude/README-openclaw.md for OpenClaw scripts and common commands
• .claude/docs/openclaw-integration.md for full setup and troubleshooting
• .claude/docs/openclaw-remote-commands.md for Discord use (slash commands, pairing, allowlists, bindings)
• docs/crewai.md for CrewAI setup, execution, and artifact workflow

Terminal Names & cca Alias

Each Claude Code session gets a random memorable name (e.g., cosmic-penguin, thunder-falcon) so you can easily identify multiple terminals. The name appears in:

• The terminal tab title
• Notification messages (so you know which terminal needs attention)
• A local identity file at .claude/terminal-identity.local.json

Launch Claude with the cca alias (added to your shell during install):

cca

This runs claude --dangerously-skip-permissions with automatic terminal naming. Equivalent to:

.claude/bin/claude-named --dangerously-skip-permissions

Note: After install, open a new shell or run source ~/.bashrc (or ~/.zshrc) to activate the cca alias.

How It Works

After installation, the kit automatically:

1. Injects autopilot — A hook detects substantive prompts and triggers the autopilot agent
2. Guards dangerous operations — Blocks rm -rf, sudo, curl|bash, auto-commits, etc.
3. Protects sensitive files — Blocks edits to .env, secrets, certs, prod configs
4. Auto-formats code — Runs Prettier/Black on edited files when configured
5. Logs everything — Prompts, commands, and responses go to .claude/logs/

If you install with OpenClaw, it also:

6. Adds gateway tooling — Remote access, channel routing, and session management
7. Adds browser tooling — OpenClaw-managed browser for local UI verification and CDP flows
8. Bootstraps repo agents — Generates root OpenClaw core files, .openclaw/ runtime state, skills, and compatibility files for multi-repo work

If you install with CrewAI, it also:

9. Scaffolds a CrewAI workspace — Generates .crewai/ with agents/tasks config and Python entrypoints
10. Creates growth artifacts — Produces go-to-market, experiment backlog, and weekly ops outputs under .crewai/reports/
11. Adds a local runner — .claude/scripts/crewai-local-workflow.sh for CLI-driven automation
12. Supports local proxy mode — .claude/scripts/crewai-cliproxyapi.sh to run Dockerized CLIProxyAPI for subscription-backed routing

Usage Guide

Ask Naturally

For most tasks, simply describe what you want:

Add a logout button to the navbar that clears the session and redirects to /login

The autopilot agent automatically:

• Explores the codebase to understand the structure
• Plans the implementation
• Makes the changes
• Verifies the result
• Reviews for issues

Structured Prompts (For Complex Tasks)

For complex or multi-step tasks, use this structure for best results:

1) GOAL
- Add user authentication with JWT tokens

2) DEFINITION OF DONE
- [ ] Login endpoint returns JWT on valid credentials
- [ ] Protected routes reject requests without valid token
- [ ] Tests pass

3) CONTEXT
- Using Express.js backend in /src/api
- User model already exists at /src/models/user.js

4) DETAILS
- Use bcrypt for password hashing
- Token expiry: 24 hours

Skip Autopilot for Simple Questions

Simple questions bypass autopilot automatically:

What files handle authentication?
How does the routing work?
Explain the database schema

Available Agents

When Autopilot Doesn't Finish

Use the autopilot-fixer subagent.

Original Task:
<<<
[paste your original prompt]
>>>

What's Still Wrong:
<<<
[paste error messages or describe the issue]
>>>

Final Verification Before PR

Use the closer subagent.

Acceptance Criteria:
<<<
- [ ] Feature works as described
- [ ] Tests pass
- [ ] No console errors
>>>

Documentation

The README is the fast path. Detailed guides live in docs/*.md and .claude/docs/*.

Core docs (docs/*.md)

• docs/README.md — documentation index
• docs/install.md — install modes, updates, flags, git hygiene
• docs/workflow.md — session persistence, notifications, guardrails, customization, plan mode
• docs/editor.md — external editor (Ctrl+G) and VS Code remote setup
• docs/troubleshooting.md — common issues and validation commands
• docs/openclaw.md — OpenClaw quick guide and hook model overview
• docs/crewai.md — CrewAI setup and marketing crew workflow
• docs/docker-openclaw-crewai.md — separate compose files for OpenClaw-only or CrewAI-only container stacks
• docs/openclaw-plugin-hooks.md — plugin hooks and wrapper design for local workflow automation
• docs/roadmap.md — roadmap for full engineer workflow enforcement

OpenClaw docs (repo-local references)

• .claude/README-openclaw.md — operator quick reference
• .claude/docs/openclaw-integration.md — full setup and operations guide
• .claude/docs/openclaw-commands.md — CLI and slash command reference
• .claude/docs/openclaw-remote-commands.md — Discord pairing, allowlists, bindings, and channel routing

Quick reminders

• Use slash commands in Discord first: /status, /help, /new
• commands.bash=true is only needed for shell passthrough (!<cmd> / /bash)
• OpenClaw plugin hooks and .claude/hooks/* are separate systems

Core Principles

1. Smallest change that satisfies the task — No drive-by refactors
2. Discovery first — Search and read before deciding
3. Always verify — Run tests/lint/build or provide manual steps
4. One bounded retry — Triage → patch → verify once if it fails
5. No destructive commands — Unless explicitly approved