Stop babysitting your AI. Start shipping while you sleep.
After 4 days fighting OpenClaw's complexity, I built a working alternative in 4 hours using Claude Code's native hooks, Discord webhooks, and a simple communication protocol. This system runs in production, coordinates multiple autonomous coding sessions, sends real-time mobile notifications, and has delivered consultant-grade engineering reports while I sleep.
AI coding tools promise autonomy but deliver constant interruptions. You end up:
- Approving every file change
- Stuck at your computer waiting for decisions
- Managing complex frameworks that fail silently
- Switching between a dozen tools that don't talk to each other
OpenClaw tried to solve this. After 4 failed installs over 4 days, I learned that sometimes the best solution is the simplest one.
Three components. Twenty minutes. Actually works.
┌─────────────────────────────────────────────────────────────┐
│ AUTONOMOUS WORKFLOW │
├─────────────────────────────────────────────────────────────┤
│ │
│ 1. Start Claude Code with task │
│ 2. Claude works autonomously │
│ 3. Claude stops at decision points │
│ 4. Hook fires → Discord notification sent │
│ 5. Phone buzzes with STATUS update 📱 │
│ 6. You respond from anywhere │
│ 7. Claude continues │
│ 8. Repeat until complete │
│ │
└─────────────────────────────────────────────────────────────┘
What makes this different:
- Actually autonomous - No approval prompts, Claude decides implementation details
- Mobile-first - Full context in phone notifications, respond from anywhere
- Production-tested - Delivered 27,000-line code analysis overnight
- Simple - Bash script, webhooks, protocol document. No complex dependencies.
- Multi-agent - Coordinate parallel development from your phone
Last night while I slept:
Opus 1 analyzed the video editor codebase:
- 27,000+ lines of code reviewed
- 15 technical issues found (severity-rated, with line numbers)
- 15 feature proposals (value/complexity matrix)
- 6-month implementation roadmap
- Delivered at 5:42 AM with mobile notification
Opus 2 analyzed the DAW system in parallel:
- 47 audio modules identified
- 12 DSP bottlenecks documented
- Performance optimization roadmap
- Timeline estimate: 10 weeks to production
Both sessions coordinated via phone notifications. Zero laptop time required.
This is the kind of report you'd pay a consultant $5,000 for. My AI did it autonomously while I slept.
Prerequisites:
- Claude Code installed
- Discord account
- 20 minutes
Setup:
# 1. Clone this repo
git clone https://github.com/AetherWave-Studio/autonomous-claude-code.git
cd autonomous-claude-code
# 2. Create Discord webhook
# Discord → Server Settings → Integrations → Create Webhook
# Copy the webhook URL
# 3. Install hook script
cp scripts/discord-notify.sh ~/.claude/
chmod +x ~/.claude/discord-notify.sh
# Edit and add your webhook URL
nano ~/.claude/discord-notify.sh
# Find: WEBHOOK_URL="YOUR_WEBHOOK_URL_HERE"
# Replace with your actual webhook URL
# 4. Configure Claude Code hooks
cp templates/settings.json ~/.claude/
# Edit ~/.claude/settings.json — replace YOUR_WSL_USER with your username
# 5. Set up protocol
cp templates/CLAUDE.md ~/.claude/
# 6. Test it
claude code
# Give it a simple task, watch for Discord notification when it stopsWindows users: See docs/SETUP.md for WSL-specific instructions.
That's it. You're autonomous.
⚠️ Security Warning: Never commit your actual Discord webhook URL to version control. The template scripts containYOUR_WEBHOOK_URL_HEREplaceholders. Your local copy with the real URL lives in~/.claude/and should never be committed to git.
- Real-time mobile notifications - Discord app on your phone buzzes when Claude needs input
- Structured status updates - Every notification includes what was done, what works, what's next
- Multi-agent coordination - Run multiple Claude sessions on different features simultaneously
- Session tracking - Unique IDs keep parallel work organized
- Non-blocking - Notifications don't interrupt Claude's work
- Cross-platform - Works on macOS, Linux, Windows (WSL)
Every notification follows this format:
**STATUS: COMPLETED**
- What was done: Authentication middleware implemented with JWT tokens
- Current state: Middleware working, 8 tests passing, integrated with login endpoint
- Next step: Ready for code review and security audit
- Session: auth-middleware-001
- Modified: auth.ts, middleware.ts, auth.test.ts
- Test: npm test && curl -H "Authorization: Bearer TOKEN" http://localhost:5000/api/protected
From your phone, you know:
- Exactly what happened
- What's working
- How to verify it
- What files changed
- What to do next
No laptop needed.
| Feature | OpenClaw | This System |
|---|---|---|
| Setup time | 4+ days | 20 minutes |
| Dependencies | Ollama, Discord bot, complex config | Bash, curl, Discord webhooks |
| Reliability | "Unreliable for mission-critical" | Built on proven Claude Code |
| Documentation | Sparse, contradictory | You're reading it |
| Debugging | Silent failures, config hell | Straightforward bash script |
| Mobile notifications | If it works | Real-time, tested |
| Cost | Free (costs your time) | Free |
| Actually works | After 4 days: No | After 20 minutes: Yes |
- Setup Guide - Complete installation walkthrough
- Troubleshooting - Common issues and fixes
- Advanced Patterns - Multi-agent, supervisor layer, MCP integrations, scaling
- Examples - Real production results
1. Claude Code Hooks
- Built into Claude Code
- Fires on Stop, Notification, Error events
- Executes shell commands
2. Discord Webhook
- Standard REST API
- Sends formatted embeds
- Delivers to mobile instantly
3. STATUS Protocol
- Trains Claude to write useful notifications
- Loaded via CLAUDE.md at session start
- Becomes the notification content
~/.claude/
├── discord-notify.sh # Hook script (executes on events)
├── settings.json # Hooks configuration
├── CLAUDE.md # Global protocol
└── commands/
└── discord-protocol.md # Slash command (optional)
sequenceDiagram
participant H as Human
participant C as Claude Code
participant S as Hook Script
participant D as Discord
participant P as Phone
H->>C: Start autonomous task
C->>C: Work autonomously
C->>C: Complete phase
C->>S: Stop event fires
S->>D: Send webhook
D->>P: Push notification
P->>H: Buzz! 📱
H->>C: Respond with decision
C->>C: Continue work
Before:
- 8-12 hours of focused work for code analysis
- 2-3 days calendar time due to context switching
- Must be at computer for all work
After:
- 5 minutes to define task
- 6 hours of AI work (while you sleep)
- 30 minutes to review results
- 10-20x productivity multiplier
Run parallel sessions:
# Terminal 1: Frontend work
claude code --dangerously-skip-permissions
"Implement real-time audio waveforms in video editor"
# Terminal 2: Backend work
claude code --dangerously-skip-permissions
"Analyze DSP pipeline for performance bottlenecks"
# Phone: Coordinate both from Discord notificationsBefore bed:
/discord-protocol Analyze the authentication system. Find security issues. Propose improvements. Create report.
Wake up to: Comprehensive security audit with specific findings and recommendations
Scenario: You're at dinner, in Vegas, on the beach
Your phone: buzzes
Notification: "Bug fix approach: A) Refactor entire module, B) Patch specific function?"
Your response: "Approach B"
Claude: Continues work
No laptop. No interruption. Just decisions.
Add the Discord MCP server to enable Claude to read your Discord replies, not just send notifications. Full async back-and-forth from your phone.
Add a third Claude instance that monitors Discord and handles routine decisions:
# Terminal 3: Supervisor
claude code
"Monitor Discord thread [URL]. Handle routine decisions. Escalate critical choices to human."
Result: 80-90% fewer notifications, only critical decisions reach you
- Z.ai Vision MCP — Claude can take screenshots and verify visual output during autonomous sessions
- Claude in Chrome MCP — Claude can navigate browsers, run end-to-end tests, verify deployments
See Advanced Patterns for setup and configuration.
Task: "Analyze video editor. Find bugs. Propose features. Create roadmap."
Result: Full example
- 27,000 lines analyzed
- Architecture mapping
- 15 bugs (C1-C2 critical, M1-M4 moderate, L1-L9 low)
- 15 features (3 tiers: value × complexity)
- 6-month strategic roadmap
- Delivered overnight via mobile notification
Quality: Consultant-grade strategic analysis
11:23 PM 🟧 Opus 1: "Waveform rendering: Canvas vs WebGL approach?"
11:24 PM Human: "WebGL for performance, Canvas as fallback"
11:47 PM 🟦 Opus 2: "Found 12 DSP bottlenecks, prioritize?"
11:48 PM Human: "Top 3, ranked by impact"
12:15 AM 🟧 Opus 1: "WebGL waveforms implemented, 60fps on test files"
1:34 AM 🟦 Opus 2: "Top 3 bottlenecks resolved, 40% faster processing"
Two autonomous builds. One phone. Zero laptop time.
Don't begin with mission-critical features. Test on:
- Documentation improvements
- Test coverage additions
- Refactoring well-tested code
- Analysis/reconnaissance tasks
AI-generated code needs verification:
- Write tests before implementation (TDD)
- Review all diffs before merging
- Use atomic commits for easy rollback
The AI will make mistakes. Your role:
- Define clear goals
- Make strategic decisions
- Catch errors before they ship
- Verify claims with tests
- Genuinely novel architectural decisions (no prior art)
- Subjective UX choices (requires user empathy)
- Business prioritization (needs market context)
- Long tasks that exceed context window
- Break large projects into phases with clear handoff docs
- Require tests before considering work "complete"
- Add supervisor layer for routine decisions
- Human handles novel/subjective/business decisions explicitly
Q: Is this safe?
A: As safe as any autonomous coding. Review all code before shipping. Use version control. Start with non-critical work.
Q: Does it work with other LLMs?
A: The hook system is Claude Code specific, but the pattern (webhooks + protocol) works with any LLM that supports hooks/callbacks.
Q: How much does it cost?
A: Free. Discord webhooks are free. Claude Code is free. The only cost is your Claude API usage.
Q: Can I use Slack instead of Discord?
A: Yes. Slack has incoming webhooks too — just change the URL and adjust the JSON payload format.
Q: Why not just use GitHub Copilot/Cursor/etc?
A: Those are assistants. This is delegation. Copilot helps you code. This codes while you sleep.
This system works. But it can be better.
Ways to contribute:
- Share your autonomous success stories (open a Discussion)
- Add protocol templates for different project types
- Improve error handling in hook scripts
- Create Slack/Teams integration variants
- Port the supervisor pattern into a reusable template
How to contribute:
- Fork the repo
- Create your feature branch
- Test thoroughly
- Submit PR with clear description
- ✅ Discord webhook notifications
- ✅ STATUS protocol
- ✅ Multi-agent coordination
- ✅ Production-tested
- Supervisor pattern templates
- Multiple protocol templates (web dev, data science, DevOps)
- Slack integration guide
- Video tutorials
- Web dashboard for monitoring
- Statistics and analytics
- Team collaboration features
MIT License - see LICENSE file for details
Built after 4 days of fighting OpenClaw taught me that simple solutions often work best.
Thanks to Anthropic for Claude and Claude Code, the Discord team for reliable webhooks, and everyone who believed autonomous AI coding could actually work.
Autonomous AI coding isn't coming. It's here. It works. And it's simpler than you think.
The question isn't "Can AI code autonomously?"
It's "What will you build while you sleep?"
Star this repo if autonomous coding changed your workflow ⭐
Share your autonomous success stories — Open a Discussion!
Built by a developer who spent 4 days fighting complexity, then 4 hours building simplicity.
Tested in production. Delivered real results. Actually works.