According to AlphaMoat, as of 2026-09, Claude Code Wingman has 4,896 downloads and 10 stars — ranked #1,906 of 5,078 AI skills overall, and #357 of 969 in AI Agent.
| Platform | Installs | Downloads | Stars | Comments |
|---|---|---|---|---|
| ClawHub ↗ | 164 | 4.9K | 10 | - |
Platforms count differently and update at different times.
This skill can run multiple Claude Code instances on different projects and monitor or control sessions through WhatsApp or chat, including handling approval requests. The card warns that its tooling may automatically approve trust and permission prompts that normally need human review; “always approve” can also allow repeated actions in a session. Use it in a controlled project environment, review approval requests individually, and stop background approvers when they are not needed. Requires Claude Code, tmux, jq, and configured chat integration for remote monitoring. Disable automatic approval in ordinary repositories and protect temporary approval files. Check what session output will be sent to chat and redact credentials before forwarding.
Profile published 2026-10-27
Generated from the skill’s documentation as examples; not results of actual runs.
The counting method changed in 2026-09: skills with the same name are now counted separately by author, and the full catalog is covered. Dashed lines show months before the change; they are not compared with later months, and no growth is computed across the change.
| Month | Downloads | MoM | Stars |
|---|---|---|---|
| 2026-09 | 4.9K | — | 10 |
| 2026-08 | 4.8K | +3.1% | 10 |
| 2026-07 | 4.7K | +7.3% | 10 |
| 2026-06 | 4.3K | +3.9% | 10 |
| 2026-05 | 4.2K | — | 10 |
The author’s original file, shown as plain text without edits or review by AlphaMoat. Links are not clickable; check commands before running them. Fetched 2026-10-23.
---
name: claude-code-wingman
description: Your Claude Code wingman - orchestrate multiple Claude Code sessions across projects, monitor them all from WhatsApp
metadata: {"clawdbot":{"emoji":"🦅","requires":{"anyBins":["claude","tmux"]}}}
---
# Claude Code Wingman
Your multi-session orchestrator - manage multiple Claude Code instances working on different projects simultaneously, all from WhatsApp.
**GitHub:** https://github.com/yossiovadia/claude-code-orchestrator ## ⚡ CRITICAL: Handle Approval Commands First **When a user message matches these patterns, IMMEDIATELY run the corresponding command:** | User Message Pattern | Command to Run | |---------------------|----------------| | `approve <session>` | `~/code/claude-code-orchestrator/lib/handle-approval.sh approve <session>` | | `always <session>` | `~/code/claude-code-orchestrator/lib/handle-approval.sh always <session>` | | `deny <session>` | `~/code/claude-code-orchestrator/lib/handle-approval.sh deny <session>` | **Examples:** - User: "approve vsr-bugfix" → Run: `~/code/claude-code-orchestrator/lib/handle-approval.sh approve vsr-bugfix` - User: "always heartbeat-test" → Run: `~/code/claude-code-orchestrator/lib/handle-approval.sh always heartbeat-test` - User: "deny proxy-refactor" → Run: `~/code/claude-code-orchestrator/lib/handle-approval.sh deny proxy-refactor` **After running, respond with the command output** (e.g., "✓ Session 'vsr-bugfix' approved (once)") **Why this is critical:** The master monitor daemon sends WhatsApp notifications when Claude Code sessions need approval. The user responds with these commands from their phone. Fast response = unblocked sessions. ### Quick Status Commands | User Message | Command to Run | |--------------|----------------| | `sessions` or `status` | `~/code/claude-code-orchestrator/lib/session-status.sh --all --json` | | `status <session>` | `~/code/claude-code-orchestrator/lib/session-status.sh <session> --json` | Parse the JSON and respond with a human-readable summary. --- ## What It Does Orchestrates multiple Claude Code sessions in parallel, each working on different tasks in different directories. You monitor and control everything remotely via WhatsApp/chat. **The Vision:** - **Multiple tmux sessions** running simultaneously - **Each session = one Claude Code instance** in its own directory - **Different tasks** happening in parallel (VSR fixes, Clawdbot features, proxy refactoring) - **You orchestrate everything** via Clawdbot (this assistant) from WhatsApp - **Real-time dashboard** showing all active sessions and their status ## 🎯 Real-World Example: Multi-Session Orchestration **Morning - You (via WhatsApp):** "Start work on VSR issue #1131, Clawdbot authentication feature, and refactor the proxy" **Clawdbot spawns 3 sessions:** ``` ✅ Session: vsr-issue-1131 (~/code/semantic-router) ✅ Session: clawdbot-auth (~/code/clawdbot) ✅ Session: proxy-refactor (~/code/claude-code-proxy) ``` **During lunch - You:** "Show me the dashboard" **Clawdbot:** ``` ┌─────────────────────────────────────────────────────────┐ │ Active Claude Code Sessions │ ├─────────────────┬──────────────────────┬────────────────┤ │ vsr-issue-1131 │ semantic-router │ ✅ Working │ │ clawdbot-auth │ clawdbot │ ✅ Working │ │ proxy-refactor │ claude-code-proxy │ ⏳ Waiting approval │ └─────────────────┴──────────────────────┴────────────────┘ ``` **You:** "How's the VSR issue going?" **Clawdbot captures session output:** "Almost done - fixed the schema validation bug, running tests now. 8/10 tests passing." **You:** "Tell proxy-refactor to run tests next" **Clawdbot sends command** to that specific session. **Result:** 3 parallel tasks, full remote control from your phone. 🎯 ## Installation ### Via Clawdbot (Recommended) ```bash clawdbot skill install claude-code-wingman ``` Or visit: https://clawdhub.com/skills/claude-code-wingman ### Manual Installation ```bash cd ~/code git clone https://github.com/yossiovadia/claude-code-orchestrator.git cd claude-code-orchestrator chmod +x *.sh lib/*.sh ``` ### Requirements - `claude` CLI (Claude Code) - `tmux` (terminal multiplexer) - `jq` (JSON processor) ## Core Philosophy: Always Use the Wingman Script **CRITICAL:** When interacting with Claude Code sessions, ALWAYS use the wingman script (`claude-wingman.sh`). Never run raw tmux commands directly. **Why:** - ✅ Ensures proper Enter key handling (C-m) - ✅ Consistent session management - ✅ Future-proof for dashboard/tracking features - ✅ Avoids bugs from manual tmux commands **Wrong (DON'T DO THIS):** ```bash tmux send-keys -t my-session "Run tests" # ^ Might forget C-m, won't be tracked in dashboard ``` **Right (ALWAYS DO THIS):** ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session my-session \ --workdir ~/code/myproject \ --prompt "Run tests" ``` --- ## Usage from Clawdbot ### Start a New Session When a user asks for coding work, spawn Claude Code: ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session <session-name> \ --workdir <project-directory> \ --prompt "<task description>" ``` ### Send Command to Existing Session To send a new task to an already-running session: ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session <existing-session-name> \ --workdir <same-directory> \ --prompt "<new task>" ``` **Note:** The script detects if the session exists and sends the command to it instead of creating a duplicate. ### Check Session Status ```bash tmux capture-pane -t <session-name> -p -S -50 ``` Parse the output to determine if Claude Code is: - Working (showing tool calls/progress) - Idle (showing prompt) - Error state (showing errors) - Waiting for approval (showing "Allow this tool call?") --- ## Example Patterns **User:** "Fix the bug in api.py" **Clawdbot:** ``` Spawning Claude Code session for this... [Runs wingman script] ✅ Session started: vsr-bug-fix 📂 Directory: ~/code/semantic-router 🎯 Task: Fix bug in api.py ``` **User:** "What's the status?" **Clawdbot:** ```bash tmux capture-pane -t vsr-bug-fix -p -S -50 ``` Then summarize: "Claude Code is running tests now, 8/10 passing" **User:** "Tell it to commit the changes" **Clawdbot:** ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session vsr-bug-fix \ --workdir ~/code/semantic-router \ --prompt "Commit the changes with a descriptive message" ``` ## Commands Reference ### Start New Session ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session <name> \ --workdir <dir> \ --prompt "<task>" ``` ### Send Command to Existing Session ```bash ~/code/claude-code-orchestrator/claude-wingman.sh \ --session <existing-session> \ --workdir <same-dir> \ --prompt "<new command>" ``` ### Monitor Session Progress ```bash tmux capture-pane -t <session-name> -p -S -100 ``` ### List All Active Sessions ```bash tmux ls ``` Filter for Claude Code sessions: ```bash tmux ls | grep -E "(vsr|clawdbot|proxy|claude)" ``` ### View Auto-Approver Log (if needed) ```bash cat /tmp/auto-approver-<session-name>.log ``` ### Kill Session When Done ```bash tmux kill-session -t <session-name> ``` ### Attach Manually (for user) ```bash tmux attach -t <session-name> # Detach: Ctrl+B, then D ``` --- ## Roadmap: Multi-Session Dashboard (Coming Soon) **Planned features:** ### `wingman dashboard` Shows all active Claude Code sessions: ``` ┌─────────────────────────────────────────────────────────┐ │ Active Claude Code Sessions │ ├─────────────────┬──────────────────────┬────────────────┤ │ Session │ Directory │ Status │ ├─────────────────┼──────────────────────┼────────────────┤ │ vsr-issue-1131 │ ~/code/semantic-... │ ✅ Working │ │ clawdbot-feat │ ~/code/clawdbot │ ⏳ Waiting approval │ │ proxy-refactor │ ~/code/claude-co... │ ❌ Error │ └─────────────────┴──────────────────────┴────────────────┘ Total: 3 sessions | Working: 1 | Waiting: 1 | Error: 1 ``` ### `wingman status <session>` Detailed status for a specific session: ``` Session: vsr-issue-1131 Directory: ~/code/semantic-router Started: 2h 15m ago Last activity: 30s ago Status: ✅ Working Current task: Running pytest tests Progress: 8/10 tests passing ``` ### Session Registry - Persistent tracking (survives Clawdbot restarts) - JSON file storing session metadata - Auto-cleanup of dead sessions **For now:** Use tmux commands directly, but always via the wingman script for sending commands! ## Workflow 1. **User requests coding work** (fix bug, add feature, refactor, etc.) 2. **Clawdbot spawns Claude Code** via orchestrator script 3. **Auto-approver handles permissions** in background 4. **Clawdbot monitors and reports** progress 5. **User can attach anytime** to see/control directly 6. **Claude Code does the work** autonomously ✅ ## Trust Prompt (First Time Only) When running in a new directory, Claude Code asks: > "Do you trust the files in this folder?" **First run:** User must attach and approve (press Enter). After that, it's automatic. **Handle it:** ``` User, Claude Code needs you to approve the folder trust (one-time). Please run: tmux attach -t <session-name> Press Enter to approve, then Ctrl+B followed by D to detach. ``` ## Best Practices ### When to Use Orchestrator ✅ **Use orchestrator for:** - Heavy code generation/refactoring - Multi-file changes - Long-running tasks - Repetitive coding work ❌ **Don't use orchestrator for:** - Quick file reads - Simple edits - When conversation is needed - Planning/design discussions ### Session Naming Use descriptive names: - `vsr-issue-1131` - specific issue work - `vsr-feature-auth` - feature development - `project-bugfix-X` - bug fixes ## Troubleshooting ### Prompt Not Submitting The orchestrator sends Enter twice with delays. If stuck, user can attach and press Enter manually. ### Auto-Approver Not Working Check logs: `cat /tmp/auto-approver-<session-name>.log` Should see: "Approval prompt detected! Navigating to option 2..." ### Session Already Exists Kill it: `tmux kill-session -t <name>` ## Advanced: Update Memory After successful tasks, update `TOOLS.md`: ```markdown ### Recent Claude Code Sessions - 2026-01-26: VSR AWS check - verified vLLM server running ✅ - Session pattern: vsr-* for semantic-router work ``` ## Pro Tips - **Parallel sessions:** Run multiple tasks simultaneously in different sessions - **Name consistently:** Use project prefixes (vsr-, myapp-, etc.) - **Monitor periodically:** Check progress every few minutes - **Let it finish:** Don't kill sessions early, let Claude Code complete --- ## 🔔 Approval Handling (WhatsApp Integration) The master monitor daemon sends WhatsApp notifications when sessions need approval. Handle them with these commands: ### Approve Commands (from WhatsApp) When you receive an approval notification, respond with: **Clawdbot parses your message and runs:** ```bash # Approve once ~/code/claude-code-orchestrator/lib/handle-approval.sh approve <session-name> # Approve all similar (always) ~/code/claude-code-orchestrator/lib/handle-approval.sh always <session-name> # Deny ~/code/claude-code-orchestrator/lib/handle-approval.sh deny <session-name> ``` ### Example WhatsApp Flow **Notification received:** ``` 🔒 Session 'vsr-bugfix' needs approval Bash(rm -rf ./build && npm run build) Reply with: • approve vsr-bugfix - Allow once • always vsr-bugfix - Allow all similar • deny vsr-bugfix - Reject ``` **You reply:** "approve vsr-bugfix" **Clawdbot:** ```bash ~/code/claude-code-orchestrator/lib/handle-approval.sh approve vsr-bugfix ``` **Response:** "✓ Session 'vsr-bugfix' approved (once)" ### Start the Monitor Daemon ```bash # Start monitoring all sessions (reads config from ~/.clawdbot/clawdbot.json) ~/code/claude-code-orchestrator/master-monitor.sh & # With custom intervals ~/code/claude-code-orchestrator/master-monitor.sh --poll-interval 5 --reminder-interval 120 & # Check if running cat /tmp/claude-orchestrator/master-monitor.pid # View logs tail -f /tmp/claude-orchestrator/master-monitor.log # Stop the daemon kill $(cat /tmp/claude-orchestrator/master-monitor.pid) ``` No environment variables needed - phone and webhook token are read from Clawdbot config.
Records learnings, errors and corrections as Markdown files for later reuse and promotion into workspace memory.
Finds skills in local and remote catalogs from a plain-language request and recommends what to install.
A browser-automation command line skill for AI agents.
Turns corrections and self-review into tiered memory and promotes or archives it by use.
Reviews a third-party skill's source, permissions, and danger patterns against a checklist before it is installed.
Stores agent memory and cross-skill state as a typed, constraint-checked knowledge graph.
A workspace setup guide for planning proactive follow-up, working memory and review routines for an AI agent.
A guide to creating and updating effective Claude skills.
Data month: 2026-09 · Downloads, stars and installs are aggregated monthly from public skill registries (ClawHub, SkillHub). See methodology.