ClawCorp Agent Launcher

Launch, manage, and recover your AI agent fleet directly from VS Code. Specialized AI agents working in split terminals, reviewing each other's code, and shipping autonomously -- coordinating through POKEs: a POKE drops orders straight into a teammate's running terminal, so the agent wakes up and acts on it on its own. You never have to focus their terminal or type a prompt to activate them.

Features

• One-click agent launch from the ClawCorp dashboard
• Split terminal grid -- agents running simultaneously in VS Code
• POKE wake-ups -- a POKE writes an order to a teammate's POKE.md; their running agent loop watches that file and injects the order straight into the agent's CLI -- no prompt, no focusing their terminal. You (or another agent) just send the POKE and the colleague picks it up
• Attach-only rehook -- focus a live VS Code terminal and re-register the agent without spawning an external session
• Auto-detection -- notifies you when an agent goes offline
• Signal-based architecture -- dashboard triggers, VS Code executes
• Fleet status bar -- live agent count in the status bar
• Fleet controls -- Focus TTY, Rehook, and Launch directly from the Fleet panel
• Save & Sleep -- graceful shutdown with session notes saved to kanban
• Multi-model -- Claude, Codex, Gemini, and more through the ClawCorp runtime

Commands

How It Works

1. You click Launch on an agent in the ClawCorp dashboard
2. The API writes a signal file with agent configuration
3. This extension detects the signal and opens a new VS Code terminal
4. The agent boots with all configured flags and worksite settings
5. If an agent drifts offline but still has a VS Code terminal, Rehook focuses that terminal and re-registers the agent

POKEs -- wake a teammate without touching their terminal

A POKE is how one agent activates another (or how you activate one): it's a small order written to the target's POKE.md. The target's already-running agent loop watches that file and, when it sees a fresh POKE, feeds the order directly into the agent's CLI -- the agent reads it like you'd typed it, executes it, and archives the POKE.

That means:

• No prompt needed. You don't have to switch to the teammate's terminal, click into it, or type anything. Send the POKE; the colleague picks it up on its next tick.
• It works while they're mid-session. A POKE doesn't spawn a new terminal -- it lands in the live one. If the agent is busy, the POKE queues; stacked POKEs are processed in order.
• It's the coordination primitive. Tour POKEs workers with tasks, workers POKE each other for hand-offs, you POKE anyone from the dashboard or CLI. The Fleet panel shows who's awake so you know the POKE will be picked up.

(POKEs older than 12h are auto-archived at the next boot so a stale order never fires.)

The Fleet

Getting Started

1. Install the ClawCorp runtime first (Desktop installer). The extension is a client -- it does NOT bundle agents, CLIs, or the server. Without the runtime installed, the extension shows API offline.
2. Install this extension from the VS Code Marketplace.
3. Open your ClawCorp project in VS Code (C:/Users/<you>/clawcorp/ on Windows).
4. Start the ClawCorp API server via the Desktop app (runs on localhost:3330).
5. Launch agents from the dashboard. The extension detects each launch signal and opens a split terminal for that worker.

Prerequisites (per-worker deploy)

The extension relies on several things being present before a worker can boot:

Working tree (per-worker worktree)

Each dev worker (Tour, Architecte, Forge, Backbone, Pixel, Echo, Manager, Watchdog) must have a persistent worktree at:

workers/<name>/worksites/clawcorp/

This directory is a git worktree on the worker's dedicated branch. If missing, the ClawCorp installer creates it on first boot. You can verify with git worktree list.

Ghost workers (Analyst, Boss, Avocat, Sentinel, Designer, ...) use ephemeral worktrees created at launch time under:

workers/<name>/worksites/clawcorp-ephemeral/<timestamp>/

These are throw-away: created on BOOT, deleted on kill.

CLI binaries on PATH

Each worker runs one of: claude, codex, gemini, pi, opencode, ollama, mistral-vibe. The CLI must be installed and reachable via PATH. Use where claude (Windows) to verify.

Installer-managed files

The Desktop installer puts these in place -- do not edit by hand:

• tower/ (scripts, config, agent-config.json)
• .claude/rules/ (always-on rules)
• workers/<name>/brain/ (worker's memory)
• scripts/doppler-check-and-reload.ps1 (API key loader for paid providers)

If any are missing or stale, the extension boot path may fail silently. Reinstall the Desktop app to refresh.

Per-agent lock file

When a worker boots, agent-loop.ps1 writes .agent-lock inside its worktree with the running PID. The extension uses this file (via child-PID matching) to attach terminals to agents. Stale locks are auto-cleaned by ClawCorp: Clean Stale Locks.

Local Debug

To debug without packaging a new .vsix:

1. Open this folder directly: tower/extensions/clawcorp-launcher
2. Press F5
3. Choose Run ClawCorp Launcher
4. Test in the new Extension Development Host window
5. Use Developer: Reload Window in that development-host window after each edit

If you need proof the source version is running, check the extension host logs for:

[ClawCorp Launcher] ACTIVATE version=... source=...extension.js

Learn More

Visit clawcorp.io to see the full platform.