Managed OpenClaw on OpenComputer

Deploy isolated, per-employee OpenClaw AI agent instances inside OpenComputer sandboxes. Each employee gets their own sandboxed agent accessible via Telegram, a web chat UI, or both.

What You Get

• Instant provisioning — new employee agents spin up in seconds from a pre-built snapshot
• Full isolation — each agent runs in its own KVM-backed sandbox with dedicated resources
• Telegram integration — connect each agent to its own Telegram bot
• Web chat UI — built-in chat proxy so employees can talk to their agent from a browser
• Auto-hibernate — sandboxes sleep when idle, wake instantly on demand
• Zero infrastructure — no servers, DNS, reverse proxies, or certificates to manage
• Fleet management — health checks, rolling updates, and config changes across all agents
• Checkpoint & restore — snapshot full agent state, fork it for new employees

Architecture

    Browser (chat UI)          Telegram
         |                        |
         v                        v
   ┌─────────────┐         ┌───────────┐
   │ Chat Server  │         │ TG Bot API│
   │ (proxy via   │         │           │
   │  OC SDK)     │         └─────┬─────┘
   └──────┬──────┘               │
          │                      │
          v                      v
   ┌──────────── OpenComputer API ────────────┐
   │                                          │
   │  ┌──────────┐ ┌──────────┐ ┌──────────┐  │
   │  │ Sandbox  │ │ Sandbox  │ │ Sandbox  │  │
   │  │ emp-001  │ │ emp-002  │ │ emp-N    │  │
   │  │          │ │          │ │          │  │
   │  │ OpenClaw │ │ OpenClaw │ │ OpenClaw │  │
   │  │ Gateway  │ │ Gateway  │ │ Gateway  │  │
   │  │ (loopback│ │ (loopback│ │ (loopback│  │
   │  │  only)   │ │  only)   │ │  only)   │  │
   │  └──────────┘ └──────────┘ └──────────┘  │
   └──────────────────────────────────────────┘

The gateway binds to loopback only — no ports are exposed externally. All chat access goes through the proxy server, which connects to sandboxes via the OpenComputer SDK.

Quick Start

# Install dependencies
npm install

# Set environment
export OPENCOMPUTER_API_KEY="your-api-key"

# 1. Build the OpenClaw snapshot (one-time, ~3 min)
npx tsx src/build-snapshot.ts

# 2. Provision an employee agent
npx tsx src/provision-claw.ts \
  --employee-id emp-001 \
  --anthropic-api-key "sk-ant-..."

# 3. Start the chat server
npx tsx src/chat-server.ts
# Open http://localhost:3000/emp-001

# 4. (Optional) Add Telegram
npx tsx src/configure-telegram.ts \
  --employee-id emp-001 \
  --tg-bot-token "123456:ABC..." \
  --tg-user-id "987654321"

# 5. Check fleet health
npx tsx src/fleet-health.ts

Provisioning Options

npx tsx src/provision-claw.ts \
  --employee-id emp-001 \
  --anthropic-api-key "sk-ant-..." \
  --model "anthropic/claude-sonnet-4-6" \   # default
  --timeout 600 \                            # idle timeout in seconds
  --memory 4096                              # sandbox memory in MB

Files

File	Purpose
src/build-snapshot.ts	Build a reusable OpenClaw snapshot with Node.js, OpenClaw, and Telegram deps pre-installed
src/build-checkpoint.ts	Alternative: build via checkpoint (step-by-step, more control)
src/provision-claw.ts	Provision a single employee's OpenClaw sandbox from snapshot
src/provision-from-checkpoint.ts	Provision from a checkpoint instead of snapshot
src/configure-telegram.ts	Add Telegram to an existing employee's agent
src/chat-server.ts	Web chat proxy — serves a chat UI per employee, proxies to sandbox via OC SDK
src/chat-ui/index.html	Chat UI frontend (served by chat-server)
src/fleet-health.ts	Health check all running agents
src/fleet-update.ts	Rolling update across the fleet
src/manage-secrets.ts	Create/update encrypted secret stores for employees
src/registry.ts	Fleet registry helpers (read/write fleet-registry.json)
src/config.ts	Shared configuration and constants

Security

• Gateway is loopback-only — no network ports exposed from any sandbox
• Chat proxy — browser never sees gateway URLs or tokens; all access goes through the server-side OC SDK
• Per-employee tokens — each sandbox gets a unique gateway auth token generated at provision time
• Tool auto-approval — exec commands are auto-approved since agents run in isolated sandboxes
• Idle timeout — sandboxes auto-hibernate after configurable idle period (default 10 min)