Run a self-hosted AI agent safely on your everyday computer.
No VPS. No Mac Mini. No server closet. Just Docker on the laptop or desktop you already have.

Clawbox is a hardened, one-command setup for OpenClaw with built-in web search.
Everything runs inside Docker containers — nothing touches your host system.

---

Why Clawbox?

Most AI agent setups assume you have a dedicated server, cloud VM or a mac mini. Clawbox is different — it's designed to run on your personal machine and locked down so you can run it safely:

• Nothing escapes Docker — the agent, its tools, and the search engine all run inside containers. Your host filesystem, apps, and data stay untouched.
• Hardened by default — localhost-only networking, command approval prompts, and sandboxed execution. No open ports, no surprises.
• Your laptop is enough — if it runs Docker Desktop, it runs Clawbox. No need to rent a VPS or dedicate hardware.

What you get

• OpenClaw — AI agent gateway with a browser-based chat UI
• SearXNG — self-hosted metasearch engine (private, no tracking, no API keys)
• Web search out of the box — the agent can search the web and read URLs immediately
• Hardened by default — localhost-only, sandboxed, with exec approvals
• Cross-platform — works on macOS, Linux, WSL, and Windows

Requirements

• Docker Desktop — download here (includes Docker Compose)
  • macOS: Docker Desktop for Mac
  • Windows: Docker Desktop for Windows (with WSL 2 backend)
  • Linux: Docker Engine + Docker Compose v2 (docker compose command)
• Git — to clone the repository
• An AI provider account — one of:
  • OpenAI (recommended) — uses OAuth login, no API key needed
  • Anthropic — requires an API key

Quick start

macOS / Linux

git clone https://github.com/iceHub82/clawbox.git
cd clawbox
./setup.sh

Windows (PowerShell)

git clone https://github.com/iceHub82/clawbox.git
cd clawbox
.\setup.ps1

What the setup script does

1. Creates configuration files
2. Pulls the Docker images
3. Starts the containers (OpenClaw + SearXNG)
4. Installs web search tools inside the gateway
5. Applies security hardening
6. Launches the configuration wizard (see below)
7. Opens your browser automatically

Configuration wizard

During setup, you'll be guided through these steps:

1. "Where will the Gateway run?" → Select Local (this machine)
2. "Select sections to configure" → Select Model (Pick provider + credentials)
3. Choose your AI provider and model
   • For OpenAI: a browser window opens for OAuth login — no API key needed
   • For Anthropic: paste your API key when prompted
4. "Select sections to configure" → Select Continue (Done)

Once complete, your browser opens to the OpenClaw chat UI — you're ready to go.

Security

Running an AI agent on your local machine carries real risk. The agent can execute shell commands, read/write files, and access your network. Clawbox takes this seriously and ships hardened by default — every layer is locked down out of the box.

What hardened mode does

Layer	Protection	Why it matters
Network isolation	Gateway bound to localhost only	No one on your network can access the agent
Exec approvals	Agent must ask before running any command	You review and approve every shell command
Sandbox	Agent tools run in isolated Docker containers	Commands can't touch your host filesystem
Internal SearXNG	Search engine not exposed to host	Only the agent can reach it, not your browser or LAN
Auth token	Gateway requires a token to connect	Prevents unauthorized access even on localhost
Docker isolation	Everything runs inside containers	Nothing is installed on your host system

Hardened vs Open mode

Setting	Hardened (default)	Open (--open)
Port binding	localhost only	LAN-accessible
Exec approvals	Required — you approve every command	Off
Sandbox	Enabled — isolated containers	Off
SearXNG	Internal only	Exposed on host

We strongly recommend running in hardened mode unless you specifically need LAN access. To disable hardening:

./setup.sh --open          # macOS / Linux
.\setup.ps1 -Open          # Windows

Security tips

• Never run --open on an untrusted network — it exposes the agent to your entire LAN
• Review exec approvals carefully — the agent will show you the command before running it
• Keep Docker Desktop updated — container isolation depends on it
• Don't store sensitive files in the workspace directory — the agent has full access to it

Architecture

Browser → OpenClaw Gateway (:18789)
              ↓ (bash tool)
           mcporter → mcp-searxng (stdio) → SearXNG (:8080)

Two persistent containers:

1. openclaw-gateway — runs the AI agent, mcporter, and mcp-searxng
2. searxng — local metasearch engine with JSON API

Common commands

Command	Description
./oc configure	Re-run the configuration wizard
./oc token	Show the gateway auth token
docker compose stop	Stop all containers
docker compose up -d	Start all containers
docker compose down	Stop and remove containers

Troubleshooting

"Pairing required" in the browser The setup script auto-approves the browser device. If it didn't work, refresh the page. If it still shows, run:

./oc devices list

Then approve the pending device ID shown.

Gateway not accessible Make sure Docker Desktop is running and the containers are up:

docker compose ps

Want to start fresh?

docker compose down -v
rm -rf data/ .env
./setup.sh

Customization

Copy .env.example to .env to customize ports, paths, and image tags. The setup script creates this automatically, but you can edit it afterwards:

Variable	Default	Description
OPENCLAW_GATEWAY_PORT	18789	Gateway UI port
OPENCLAW_BRIDGE_PORT	18790	Bridge port
SEARXNG_PORT	8888	SearXNG port (open mode only)
BIND_HOST	127.0.0.1	127.0.0.1 = localhost, 0.0.0.0 = LAN
OPENCLAW_IMAGE	ghcr.io/openclaw/openclaw:latest	Docker image

Adding more MCP tools

You can extend the agent with additional MCP tools:

# Install the MCP server package
docker exec -u root clawbox-openclaw-gateway-1 npm install -g <mcp-package>

# Register it with mcporter
docker exec clawbox-openclaw-gateway-1 mcporter config add <name> \
  --command <mcp-binary> --scope home

The agent can then use it via mcporter call <name>.<tool> key=value.

License

MIT