OpenClaw Setup Guide

Complete setup and configuration guide for OpenClaw — the agent with the fastest GitHub star growth in history. Skills & Tools model, NVIDIA NemoClaw, Pi SDK engine, security hardening.

June 9, 2026
openclawself-hostedagent-configurationskillstoolsnemo-claw

OpenClaw

OpenClaw crossed 250,000 GitHub stars faster than any AI framework in history. NVIDIA built NemoClaw on top of it. JustPaid ran 7 OpenClaw agents 24/7 and shipped 10 features in a month. It's the most popular open-source agent framework by GitHub metrics.

Under the hood, OpenClaw uses Pi Coding Agent's SDK as its agent engine. Skills & Tools are the atomic units of capability — install from the marketplace, configure with YAML, and the agent discovers them at runtime.

For an introductory overview, see the OpenClaw Guide blog post. This page covers production setup and security hardening.

Note:

Security alert. OpenClaw has documented CVEs related to unauthenticated MCP tool access and malicious plugin injection. Zapier's enterprise guide calls it out specifically. For production deployments, follow the security hardening steps below. Consider Hermes Agent if security is your primary concern.

Installation

1

Clone and Install

git clone https://github.com/OpenClaw/OpenClaw.git
cd OpenClaw
npm install
2

Configure Environment

cp .env.example .env

At minimum: one LLM provider API key and one messaging platform.

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
TELEGRAM_BOT_TOKEN=...
3

Start

npm run start

Skills & Tools

Skills define what the agent can do. Tools are the atomic actions skills use. The marketplace has thousands of community skills. Install them declaratively.

# skills.yml
skills:
  - name: web-search
    tools: [brave-search, firecrawl]
    instructions: "Search the web using Brave. Extract pages with Firecrawl."

  - name: code-execution
    tools: [run-python, run-bash, read-file, write-file]
    instructions: "Execute code in a sandbox. Read and write files."

  - name: github
    tools: [github-issues, github-prs, github-search]
    instructions: "Create issues, review PRs, search repositories."
# Install a community skill
openclaw skills install openclaw/skills/email-automation

# Enable it
openclaw skills enable email-automation

# List active skills
openclaw skills list

Essential Skills

openclaw/skills/web-search
Brave Search + Firecrawl + Jina Reader. Universal web research toolkit.

Values: 95K installs

openclaw/skills/github
Full GitHub integration. Issues, PRs, search, code browsing.

Values: 88K installs

openclaw/skills/code-exec
Sandboxed Python, Bash, and Node.js execution.

Values: 80K installs

openclaw/skills/email-automation
Gmail + Outlook. Read, categorize, draft, send.

Values: 65K installs

openclaw/skills/data-pipeline
SQL queries, CSV/JSON parsing, chart generation. Postgres, SQLite, BigQuery.

Values: 52K installs

NVIDIA NemoClaw

NVIDIA built NemoClaw on top of OpenClaw for GPU-optimized agent inference. It adds:

  • GPU-accelerated inference — Runs models directly on NVIDIA GPUs via TensorRT-LLM
  • Microservice architecture — Each agent skill runs as an independent NIM (NVIDIA Inference Microservice)
  • Multi-GPU scaling — Distribute agent workloads across GPU clusters
# NemoClaw uses the same skills as OpenClaw but with GPU acceleration
nemoclaw run --skill web-search --gpu a100-80gb

Note:

When to use NemoClaw. If you're self-hosting model inference on NVIDIA GPUs (not using cloud APIs), NemoClaw gives you 3-5x faster token generation compared to running the same model on CPU or through Ollama. For API-based workflows (Claude, GPT-4o via API), standard OpenClaw is sufficient.

Security Hardening

1

Restrict Tool Access

Never expose all tools to a single agent instance. Define tool allowlists per agent.

# harden.yml
tool_policy:
  allowlist: [web-search, file-read, github-read]
  blocklist: [file-delete, shell-exec, db-write]
2

Sandbox Execution

Run OpenClaw inside a container with minimal filesystem access.

docker run -v $(pwd)/workspace:/workspace:ro \
  -v $(pwd)/skills:/skills:ro \
  openclaw:latest

Mount the workspace as read-only. Only mount write access to a dedicated output directory.

3

Audit Plugin Sources

Skills from the marketplace are community-contributed. Audit before installing.

# Review a skill before enabling
openclaw skills inspect openclaw/skills/email-automation

# Pin to a specific version
openclaw skills install openclaw/skills/[email protected]
4

Enable MCP Authentication

If using MCP tools, require authentication.

mcp:
  auth:
    required: true
    tokens:
      - key: mcp-token-1
        scopes: [read, search]

Note:

Do not run OpenClaw with --no-sandbox in production. The flag disables all tool restrictions and allows the agent to execute arbitrary commands on the host system. Only use it in isolated development environments.

Provider Configuration

Same providers as Hermes Agent. Configure in providers.yml:

providers:
  anthropic:
    api_key: ${ANTHROPIC_API_KEY}
    default_model: claude-sonnet-4-20250514

  openai:
    api_key: ${OPENAI_API_KEY}
    default_model: gpt-4o

  ollama:
    endpoint: http://localhost:11434
    default_model: llama3.2:3b

Running Blueprints on OpenClaw

Package any blueprint as an OpenClaw skill:

# Create a skill from the research agent blueprint
mkdir -p ~/.openclaw/skills/research-agent
cp blueprints/research-agent/* ~/.openclaw/skills/research-agent/

# Register tools from the blueprint
openclaw tools register brave-search firecrawl fact-check cite-source

# Enable the skill
openclaw skills enable research-agent

Key Takeaway

OpenClaw is the right choice when you need the largest skill ecosystem and community support. Its marketplace has thousands of pre-built integrations. The tradeoff is security — you must harden it before production use. For teams prioritizing security over ecosystem size, Hermes Agent is the safer choice.