Setup Guide: Self-Hosted Agents with OpenClaw

The short version

OpenClaw is the self-hosted agent runtime for teams that can't use Claude's managed setup for compliance, security, or regulatory reasons. Same agents as the Claude setup, same MCP data source connectors, but everything stays inside your firewall on Docker. The architecture has five components: gateway, brain (reasoning engine with pluggable LLMs), memory, skills (MCP servers), monitoring. Prerequisites: Docker, 8GB+ RAM, network isolation, API keys. Run on stable on-prem hardware. Pick OpenClaw if data cannot leave your network or you need custom model routing. Pick Claude's managed setup otherwise. Note: OpenClaw shipped with security vulnerabilities through 2025-2026; audit every plugin and run in a network-isolated environment.

OpenClaw is the self-hosted agent runtime. Same agents from the AI Agent Army. Same data source connectors. Same cron-based scheduling. Your infrastructure. Your network. No data leaving your systems.

This guide is for teams that picked Claude's setup but can't use it for compliance, security, or regulatory reasons. If that's you, this is how you get the same result on your own hardware.

When to Use OpenClaw vs Claude

Pick Claude if:

• You want managed infrastructure and automatic patching.
• Your data isn't restricted from cloud processing.
• You need the most capable models out of the box.
• You don't want to run a data center.

Pick OpenClaw if:

• Your data cannot leave your network (compliance, regulated industries).
• You need custom model routing or inference control.
• You want to reduce API costs at scale.
• You're auditing every line of reasoning code.

Architecture: What's Different

Claude setup: Your agents live on Anthropic's servers. You send data via API. Results come back.

OpenClaw setup: Your agents live on your hardware. Everything stays inside your firewall. You manage the infrastructure.

The architecture has five components:

• Gateway: Routes agent requests.
• Brain: Reasoning engine with pluggable LLMs.
• Memory: Stores context and tool results.
• Skills: Tools that connect to your data sources (MCP servers).
• Monitoring: Health checks and audit logs.

All MCP servers from the Claude setup work here unchanged. Jira, Slack, Zendesk, Salesforce, GitHub, Google Calendar, Gong, Notion, analytics - same connectors.

Prerequisites

You need:

• Docker and Docker Compose on a stable server.
• 8GB+ RAM (more if running large models locally).
• Ability to run a persistent container.
• API keys for each data source (same ones as Claude setup).
• Network isolation (firewall rules so agents don't leak data).

Installation and Basic Config

Clone the OpenClaw repo and set up your .env file:

OPENCLAW_MODEL=mistral
OPENCLAW_PORT=8000
STORAGE_PATH=/data/openclaw
VAULT_PATH=/data/vault

Run the Docker container:

docker-compose up -d

Verify with: curl http://localhost:8000/health

See the full artifact for Dockerfile templates and network setup.

Connecting Data Sources

The same connectors work here. Start with three:

Jira: Configure MCP server with your instance URL and API token. Agents can now read tickets, update sprints, track velocity.

Slack: Use Slack MCP with your workspace token. Agents can post updates, respond to mentions, deliver scheduled reports.

GitHub: Connect with personal access token. Agents review PRs, track releases, manage deployments.

Then add the others as needed:

• Salesforce (pipeline sync, forecast updates)
• Zendesk (ticket routing, knowledge base)
• Google Calendar (scheduling, meeting context)
• Gong (call insights)
• Notion (documentation, templates)
• Analytics (KPI tracking)

Each takes 15-40 minutes to connect. See the artifact for step-by-step MCP server setup for each.

Scheduling and Delivery

Same approach as Claude setup. Define cron schedules in your config:

agents:
  - name: daily_standup
    schedule: "0 9 * * 1-5"
    delivery: slack
    channel: #updates

Agents run on schedule, gather data from your sources, and deliver to Slack. No external webhooks. No data leaving your network.

Security: Your Responsibility

You own the security stack:

• Encrypt all credentials at rest (use a secrets manager).
• Run OpenClaw on an isolated network segment.
• Set up audit logging for all agent actions.
• Monitor resource usage. Agents with access to credentials are high-value targets.
• Patch dependencies regularly.
• Rotate API keys annually.

Data never leaves your network. Audit logs stay on your servers. You control who can access agent outputs.

Claude vs OpenClaw: Quick Decision Matrix

Next Steps

Ready to set up? Download the artifact for detailed Dockerfile templates, MCP server configs, and security checklists.

Want the simpler path? See the Claude MCP setup guide. Zero infrastructure overhead. Your tradeoff is data flows through Anthropic's servers.

Then explore the AI Agent Army overview to understand which agents work best for your workflows. Each agent page has a "Connect via OpenClaw" section.

Start with Jira, Slack, and GitHub. Get three agents running. That's your baseline. Expand from there.