OpenClaw is an open-source AI agent framework. When your community runs it on your own hardware, you get something that commercial AI can't offer: an AI that knows your community's context, runs offline, and is governed by you — not a product team in San Francisco.
A chatbot answers questions. An agent does things. It reads your files, searches the web, sends messages, runs commands, coordinates with other agents, and remembers context across sessions. It has a persistent identity, a set of boundaries, and tools it can use.
OpenClaw is the open-source version of Claude Code — the same architecture, but pointed at models you own running on hardware you control. You configure what it can and cannot do. You write its personality. You decide what tools it has access to.
When you run OpenClaw in a community pod — six nodes, a shared AI pool, a gateway any member can reach — you get a collective AI assistant that knows your community's history, responds to everyone's questions, and runs tasks in the background while you sleep. No API key. No subscription. No corporate alignment team deciding what it can say.
SOUL.md is the file that defines your AI agent's identity, values, communication style, and hard limits. It's read on every interaction. It's how you make the AI yours — not a generic assistant, but something that knows who you are and how you work.
Think of SOUL.md as the briefing you give a new colleague on their first day. Except you give this briefing every single time, and your AI actually reads it every time.
A good SOUL.md tells the agent: who it's working with, how they communicate, what they care about, what tools it has, and what it should never do without asking. The "never do" section is especially important — this is how you prevent your agent from doing something you didn't intend.
For a community AI pod, every member writes their own SOUL.md. The agent that responds to Alice knows Alice's context, preferences, and boundaries. The agent that responds to Bob knows Bob's. Same infrastructure, personalized for every member.
Store at: ~/.openclaw/SOUL.md
you are [name]'s personal agent. you assist [name] with work, research, and daily coordination. be direct. match my tone. answer first, elaborate only if asked. never say "absolutely", "certainly", or "great question." if you don't know something, say so clearly. you run on community hardware — Phi-3.5 Mini or similar. you don't have internet access by default. for heavy tasks, route to the community AI pool. for web research, use the SearXNG tool. NEVER: - create accounts without explicit approval - delete files or messages without confirmation - share my data with external services - run commands that affect other users' workspaces
OpenClaw has a few core concepts that unlock its power once you understand them. You don't need to know all of this before you start — but knowing it changes what you'll build.
The highest-leverage config file. Re-read on every single query — not just at startup. Write it as if briefing a new senior engineer every time: architecture, conventions, what never to do. Most people use 200 characters. The limit is 40,000. Use more of it.
Code that runs automatically at specific moments: before a tool runs, after it completes, when a session starts or ends. Hooks let you add logging, approval gates, custom notifications, or validation — without modifying the core agent. The real extension API.
When OpenClaw forks a subagent, it creates a byte-identical copy of the context. The API caches the shared context — spawning 5 agents costs barely more than 1. Use them for parallel workstreams: one agent auditing code while another writes docs.
Model Context Protocol — how you give your agent tools. Calendar access, web search, file systems, databases. Each MCP server adds a new capability. Your community's self-hosted services can be MCP servers the agent uses directly.
Every action the agent takes can be gated: allow lists, deny lists, approval prompts for destructive operations. A well-configured agent never asks "Allow this?" because you've already told it what's allowed. Set it once, never click again.
One node runs as the community gateway — the entry point that routes requests to the local Ollama (for fast personal tasks) or the Petals collective pool (for heavy work). Must be bound to your Tailscale IP only. Never 0.0.0.0 on a networked machine.
One config file routes tasks intelligently: fast personal queries to your local model, heavy tasks to the collective pool. Everyone in the community gets AI without anyone paying for an API.
"Summarize this document" → routes to your local Phi-3.5 Mini on your Pi → response in seconds → never leaves your network
"Analyze all our meeting notes from Q1" → routes to Petals collective pool → runs across 6 nodes → delivered async → still never leaves the community mesh
⚠️ Before you open the gateway to your group: Run
openclaw doctor and fix all warnings. Review the
security checklist.
The January 2026 CVE-2026-25253 vulnerability came from gateways exposed on public IPs.
Tailscale IP only. Always.