The Ultimate Guide to OpenClaw

OpenClaw is an open-source AI agent framework that runs persistent AI agents with access to tools, memory, and multiple communication channels. It's the underlying framework that DenchClaw is built on — the "React" to DenchClaw's "Next.js."

This guide covers what OpenClaw is, how it works architecturally, and how to build applications on it.

What Is OpenClaw?#

OpenClaw is a runtime for AI agents. It provides:

• Session management: Persistent conversation sessions with history and context
• Channel routing: Connects agents to Telegram, WhatsApp, Discord, web chat, and dozens of other messaging platforms
• Tool execution: Agents can call tools (file I/O, web requests, code execution, database queries)
• Memory system: Persistent state across sessions via markdown files
• Multi-agent support: Spawn subagents for parallel work; results auto-announce back to parent
• Skills architecture: Capabilities defined in SKILL.md files, discoverable and composable

OpenClaw is MIT-licensed and available at github.com/openclaw/openclaw. It was created by Peter Steinberger, formerly of Clawdbot/Clawd.

OpenClaw Architecture#

The Gateway#

The OpenClaw Gateway is the core process. It:

• Manages sessions and message routing
• Handles channel integration (Telegram, WhatsApp, Discord, etc.)
• Executes tool calls
• Maintains session state

The Gateway runs as a background daemon:

openclaw gateway start
openclaw gateway status
openclaw gateway stop

Profiles#

OpenClaw runs as profiles — isolated configurations. Multiple profiles can run simultaneously on different ports.

openclaw gateway status  # shows active profiles

DenchClaw creates its own profile (dench) running on port 19001 when you run npx denchclaw.

Sessions#

A session is an active conversation context. Each session has:

• A unique session key
• Message history
• Active model configuration
• Memory scope (which files the agent reads)

Sessions can be:

• Main sessions (direct chat with a user)
• Subagent sessions (spawned by a parent agent for parallel work)
• ACP sessions (coding agents via the ACP protocol)

Channels#

Channels are how users communicate with agents. OpenClaw supports:

Each channel is a plugin. Adding a new channel requires creating a channel plugin, not modifying core.

Tools#

Tools are the actions the agent can take. OpenClaw provides:

• exec: Run shell commands
• read/write/edit: File operations
• web_fetch/web_search: Web access
• browser: Browser control
• canvas: UI rendering
• message: Send messages to channels
• sessions_spawn: Create subagents
• image/pdf: Analyze media

Custom tools can be added through the tool registry.

Skills#

Skills are SKILL.md files that describe capabilities to the agent. The agent reads installed skills to understand what it can do. Skills live in the workspace skills directory or global skills directory.

# List installed skills
ls ~/.openclaw-dench/workspace/skills/
 
# Install from clawhub
clawhub install apollo

Building on OpenClaw#

Creating a Basic OpenClaw Application#

1. Install OpenClaw:

npm install -g openclaw

2. Create a profile:

openclaw profile create myapp

3. Configure your workspace (AGENTS.md, SOUL.md):

# AGENTS.md
You are [agent name]. Your purpose is [description].

4. Start the gateway:

openclaw gateway start

5. Connect a channel (Telegram example):

• Create a bot via BotFather
• Add the token to your OpenClaw config
• Start chatting

Creating Custom Tools#

Tools are TypeScript functions registered with the gateway:

// my-tool.ts
export const myTool = {
  name: "my_tool",
  description: "Does something useful",
  parameters: {
    type: "object",
    properties: {
      input: { type: "string", description: "Input data" }
    },
    required: ["input"]
  },
  execute: async ({ input }) => {
    // Tool implementation
    return { result: `Processed: ${input}` };
  }
};

Creating Skills#

Skills are SKILL.md files:

# My Skill
 
## What This Does
Describe the capability.
 
## When to Use
Situations where this skill applies.
 
## How to Use
Instructions for the agent.
 
## Operations
Specific commands or API calls the agent should use.

Place in ~/.openclaw-[profile]/workspace/skills/my-skill/SKILL.md.

Building DenchClaw-style Applications#

DenchClaw is an opinionated application on top of OpenClaw. Building a similar application:

1. Choose a data model: DenchClaw uses DuckDB with EAV schema. You might use SQLite, a flat file, or a different schema.

2. Define your objects: What are the core entities in your domain?

3. Write the skill suite: Create SKILL.md files for each major capability area.

4. Configure the workspace: AGENTS.md, SOUL.md, USER.md for your agent's identity and context.

5. Build the UI: OpenClaw provides a web interface; you can extend it or replace it.

OpenClaw vs. Other Agent Frameworks#

OpenClaw vs. LangChain#

LangChain is a Python framework for building LLM-powered applications. More general-purpose, but:

• No built-in channel integration
• No persistent memory system
• More code required for basic agent functionality
• Better for programmatic ML pipelines

OpenClaw is better for: persistent personal/business assistants with messaging channel integration.

OpenClaw vs. AutoGPT#

AutoGPT is a general-purpose autonomous agent. Less production-ready, more experimental. OpenClaw is more stable and purpose-built for productive agent applications.

OpenClaw vs. Dify / Flowise#

Dify and Flowise are visual workflow builders for AI. Lower code floor but less flexible. OpenClaw requires more technical setup but is more powerful for complex agent behavior.

OpenClaw for Developers#

OpenClaw is most interesting for developers who want to build AI-native applications:

Use cases for building on OpenClaw:

• Personal productivity agents with domain-specific knowledge
• Team automation agents (stand-up bots, PR review, deployment monitoring)
• Domain-specific applications (research assistant, legal intake, customer support)
• Developer tools (code review agent, documentation agent, test generation)

Skills for the OpenClaw developer:

• Node.js experience (OpenClaw is Node.js-based)
• Familiarity with AI/LLM concepts
• Understanding of the filesystem-as-memory model
• Knowledge of the channel plugins you want to integrate

DenchClaw as an OpenClaw Profile#

DenchClaw installs as an OpenClaw profile with:

• Custom gateway config on port 19001
• DenchClaw-specific skills (crm, app-builder, browser, gstack)
• Pre-configured AGENTS.md with CRM agent identity
• DuckDB database initialized
• Web frontend at localhost:3100

This is the "Next.js" layer on top of OpenClaw's "React." All of DenchClaw's capabilities come from OpenClaw's primitives; DenchClaw is the opinionated configuration that makes them work for the CRM use case.

Frequently Asked Questions#

Is OpenClaw production-ready?#

OpenClaw powers DenchClaw which has real users. The core agent runtime is stable. The channel integrations vary in maturity. For personal productivity and small business use: yes. For enterprise deployments: evaluate carefully.

What models does OpenClaw support?#

Any model accessible via OpenAI-compatible API (OpenAI, Anthropic, Gemini, Ollama local models, and others). Configure the model in your profile settings.

Can OpenClaw run on a server, not just a local Mac?#

Yes. OpenClaw runs on Linux (common for server deployments), macOS, and Windows. Many users run it on a VPS or Raspberry Pi for always-on access.

What's the difference between OpenClaw and DenchClaw?#

OpenClaw is the framework (handles sessions, tools, channels, memory). DenchClaw is an application built on OpenClaw (adds CRM schema, DuckDB, app platform, web UI). You use OpenClaw to build applications; you use DenchClaw as your CRM.

Where can I get help building on OpenClaw?#

The OpenClaw Discord community (discord.com/invite/clawd), GitHub discussions on the OpenClaw repo, and the clawhub.ai community. DenchClaw documentation covers the DenchClaw layer.

Ready to try DenchClaw? Install in one command: npx denchclaw. Full setup guide →