OpenClaw

Made by: Community Type: Open-Source Agent Harness / Framework Best for: Building custom agents with full control, multi-provider deployments, self-hosted setups

What it is

OpenClaw is an open-source agent harness — the layer between your AI model and the real world. It provides the scaffolding that turns a language model into a working agent: tool management, conversation history, memory, provider switching, and execution loop.

Think of it as the plumbing. You bring the model, you bring the business logic, OpenClaw handles the infrastructure that makes the agent actually run.

The provider landscape

One of OpenClaw's biggest strengths is 50+ provider integrations out of the box:

AI Models

• Anthropic (Claude family)
• OpenAI (GPT-4, o1/o3)
• Google (Gemini)
• Mistral, Cohere, and others
• Local models via Ollama

Channels (where agents receive and send messages)

• Telegram, Slack, Discord
• Email (IMAP/SMTP)
• WhatsApp
• HTTP webhooks
• CLI

Data & APIs

• PostgreSQL, MySQL, SQLite
• Redis
• REST APIs with auto-auth
• File system
• Web scraping

Architecture

OpenClaw is built around a plugin architecture. Every capability — a tool, a model provider, a communication channel — is a plugin. Plugins are loaded at runtime from a manifest file.

# openclaw.yaml
agent:
  name: my-sales-agent
  model: claude-sonnet-4-5
 
channels:
  - type: telegram
    bot_token: "${TELEGRAM_BOT_TOKEN}"
  - type: slack
    token: "${SLACK_BOT_TOKEN}"
 
tools:
  - type: web_search
  - type: database
    connection: "${DATABASE_URL}"
  - type: email
    smtp: "${SMTP_CONFIG}"
 
memory:
  type: sqlite
  path: ./memory.db

Multi-agent support

OpenClaw supports sub-agents — agents that can spawn and coordinate with other agents. This enables patterns like:

• A supervisor agent that breaks a task into subtasks and delegates each to a specialist
• A research agent and a writing agent working in sequence
• Parallel agents running the same task across different data sources

tools:
  - type: delegate
    agents:
      - research-agent
      - analysis-agent
      - report-writer

Memory system

OpenClaw ships with a built-in memory system using SQLite with FTS5 full-text search. Agents can:

• Store facts from conversations
• Retrieve relevant memories based on semantic similarity
• Build up knowledge over time without re-reading source documents

[Agent searches memory for "customer pricing objections"]
→ Finds 12 relevant past conversations
→ Uses them to inform the current response

Scheduling

Agents can be configured to run on a schedule without external cron infrastructure:

schedule:
  - cron: "0 8 * * 1-5"   # 8am weekdays
    task: "Check overnight alerts and send morning briefing"
  - cron: "*/30 * * * *"   # Every 30 minutes
    task: "Poll for new support tickets"

Deploying OpenClaw

OpenClaw is a single binary (cross-platform). Deployment is simple:

# Run directly
openclaw start --config openclaw.yaml
 
# Or via Docker
docker run -v $(pwd)/config:/config openclaw/openclaw start

For production, run it as a systemd service or a Kubernetes Deployment. Because it's a single binary with no external dependencies (SQLite is embedded), it's easy to manage.

When to choose OpenClaw

Getting started

# Install the binary
curl -fsSL https://openclaw.dev/install | bash
 
# Create a project
openclaw init my-agent
cd my-agent
 
# Edit openclaw.yaml, then run
openclaw start

Documentation and examples: openclaw.dev/docs (opens in a new tab)