Build Your First AI Agent with OpenClaw

This guide explains the core concepts and what building an agent looks like. For exact CLI commands and API syntax, refer to the official OpenClaw documentation – it stays current as the framework evolves.

An OpenClaw agent consists of four pieces that snap together:

A minimal agent connects one channel (say, Telegram), adds a couple of skills, and uses local memory. A production agent like Amplify connects four channels, runs 32 skills, and uses Supabase-backed memory.

Channels

Channels are how your agent receives and sends messages. OpenClaw ships with plugins for Telegram, WhatsApp, Discord, and Slack. Each channel plugin handles the platform-specific protocol – your agent logic stays the same regardless of which messenger a user writes from.

When a user sends "What's the weather in Berlin?" via Telegram, the Telegram channel plugin receives the message, passes it to the agent core, and routes the response back to the same Telegram chat.

Skills

Skills are modular, sandboxed capabilities. A skill defines:

The LLM decides when to invoke a skill based on user intent. If someone asks about weather, the LLM recognizes the intent and calls the weather skill automatically.

Skills run in sandboxed environments for security – a skill can't access other users' data or interfere with other skills.

Memory

Memory persists across conversations. If a user says "My name is Alex and I live in London" today, the agent remembers tomorrow. Memory is cross-channel: context from a Telegram conversation is available when the same user messages via WhatsApp.

For development, you can use local file-based memory. For production, OpenClaw supports database-backed providers (Supabase, PostgreSQL).

LLM Routing

OpenClaw routes LLM requests through OpenRouter by default, giving access to 100+ models (OpenAI, Anthropic, Google, Mistral, and more) with a single API key. The framework supports dynamic model selection – choosing the best model for each task automatically. You can also connect your own API keys directly.

At a high level, getting a basic agent running involves:

The exact commands, configuration syntax, and project structure are in the OpenClaw documentation. We keep the docs authoritative rather than duplicating them here — the framework evolves and articles go stale.

To give you a sense of the skill model, here's the conceptual structure (not runnable code — see docs for exact syntax):

A weather skill would define:

When a user asks "What's the weather in Tokyo?", the agent's LLM recognizes the intent, calls the weather skill with location: "Tokyo", gets the result, and formats a human-readable response.

You can build skills for anything: calling external APIs, processing files, querying databases, generating media, or running computations. Each skill is isolated and modular – add or remove them without affecting the rest of the agent.

The framework itself is free (open-source). Your costs:

A personal agent serving one user typically costs $5-15/month in total infrastructure and API fees.

You could build a Telegram bot that calls OpenAI directly — many people do. OpenClaw gives you the layer above that:

If you're building a quick prototype, the raw API is fine. If you're building something you'll use daily – with memory, multiple channels, and growing capabilities – OpenClaw saves you from rebuilding infrastructure that's already solved.

Frequently Asked Questions