OpenClaw: Deploy Your Personal AI Assistant on WhatsApp & Telegram

OpenClaw is an open-source framework for deploying personal AI assistants across messaging platforms. Unlike generic chatbots, OpenClaw creates a persistent, personality-driven AI that knows you, remembers context across conversations, and can execute real-world actions -- from browsing the web to managing your calendar.

It is built on a distributed architecture with a central gateway, multiple nodes (each running on different machines), and channels (WhatsApp, Telegram, Slack, etc.). This design means your AI assistant can run on your laptop, a VPS, or a Raspberry Pi, and you interact with it from any messaging app.

What makes OpenClaw unique is its personality system. Through AGENTS.md, SOUL.md, and MEMORY.md files, you define not just what your assistant can do, but who it is -- its communication style, values, knowledge areas, and behavioral boundaries.

The architecture supports multiple concurrent users, each with their own agent personality and memory. A single gateway can route messages to different nodes based on user identity, message type, or load conditions.

Nodes communicate with the gateway via WebSocket connections, enabling real-time message delivery and bidirectional streaming for voice interactions. The protocol is designed for unreliable networks -- messages are queued and retried if a node disconnects temporarily.

OpenClaw runs on Node.js 24 (recommended) or 22.16+ and requires a few environment variables for API keys and channel credentials. The installation process is streamlined for developers familiar with npm:

Configuration is split into two areas: infrastructure (gateway ports, node addresses, channel credentials) in .env and the environment config files, and personality/behavior in the Markdown files (AGENTS.md, SOUL.md, MEMORY.md).

The personality system is what transforms a generic AI into your personal assistant. Three Markdown files define the complete behavioral profile:

Channel setup involves configuring the messaging platform API and connecting it to your OpenClaw gateway. Each platform has different requirements and capabilities:

Both channels support text, voice messages (transcribed via Whisper or similar), images (analyzed via vision models), and file attachments. The channel adapter normalizes all message types into a unified format before passing them to the agent.

For WhatsApp, voice messages are particularly powerful -- you can speak naturally to your AI assistant while driving or walking, and it responds with synthesized voice or text based on your preference.

OpenClaw supports full voice interaction: speech-to-text for incoming voice messages and text-to-speech for responses. The voice pipeline integrates with multiple providers:

• Speech-to-Text -- OpenAI Whisper (local or API), Google Speech-to-Text, or Deepgram for real-time transcription
• Text-to-Speech -- Qwen3-TTS (local, voice profiles), LuxTTS (voice cloning, 150x realtime), OpenAI TTS, or Google TTS for response synthesis
• Voice mode -- Configure the agent to always respond with voice, only when receiving voice, or based on message length and context

Voice interactions feel natural because the agent maintains conversational context. You can start a voice thread about a topic, switch to text, and the agent seamlessly continues the conversation across modalities.

Skills are the building blocks of your agent's capabilities. Each skill is a self-contained module that the agent can invoke when needed. Skills can range from simple (weather lookup) to complex (multi-step browser automation with data extraction).

Creating a custom skill involves defining a tool schema (name, description, parameters) and an execution handler. Skills are registered in AGENTS.md and can declare dependencies on other skills or external APIs.

OpenClaw's cron system lets your agent perform tasks on a schedule without any user interaction. Cron jobs are defined in the configuration and can trigger any skill or sequence of skills:

• Daily briefings -- Morning summary of weather, calendar, news, and pending tasks sent to your WhatsApp
• Health reminders -- Supplement reminders, water intake alerts, workout prompts at configured times
• Monitoring checks -- Hourly server health reports, SSL certificate expiry warnings, uptime monitoring
• Data collection -- Scrape prices, track deliveries, monitor stock levels, and report changes

Cron jobs use standard cron syntax and can be conditional -- for example, only send the morning briefing on weekdays, or only alert about server issues if CPU exceeds 80%.

Browser automation gives your agent the ability to interact with any website as a user would. Built on Puppeteer, the browser skill can navigate pages, fill forms, click buttons, extract data, and take screenshots.

Common use cases include:

• Filling out repetitive forms (government portals, registrations)
• Extracting data from websites that lack APIs
• Monitoring price changes on e-commerce sites
• Taking screenshots for visual verification or reporting
• Automating multi-step web workflows (booking, ordering, submissions)

The agent can combine browser automation with its language understanding -- you describe what you need in natural language, and it figures out which buttons to click, which forms to fill, and how to navigate complex multi-page workflows.

The Canvas UI is an optional web interface that provides rich, visual interactions beyond what messaging apps can display. It connects to the same gateway and shares context with your messaging channels, so you can start a conversation on WhatsApp and continue exploring results on the canvas.

Key capabilities include:

• Interactive data tables and filterable lists for structured information
• Chart rendering for time-series data, health metrics, or financial tracking
• Form-based workflows for multi-step data entry that is awkward via chat
• Code editor panels with syntax highlighting for reviewing and editing files
• Screenshot previews and visual diff views from browser automation results

The canvas is served by the gateway on a configurable port and is protected by the same authentication layer. It is fully optional -- your agent works perfectly fine through messaging channels alone.

OpenClaw's hooks system provides event-driven extensibility. Hooks let you run custom logic before or after key events in the agent lifecycle, without modifying the core codebase. They are defined in the node configuration and can be written in JavaScript or as shell scripts.

Available hook events include:

• onMessageReceived -- Triggered before the agent processes a message. Use it for logging, content filtering, or routing overrides
• onResponseReady -- Triggered after the agent generates a response, before it is sent. Use it for output sanitization, translation, or analytics
• onSkillExecuted -- Triggered after a skill completes. Use it for audit trails, cost tracking, or chaining follow-up actions
• onNodeConnect / onNodeDisconnect -- Triggered when a node joins or leaves the gateway. Use it for alerting, failover logic, or resource rebalancing
• onCronTick -- Triggered on each cron evaluation cycle. Use it for dynamic scheduling or conditional job suppression

Hooks receive a context object with the event payload and can modify it in-place (for pre-hooks) or trigger side effects (for post-hooks). They run synchronously in the order they are declared.

All infrastructure configuration in OpenClaw is managed through environment variables, typically loaded from a .env file at the project root. This keeps secrets out of version control and makes deployments portable across machines.

Key environment variables include:

The .env.example file in the repository documents every variable with defaults and descriptions. Variables can also be set directly in the shell environment or through your deployment platform's secrets manager.

OpenClaw supports running multiple nodes connected to a single gateway. Each node is a separate machine (or process) that registers with the gateway using a pairing token. This lets you distribute workloads and capabilities across devices.

To pair a new node:

Each node declares its capabilities (e.g., browser automation, file system, audio processing) during registration. The gateway uses this information to route tasks intelligently -- a browser task goes to the node with Chromium installed, a file operation goes to the node with filesystem access.

Multi-device setups are useful for separating concerns: run your personal node on your laptop for local file access, and a cloud node for always-on availability and heavy compute tasks. Both nodes share the same agent identity and memory.

OpenClaw now supports 20+ messaging channels, making it one of the most versatile AI agent frameworks for multi-platform deployment. Beyond the original WhatsApp and Telegram, the channel adapter system now includes: Slack, Discord, Google Chat, Signal, iMessage, BlueBubbles, IRC, Microsoft Teams, Matrix, LINE, Mattermost, Nextcloud Talk, Nostr, Twitch, and WebChat. Each channel adapter normalizes platform-specific features (reactions, threads, file types, rich messages) into OpenClaw's unified message format.

OpenClaw supports multi-stage Docker builds with minimal runtime images. The build stage compiles TypeScript and installs dependencies, while the runtime stage uses a slim Node.js base image with only production dependencies. This reduces the final image size significantly, speeds up container startup, and minimizes the attack surface for production deployments. Docker Compose configurations are provided for running the gateway, node agent, and voice services as a coordinated stack.

OpenClaw supports configurable group broadcast dispatch with observer-session isolation and cross-account deduplication. When an agent is added to a group chat (Telegram, Discord, Slack), the broadcast system ensures that each message is processed exactly once, even when multiple agent instances observe the same group. Observer sessions are isolated -- each agent maintains its own context and memory for the group conversation without interfering with other agents.

Cross-account deduplication prevents duplicate responses when the same user interacts with the agent from different accounts or platforms. The broadcast dispatcher routes messages based on configurable rules: round-robin, capability-based, or affinity-based routing. This enables multi-agent setups where different agents specialize in different topics within the same group chat.

Jose runs OpenClaw 2026.4.14 on an MSI Raider 18 HX as a production AI assistant accessible via WhatsApp and Telegram. This is not a demo -- it is a fully autonomous system with 8 always-on services, 4 automated daemons, and 6 local AI models running on an RTX 5090. Here are the exact details: