What is OpenClaw, really?

OpenClaw is an open-source AI agent runtime. It runs on your own machine — laptop, VPS, Raspberry Pi — and turns an LLM (Claude, GPT, a local model via Ollama) into a teammate that can actually do things, not just talk about them.

The shorthand: ChatGPT lives in a browser tab. OpenClaw lives on a computer you own. That changes what's possible. An OpenClaw agent has its own files, its own memory, its own scheduled jobs, and its own connections to the apps you already use — Slack, Telegram, Discord, WhatsApp, email. It can browse the web in a real Chrome browser, fill forms, read dashboards, and write back to you in the same channel where you talk to your coworkers.

Three layers. Gateway is the long-running process: receives messages from your channels (Slack, Telegram, etc.), schedules jobs, manages sessions. Agent is the model loop — given a turn, what should I do next, which tool, with what arguments. Skills + tools are the things the agent can call: read a file, run a command, browse a URL, send an email, post to a channel.

When you message the agent on Telegram, the gateway routes the message into a session. The agent reads the session history, loads the relevant memory files, decides what to do, calls skills as needed, and posts the reply back through the gateway. The whole loop is open-source — you can read every line of it on GitHub.

The honest answer: anything you can describe in writing that doesn't need a human in the loop. The most common things people run agents for in 2026:

• Inbox triage — read inbound mail, classify, draft replies for human review, auto-reply to obvious ones.
• Daily digests — every morning, summarize yesterday's Slack, GitHub activity, and metric dashboards into a single message.
• Research — given a topic, browse 20 sites, read them properly, return a structured brief with citations.
• Form filling — fill a 30-field application with data from your CRM, screenshot the result, save the URL.
• Personal assistant — schedule meetings, remember birthdays, draft messages, manage subscriptions.
• Always-on monitoring — watch a status page, ping you the moment something changes.

What it's not good at: anything requiring real-time physical world interaction, anything where the cost of being wrong is irreversible (think wire transfers), anything that needs sustained creative judgment over weeks. Hire it for execution, keep humans on strategy.

OpenClaw's biggest practical advantage over generic LLM wrappers is that you talk to it where you already are. No new interface to learn, no app to download. The supported channels: Telegram, Slack, Discord, WhatsApp, iMessage, email, and an embeddable web chat widget. One agent, multiple front doors, shared memory.

openclaw channel add whatsapp

Picking a channel matters more than picking a model. A team already on Slack should set up Slack first; a personal user with a phone in their pocket should set up Telegram or WhatsApp. The agent doesn't change.

A skill is a packaged capability — at minimum a folder with a SKILL.md file describing what the skill does, when to use it, and how. The agent reads that description to decide whether to invoke it.

ClawHub is the public registry — about 13,700 community-built skills as of mid-2026. Things like imap-smtp-email for email, playwright for advanced browser control, tavily-search for high-quality web search. Browse, install with one command, the agent picks them up.

Read the full breakdown on the OpenClaw skills page — what they are, how authoring works, and which ones are worth installing first.

The single biggest difference between an LLM and an agent is persistence. OpenClaw stores memory as plain Markdown files in your workspace. There's no hidden database — you can open the files in any editor.

• MEMORY.md is the durable long-term memory. Loaded at the start of every session.
• memory/YYYY-MM-DD.md is the daily note. Today's and yesterday's load automatically.
• A SQLite vector index (~/.openclaw/memory/{agentId}.sqlite) powers semantic recall via memory_search.
• Before context compaction, OpenClaw runs a "memory flush" — the agent saves anything important before its history gets summarized.

Full deep-dive on the OpenClaw memory guide.

The OpenClaw runtime is free. The bill comes from the LLM API calls. Below is a rough sense of where you'll land — drag the sliders to match your usage.

Real-world ranges from setups we've benchmarked: $10–30/mo for a personal assistant on Sonnet with reasonable caching, $30–80/mo for a browser-active agent doing real research, $300–600/mo for a 24/7 multi-agent setup running on Opus without optimization. The cost optimization guide walks the seven settings that drop most bills 60–90%.

The fastest path from zero to a working agent. Mac and Linux are similar; Windows needs WSL2 first.

Full step-by-step with platform-specific commands on the install OpenClaw guide.

Self-hosted OpenClaw is a great choice if you enjoy owning the stack — you control the model, the data, the uptime. Most individual users land here.

Where managed (Provision) wins: you don't want to babysit a VPS, you want a sandboxed browser with proxy rotation, you want email infrastructure that doesn't get marked as spam, or you want a team-friendly UI for managing multiple agents. We run OpenClaw for you, the agents land in the same Slack and Telegram, and there's nothing to update on your side.

See how Provision compares →