Knowledge Base - What is OpenClaw?

What is OpenClaw?

Published on March 24, 2026 by Pavel Admin

Last revised March 24, 2026

An introduction to the OpenClaw tool for running multiple AI agents - feeding and care thereof.

What is OpenClaw?

OpenClaw is an open‑source “doer” agent that (mostly) lives on your own infrastructure, connects to your chat apps, and can actually perform tasks and automations rather than just answer questions. It runs behind a gateway that can talk to platforms like WhatsApp, Telegram, Discord, Slack, and Signal, turning your messages into actionable workflows.

OpenClaw runs as an always‑on AI assistant that sits behind a “gateway” and listens to channels like WhatsApp, Telegram, Discord, Slack, Signal, and others. When you message it, the gateway routes your request to the right agent, waits for a response, and sends the reply back to the app you came from.

Instead of being just a chatbot, OpenClaw can trigger tools and workflows: searching the web, reading and writing files, calling APIs, posting messages, and more. It also keeps a workspace and session memory, so it can remember context over time, not just inside a single conversation.

At a high level:

The Gateway is the control plane that connects all your messaging platforms and routes messages.
Agents are AI “brains” with their own configuration, tools, and personality (“soul”).
Tools/skills are the specific actions agents can take: web search, file access, APIs, etc.

A simple example is a “research and report” workflow: you message OpenClaw on Telegram, the gateway passes the request to a main agent, which then delegates to sub‑agents to collect data, analyze it, and write a report before sending it back to you.

Where You Can Run OpenClaw?

OpenClaw is self‑hosted software, so you choose where it runs: on your own machine, on a Virtual Private Server (VPS), or via a fully managed service that gives you an OpenClaw instance.

Local install on your own machine

You can install OpenClaw on your laptop or desktop -directly or in a Docker container, as long as you meet basic hardware requirements (2+ CPU cores, 4+ GB RAM, modern Linux recommended but all major operating systems supported). This is usually done through a CLI workflow that walks you through configuring model providers, API keys, and integrations.

Typical characteristics:

Runs only when your machine is on and not sleeping.
Great for experimenting, development, and personal use where uptime is not critical.
Best on a separate machine if you allow it to touch files or run commands, to limit risk.

VPS (Virtual Private Server)

Many guides and vendors explicitly recommend a VPS as the “serious” option for OpenClaw. A VPS is a small always‑on server in the "cloud" that you manage yourself.

Key points:

Always online, no sleep, suitable for 24/7 automations and monitoring.
Typical baseline: Ubuntu 24.04 LTS, 2 vCPU, 4 GB RAM, 20–100 GB SSD.
Commonly deployed via Docker so all dependencies are packaged and updates are simpler.

Some hosting providers offer VPS images with OpenClaw pre‑installed, so you get a ready‑to‑use instance with a web dashboard and sane defaults out of the box.

Fully hosted / managed OpenClaw

While the core project is self‑hosted, a growing ecosystem of hosting companies sells fully configured OpenClaw instances: you buy a plan and get a dashboard and endpoint rather than touching the server yourself. Under the hood this is still OpenClaw on a VPS, but backups, monitoring, OS patches, and sometimes basic security hardening are handled for you.

For a semi‑technical user, this feels closer to a typical SaaS: you focus on configuring agents and tools, not Linux and Docker.

How Agents Use Models: Local vs API vs Mixed

OpenClaw separates the “agent” from the actual model, so you can point an agent at different backends: paid cloud APIs, local models, or a mixture. This is one of the main advantages of the platform: you are not tied to a particular provider or LLM model.

All paid models via API

Here, every agent uses cloud model providers such as OpenAI, Anthropic, or Google Gemini via API keys you configure.

Characteristics:

Strong reasoning and large context windows with frontier models.
Pay‑per‑use billing: you pay per token, often with better performance and reliability.
Data leaves your server to the provider, though many offer enterprise and “no‑training” modes (presumably keeping your privacy).

This is the simplest to set up and usually the most capable for complex automations.

All‑local models

You can instead run local models via Ollama or other OpenAI‑compatible local servers, so the agent sends requests to your own model running on local hardware.

Characteristics:

No per‑token cost, full data locality (everything stays on your machine or VPS).
Requires substantial hardware; community experience suggests that once you go beyond simple tasks you need 32B‑parameter‑class models and at least 24 GB of VRAM to match cloud agents for multi‑step workflows.
You won’t match frontier cloud models on the hardest reasoning tasks or very long contexts.

This option is attractive for stricter privacy and cost predictability if you have the hardware.

Mixed: local for routine, API for heavy tasks

Because OpenClaw supports multiple agents and model backends, many teams use a hybrid pattern: lightweight tasks go to fast, cheap local or smaller API models, while complex reasoning or high‑stakes tasks go to premium cloud models.

For example:

A “notification formatter” sub‑agent might use a small, cheap model.
A “strategy advisor” or “data analyst” sub‑agent might use a high‑end cloud model with better reasoning.

Main Agent, Sub‑Agents, and Evolution Over Time

OpenClaw is built around multi‑agent orchestration, where a primary agent coordinates a team of specialized sub‑agents.

Main agent vs sub‑agents

The main agent is the entry point that receives user messages from the gateway and is responsible for understanding the high‑level task. It often plays the role of “project manager”: it decomposes work, chooses tools, and delegates tasks to sub‑agents.

Sub‑agents are specialized workers configured in their own files with defined roles and skills. Examples include a web‑scraper, a data‑analyst, and a report‑writer, each using different models and tools but operating under one team configuration.

In OpenClaw’s documented depth model:

Depth 0: the main agent session (agent:<id>:main).
Depth 1: sub‑agents spawned by the main agent (they can sometimes spawn further sub‑agents if allowed).
Depth 2: sub‑sub‑agents (leaf workers) that cannot spawn new agents.

Results flow back “up the chain”: workers announce results to their parent, the parent synthesizes and returns to the level above, and eventually the main agent returns the final answer to the user.

How agents evolve over time

Even though each model call is stateless, OpenClaw gives agents a persistent workspace and session memory. Over time, this lets your setup evolve in several ways:

Configuration evolution: you refine configs for the main agent and sub‑agents as you see what works (e.g., tightening instructions, changing models, adding tools).
Skill/tool evolution: you plug in new tools or external APIs so agents can act on more systems: CRMs, calendars, lead‑gen tools, file stores, etc.
Memory and behavior evolution: as OpenClaw accumulates session history and workspace data, it can adapt responses and workflows to your patterns within the constraints of its design and your retention settings.

Practically, you can start with one generalist main agent and a couple of simple sub‑agents, then gradually split responsibilities (e.g., separate research, analysis, writing, ops) as your use cases grow.

Practical Privacy and Security Advice

Because OpenClaw can touch sensitive files, send messages, and call APIs, you need to treat it like any other powerful automation system.

Where to run it, from a risk perspective

Avoid running a powerful, tool‑enabled agent on your primary work laptop; if it can run shell commands or read arbitrary files, a misconfigured prompt could expose data or damage the system.
A VPS or a dedicated machine with a limited set of permissions is generally safer, following the principle of least privilege: give OpenClaw only the access it really needs (specific folders, specific APIs, specific accounts).
For fully hosted offerings, ask about data handling (logging, retention, encryption, geographic location) and whether they support “no‑training”/no‑log modes for model providers.

Data flow and model providers

When you use API‑based models, message content and sometimes tool outputs are sent to the model provider. To manage this:

Use providers and modes that commit not to train on your data when that matters, and be cautious about sending extremely sensitive content.
For regulated or confidential work, keep the most sensitive processing on local models and reserve cloud APIs for less sensitive tasks.

Operational hygiene

Put OpenClaw behind HTTPS (reverse proxy with TLS, such as Nginx), especially if it is reachable over the internet.
Keep your OS, Docker images, and dependencies updated, and rotate API keys periodically.
Start with conservative tool access (e.g., read‑only file tools, limited shell commands), then expand privileges as you gain confidence.

Practical starting advice

For a semi‑technical user who cares about privacy but does not want to run servers full time:

Start with a small VPS that runs OpenClaw via Docker, with API‑based models for convenience.
Configure one main agent and 2–3 sub‑agents (e.g., web‑scraper, analyst, writer) using different models as needed.
Limit tools at first (e.g., web search and a couple of API integrations) and avoid broad shell/file access until you are comfortable with how the agents behave.

That gives you an always‑on, reasonably secure, multi‑agent OpenClaw setup that can grow over time as you add more tools, refine agent roles, and adjust your balance of local versus cloud models.

What is OpenClaw?

What is OpenClaw?

Where You Can Run OpenClaw?

How Agents Use Models: Local vs API vs Mixed

Main Agent, Sub‑Agents, and Evolution Over Time

Practical Privacy and Security Advice

Practical starting advice

Further Reading

About this article