10 min read

OpenClaw alternative: 7 options, from open source to ready-made

In short: people leave OpenClaw for three reasons: setup (command line, API keys, a service to keep running), security (exposed installations and malicious skills made headlines in 2026), and autonomy (an agent that acts on its own is not what everyone wants). If you want to stay open source, Hermes Agent is the closest replacement and NanoClaw or ZeroClaw are the safer, lighter ones. If you mainly want to message your computer from your phone, Claude Code Channels, Codex remote in the ChatGPT app or a remote MCP server do it without hosting an agent. Below: a comparison, how each one works, and how to run OpenClaw more safely on Windows, Telegram and WhatsApp if you decide to keep it.

Why people look for an OpenClaw alternative

OpenClaw is a self-hosted personal agent: you install it on a computer or a server, connect one or more AI models with your API keys, and talk to it from messaging apps. It runs as a background service, remembers past conversations, and can be extended with skills. It became one of the most popular open-source projects of 2026, and it is genuinely capable. These are the reasons that push people elsewhere:

The alternatives compared

For every way of letting AI use a computer or a phone, not only the OpenClaw-like ones, see computer use agents: the complete guide.

Open-source alternatives

Hermes Agent

Built by Nous Research and released under the MIT license, Hermes Agent is the closest open-source rival: messaging channels, scheduled jobs, persistent memory and support for a very wide range of models. Its distinctive feature is that it turns workflows you repeat into reusable skills. It also includes a migration command that imports an existing OpenClaw setup. Watch out for: it is a large install and, unless you choose one of its sandboxed backends, it runs commands directly on the host. Pick it if you like OpenClaw and want the same power with a clearer security model.

NanoClaw

Born as a direct response to OpenClaw's security problems: it is small enough to read, and every agent session runs in its own container, so a bad skill or a manipulated prompt cannot reach the rest of the machine. Watch out for: it is built around Claude and needs Docker. Pick it if isolation matters more to you than features.

ZeroClaw

A single compiled program with a small memory footprint and approvals for actions turned on by default. Pick it if you want a light, always-on agent on a small server.

Official features that cover the same need

Many people install OpenClaw for one thing: sending a message from the phone and having the computer at home do the work. Today the AI companies offer that directly.

Claude Code Channels

A Claude Code feature that lets you message a running session through a Telegram or Discord bot: Claude does the work on your computer and replies in the same chat. Access is locked with a pairing code and an allowlist of your accounts. Watch out for: it is a research preview built for developers, and the session must be running. Close to the “OpenClaw on Telegram” experience, without hosting a separate agent.

Codex remote in the ChatGPT app

The ChatGPT phone app pairs with your Windows PC or Mac with a QR code: you start a task, approve the commands it wants to run and check the result, while computer use lets it operate the programs. Watch out for: on Windows it takes over the screen and the PC must stay unlocked. More in control your PC from your phone with AI.

Perplexity Personal Computer

The closest commercial product to OpenClaw's always-on idea: the Perplexity Mac app works with local files and apps and takes tasks from the phone, ideally on a Mac mini left running. It also runs on Windows, but control of installed apps is Mac-only; it needs a Pro, Max or Enterprise plan. More in Perplexity Personal Computer on Windows.

A ready-made option for Windows: MCP DeskManager

DeskManager is our product, and it is not an agent: it makes your Windows PC a tool of the Claude or ChatGPT chat you already use. You install it, paste one address into your assistant as a connector, and from then on any chat, the phone apps included, can use the PC's terminal, files, your real Chrome with your logins and programs without an API. There are no API keys to store on the computer, no ports to open and no bot to host; it acts only when you ask, and irreversible actions need your confirmation.

It fits if you wanted OpenClaw to reach your Windows PC from your phone but do not want an autonomous agent with your keys running unattended. It does not fit if you want jobs that start on their own at night, a model other than Claude or ChatGPT, or a Mac (the macOS version is in development).

OpenClaw on Windows

OpenClaw now supports Windows natively for the command line and the gateway, but its documentation still recommends WSL2 (Linux inside Windows) for the full experience. In practice, on WSL2:

• install WSL2 with Ubuntu from an administrator PowerShell, and turn on systemd inside it;
• install OpenClaw with the Linux installer inside WSL, then run the onboarding to set up the gateway as a service;
• WSL does not start by itself when Windows boots: if the agent must always be there, you need a scheduled task that starts it;
• keep the project files inside the Linux file system for speed, and keep the gateway on localhost.

If what you want on Windows is mainly to use the PC's programs and files from a chat, a tool that installs as a normal Windows program saves most of these steps: see MCP server on Windows: every option compared.

OpenClaw on Telegram and WhatsApp

Telegram is the easiest channel: you create a bot with BotFather, paste its token into OpenClaw and restrict who may write to it. WhatsApp takes more steps: OpenClaw links to your account like WhatsApp Web, by scanning a QR code, through an unofficial library rather than WhatsApp's business API, and the phone must come online from time to time to keep the link alive. Whichever channel you use, limit the allowed senders to yourself: anyone who can write to the bot can give orders to the agent.

If chatting from the phone is the main point, remember that the Claude and ChatGPT apps can already reach your computer through Claude Code Channels, Codex remote or a connector, without a bot in between.

If you keep OpenClaw: run it more safely

In early 2026 internet scans found more than 21,000 OpenClaw gateways reachable from the open internet, a serious vulnerability was disclosed and patched within days, and a security firm identified 341 malicious skills in the ClawHub marketplace, most from a single campaign. The project has shipped many fixes since; what remains is mostly up to how you set it up:

• never expose the gateway to the internet: keep it on localhost and reach it through SSH or an authenticated tunnel;
• update often, and check the version you run against the security advisories;
• treat every skill as untrusted code: install few, read them, and never paste commands a skill page asks you to run;
• restrict channels to your own accounts with an allowlist;
• give it separate accounts and low-value keys, with spending limits on the model provider;
• keep it away from systems with sensitive or regulated data: the Dutch data protection authority, among others, has advised organizations against it.

How to choose

• I like OpenClaw, I want something more solid → Hermes Agent.
• Security first → NanoClaw (containers) or ZeroClaw (approvals on).
• I just want to text my computer from the phone, and I code → Claude Code Channels.
• I use ChatGPT → Codex remote with computer use.
• I want an always-on assistant and have a spare Mac → Perplexity Personal Computer.
• I want my Windows PC from any Claude or ChatGPT chat, with no agent running on its own → MCP DeskManager.

Frequently asked questions

What is the best OpenClaw alternative?

It depends on why you are leaving. For the same features in open source, Hermes Agent; for isolation, NanoClaw; to message your computer from the phone without hosting an agent, Claude Code Channels or Codex remote; for Windows from any Claude or ChatGPT chat, a remote MCP server.

Is OpenClaw free?

The software is free and open source; you pay the AI model's usage through your API keys, and an always-on agent can use a lot of it.

Does OpenClaw work on Windows?

Yes: natively for the command line and the gateway, with some limits, and fully through WSL2, which its documentation recommends.

Is OpenClaw safe?

It can be run safely, but it requires care: keep the gateway off the internet, update it, install only skills you have checked, and restrict who can message it.

Can I use OpenClaw with WhatsApp?

Yes, by linking it to your account like WhatsApp Web through a QR code. It is not an official WhatsApp integration, so limit the allowed senders and keep the phone online now and then.

Further reading

• Computer use agents: every way to let AI control your PC, Mac or phone
• Control your PC from your phone with AI: 4 ways
• MCP server: what it is, how it works and how to use it