OpenClaw: what it is, how to install it, and connect it to Telegram

OpenClaw showed up in my feed last winter, when the project changed its name twice in a week and gained hundreds of thousands of stars on GitHub. Since then, people have increasingly been searching for “openclaw installation” and “openclaw telegram.”

Below, I go through the official documentation: what this thing is, how it differs from Claude Code, how to install it on Windows, macOS, or Linux, how to choose a model, and how to connect a Telegram bot. I also cover security separately, because OpenClaw has been involved in a couple of high-profile incidents.

Learn how to work with the service in the Ailoot documentation.

What is OpenClaw in plain language?#

OpenClaw connects an AI model to your Telegram bot. With the relevant tools and permissions configured, it can summarize messages or schedule a reminder. The gateway process and its local configuration run on your computer or server. That does not make every conversation local: when you select a cloud model, the request and relevant conversation context are sent to that provider. If you connect through Ailoot, the request goes through Ailoot to the selected model provider. Local model inference requires a local model and no cloud fallback; Telegram messages still travel through Telegram, and enabled tools may contact other services.

The project is developed by the OpenClaw open-source community. The GitHub repository was created on November 24, 2025, and as of August 22, 2026, it has 387 thousand stars and 81 thousand forks—one of the fastest-growing open projects of the year. The language is TypeScript.

The project has a funny naming history, and it’s worth knowing so you don’t get confused by older articles. It started as Warelay, a gateway for WhatsApp. Then it was Clawdbot. In January 2026, Anthropic asked for a name change because of a trademark, and the project spent a couple of days as Moltbot. On January 30, 2026, it got its final name: OpenClaw. So if you come across instructions about “Clawdbot” or “Moltbot,” it’s the same project; those articles were just written before the renaming.

The project’s mascot is Molty the lobster, and “molting” (changing its shell) became the official metaphor for each renaming.

How does OpenClaw compare with the AI agent Hermes and Claude Code?#

I get asked about this difference all the time because, from the outside, they all look like “AI agents.” Here’s a comparison table:

How to read this. I use Claude Code for development: it’s strongest when you need a powerful model and work with code. OpenClaw and Hermes are for something else: an agent that runs all the time, hangs out in Telegram, remembers yesterday’s conversation, and gets things done on a schedule. The choice between OpenClaw and Hermes is subtler; there’s already a Hermes guide with installation and setup.

If you’re just choosing a tool, start with the overview: what people code with in 2026 and which tasks to delegate to an AI agent. I also have a separate article on installing Claude Code.

What do you need before installing OpenClaw?#

Check your Node version ahead of time—it’s the only strict requirement:

node --version

Versions 22.22.3+, 24.15+, and 25.9+ are supported; Node 26 is recommended. If the command returns “command not found” or the version is old, don’t worry: the official installer will download a recent Node version for you, so you don’t need to install anything manually.

The remaining question is the model. OpenClaw doesn’t think on its own: it calls the model you choose. Here are the cost options: an Anthropic, OpenAI, or Google key—you pay the provider for actual usage; or a local model through Ollama—free in terms of money, but demanding on hardware. I cover both options below in separate sections.

How do you install OpenClaw on macOS and Linux?#

The recommended route is the official installer:

1. Run the installer

   curl -fsSL https://openclaw.ai/install.sh | bash
   

   The script detects your OS, checks for Node (and installs it if needed), installs OpenClaw, and then opens the setup wizard. If you want to look around before running the wizard, add the flag: bash -s -- --no-onboard.

2. Check that everything installed

   openclaw --version
   openclaw doctor
   openclaw gateway status
   

   The first command shows the version, the second checks the configuration, and the third confirms that the gateway is running and listening on port 18789.

If you use npm and want to manage the installation yourself:

npm install -g openclaw@latest --allow-scripts=openclaw
openclaw onboard --install-daemon

There’s a recent pitfall here, covered in a separate chapter of the documentation. npm 12 blocks package installation scripts by default, and without the --allow-scripts=openclaw flag, installation will stop with an allowScripts message. On npm 11.15 and earlier, the flag isn’t needed—there’s no such restriction yet. Check your version with: npm --version.

The --install-daemon flag in the second command sets up autostart: on macOS, it’s a LaunchAgent; on Linux, a systemd service. It starts the gateway after a reboot or login, subject to the service manager’s conditions. It does not prevent sleep or keep the gateway running when the computer is off. Telegram replies require the host to stay awake and connected to the network; use an always-on server if you need replies while your laptop is asleep or off.

How do you install OpenClaw on Windows?#

The first option is for people who prefer windows to the terminal. Windows Hub is a native OpenClaw app for Windows 10 (20H2+) and Windows 11, currently in beta. Go to the Windows Hub releases page and download the installer for your processor: OpenClawCompanion-Setup-x64.exe for regular computers, or arm64 for ARM. After installation, OpenClaw Companion will appear in the Start menu, with a tray icon, chat window, and settings.

On first launch, the app will offer “Set up locally”—accept. It will create a dedicated WSL environment called OpenClawGateway, install the gateway inside it, and connect to it. It won’t touch your existing Ubuntu installation, if you have one.

The second option is through PowerShell:

iwr -useb https://openclaw.ai/install.ps1 | iex
openclaw gateway install

The first line installs OpenClaw; the second registers autostart through the Windows Task Scheduler. If your computer’s policy blocks task creation, OpenClaw will switch to a fallback on its own: a shortcut in the Startup folder, so the gateway starts when you sign in.

The third option is WSL2, which the documentation calls the most compatible runtime on Windows. Install Ubuntu, enable systemd, then follow the Linux instructions from the previous section:

wsl --install -d Ubuntu-24.04

Inside WSL, enable systemd in /etc/wsl.conf (the [boot] section, the systemd=true line), restart WSL by running wsl --shutdown from PowerShell, then install OpenClaw with the Linux command. If “distribution” and “systemd” mean nothing to you, use Windows Hub; it does all of this for you.

How do you choose a model and complete the initial setup?#

The model is the assistant’s engine. The setup wizard will offer a list of providers, and here are three practical options:

There’s a nice detail about Anthropic: if Claude Code is already installed on your machine and you’re signed in, OpenClaw can use that login. First, run claude auth login on the machine, then tell OpenClaw to use it:

openclaw models auth login --provider anthropic --method cli --set-default

If the gateway runs as a background service (which it does with --install-daemon), the key must be available to the service. To do that, put it in ~/.openclaw/.env—the wizard does this for you, so you don’t need to edit the file manually. After any changes, restart and check:

openclaw gateway restart
openclaw models status

If paying is difficult, a local model from the Ollama section is the way to go.

How do you connect OpenClaw to Telegram?#

Telegram is the fastest channel to connect: you only need a bot token, with no business account verification. The documentation calls it the easiest way to start messaging your assistant from your phone. Here’s the full process:

1. Create a bot with BotFather

   Open a Telegram chat with @BotFather (make sure it’s that exact username—there are plenty of fakes), send /newbot, and choose a name and username ending in bot. You’ll get a token like 123456789:ABCdef... in response—it’s the key to your bot, so keep it like a password.

2. Enter the token in OpenClaw

   The easiest way is through the wizard: run openclaw configure and select the Telegram channel. The manual alternative is the channels.telegram.botToken setting or the TELEGRAM_BOT_TOKEN environment variable. The documentation specifically says that the openclaw channels login telegram command doesn’t exist; Telegram connects only through a token.

3. Start the gateway and message the bot

   openclaw gateway
   

   Open your bot in Telegram and send it something. It will reply with an eight-character pairing code. This is a safeguard: messages won’t be processed until you approve the sender.

4. Approve yourself

   openclaw pairing list telegram
   openclaw pairing approve telegram CODE
   

   The code is valid for one hour, so don’t wait too long. The first approved sender automatically becomes the owner of privileged commands.

There are two tricky points that people often run into.

The first is groups. OpenClaw blocks all groups by default: you need to explicitly allow the bot in the channels.telegram.groups settings and leave requireMention: true mode enabled inside the group. Then it will reply only when called by username, and won’t jump into the general conversation. Telegram also has its own bot privacy mode: in a group, a bot sees only commands by default. Turn it off with /setprivacy through @BotFather. After changing it, you need to remove the bot from the group and add it back; otherwise, Telegram won’t apply the setting.

The second is not to open access to everyone. The settings include a dmPolicy: "open" mode with an asterisk in the access list: anyone who finds the bot by username can command it. For a personal assistant, that’s a direct path to giving someone else control of your machine. Leave pairing-by-code enabled or enter your numeric ID in allowFrom.

How do you run OpenClaw with a local model through Ollama?#

This option keeps model inference on your hardware when you select a local-only route without a cloud fallback, and it needs no paid model endpoint. It does not make the whole Telegram workflow local: messages still pass through Telegram, and enabled online tools can send data to other services.

1. Install Ollama

   Download the installer from ollama.com/download—there are versions for macOS, Windows, and Linux.

2. Download a model

   ollama pull gemma4
   

3. Connect Ollama to OpenClaw

   openclaw onboard
   

   Choose the Ollama provider and a mode: Local only (your own models), Cloud only (models from ollama.com), or Cloud + Local. The wizard will find the local server and show the installed models.

4. Check

   openclaw models list --provider ollama
   

Here’s the honest caveat, without which this section would read like an ad. OpenClaw’s documentation says directly that comfortable agent use with a local model requires hardware on the level of two fully loaded Mac Studios, while a 24 GB graphics card can handle only lightweight requests, with delays. Small and heavily compressed models truncate context and are less resistant to malicious instructions hidden in someone else’s text. In everyday terms: a local model on a good laptop is enough to “try it and understand what it’s about,” but a cloud model is still much more reliable for daily work with an assistant that uses tools.

One more warning from the documentation, for anyone putting together a WSL2 + Ollama + NVIDIA GPU setup: the official Linux Ollama installer enables autostart with a “always” restart policy, and on WSL2 with a GPU this can put the machine into a restart loop. The fix is to disable service autostart. If you run into this, see the troubleshooting section on the Ollama page in the documentation.

How safe is it to run OpenClaw?#

I’ll start with the main point, which not everyone understands. OpenClaw has access to your machine: files, commands, and browser. The documentation puts it plainly: the project is designed for one trusted user, and you shouldn’t give outsiders access to your bot. If someone can message your agent, they can use it to act on your computer.

The practical minimum:

1. Keep pairing by code enabled or use an access list—I wrote about setting this up in the Telegram section.
2. Don’t expose the gateway externally unless you need to. By default, it listens only on your machine; leave it that way.
3. Run the built-in check after any configuration changes:

openclaw security audit

It checks for open access, overly broad tool permissions, weak tokens, and exposed ports. With the --fix flag, some findings can be fixed automatically.

A separate note about scams—and this isn’t theoretical. When the project changed its name in a panic in January 2026, squatter bots took over the old addresses in seconds and immediately posted a crypto wallet; fake “OpenClaw developers” began promoting a token scam. That leads to two hard-and-fast rules. Download OpenClaw only from openclaw.ai or github.com/openclaw/openclaw. And the project has no cryptocurrency: any “OpenClaw token” is a scam. Don’t send money to it under any circumstances.

What can OpenClaw do after installation?#

Here’s what I’d try, in order:

1. Gateway and model. Run openclaw gateway status, then open openclaw dashboard. A loaded panel confirms that the Control UI is reachable. Send “Reply with: model connection test” in its chat and check that an AI reply arrives; loading the page alone does not test the model.
2. Telegram and the intended task. After configuring the bot and approving your own pairing, send a new message to that bot in Telegram and check that its reply arrives there. Then try one bounded task with a visible result, such as summarizing a harmless test folder you have allowed it to read. For a reminder, check the saved schedule and the delivered reminder separately. A dashboard reply does not establish Telegram delivery or scheduled execution.
3. Scheduled tasks. The assistant can run jobs on a schedule—a morning summary in Telegram, regular email checks. That’s why you keep the gateway running around the clock.
4. Memory. OpenClaw stores long-term memory in ordinary Markdown files in its working folder: who you are, how to respond to you, and what it has remembered from earlier conversations. You can open the files and read them yourself—there’s no black box.
5. Skills. The project has a catalog of ready-made skills and plugins with built-in security checks. Someone else’s skill is someone else’s code on your machine, which is why the catalog checks them.

There’s one rule here: don’t turn everything on at once. Start with one stable conversation, then Telegram, then schedules. That way, each failure shows up at its own step instead of turning into “everything broke, and I don’t know where.”

What should you do if installation goes wrong?#

Let’s go through three common scenarios.

“openclaw: command not found” right after installation. It’s almost always a path issue: the npm global packages folder isn’t in your terminal’s PATH. The documentation’s diagnostic commands are:

node -v
npm prefix -g
echo "$PATH"

The path from the second command should appear in the third. If it doesn’t, add it to your shell settings, or simply close and reopen the terminal: sometimes the Node path only appears in a new session after installation.

The gateway is running, but the assistant is silent. The symptom from the documentation FAQ: there’s no reply to the first message, and the token counter is at zero. That means the agent hasn’t even started—most often, the model isn’t configured. Check in this order:

openclaw gateway restart
openclaw status
openclaw models status

If a model is configured and there’s still no reply, openclaw logs --follow will show the live log and where things are getting stuck.

The Telegram bot isn’t replying. Check in this order: the token belongs to the bot you’re messaging; the gateway is running (openclaw gateway status); and your pairing has been approved (openclaw pairing list telegram). A pairing code is valid for one hour—if you messaged the bot yesterday, simply message it again and approve the new code.

When you’re completely stuck, the documentation suggests something unexpected but sensible: give the diagnostics to a local AI agent that can see your machine. Claude Code or Codex, given the output of openclaw status --all and openclaw doctor, can make sense of the configuration faster than support chat. Full circle—AI fixes AI.

What else do people ask about OpenClaw?#

Get to know Ailoot.