Search for openclaw install windows or openclaw install mac and the first page is vendors: Kimi's guide ends at its own cloud and still says Node 22.14, Parallels wants you in a VM, hosting blogs want you on their server. The official answer is spread over five docs pages. This one puts it in order: what the two installer scripts do step by step, the requirements that changed in September, the errors people hit with their real text, and a checklist for the end. Everything about OpenClaw is from the OpenClaw docs and the scripts themselves, read on 20 October 2026 (release 2026.10.5).

Requirements, as of October 2026

• Node 24.16 or newer, or 26.1 or newer. Node 22, 23 and 25 are unsupported since 2026.9.3, because their node:sqlite silently truncates text at an embedded NUL (Node.js compatibility). The installer adds Node when it is missing: 26 via Homebrew on a Mac, the LTS line via winget on Windows.
• Windows: PowerShell 5 or newer for the script; Windows 10 20H2 or 11 for the Windows Hub app. WSL2 is optional, not required (Windows).
• Mac: the CLI runs where Node runs, macOS 13.5 or newer; the menu bar app needs macOS 26.2 Tahoe (macOS app).
• Git, even for an npm install: the installer checks for it so dependencies from git URLs do not fail with spawn git ENOENT.
• A model: a provider API key, a Claude Code or Codex login, or Ollama. Every route, including the free ones: OpenClaw without an API key.

Which path

On Windows the docs recommend the Windows Hub app (OpenClawCompanion-Setup-x64.exe, 2026.9.8-1 on 7 October): no administrator rights, and Set up locally creates its own WSL distro, OpenClawGateway. The PowerShell script gives a native CLI and a gateway kept alive by a Scheduled Task; WSL2 runs the Mac/Linux script. On a Mac: the menu bar app (2026.10.5 ships Apple Silicon, Intel and Universal DMGs) or the curl script, which installs a LaunchAgent. This page follows the scripts.

What install.sh does on a Mac

About five minutes by the getting started page. The flow, from installer internals and the script:

1. Run it in Terminal:

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

2. Node. The script checks the version and the linked SQLite. If Node is missing it installs Homebrew first (that needs an Administrator account) and then brew install node, which is Node 26. An nvm install is reused when compatible; with nvm present but no compatible version it offers nvm install 26 and never installs a second nvm.

3. Git via Homebrew if missing, then npm install -g openclaw@latest, with --allow-scripts=openclaw added on npm 11.16 and newer. A failed build pulls in Xcode Command Line Tools and retries.

4. Onboarding. On a fresh install it hands over to openclaw onboard; Quick start finds a Claude Code or Codex login or a key, checks it with a real completion and opens the dashboard. On an existing install it runs openclaw doctor --fix and restarts the gateway service instead.

5. Keep it running. Quick start leaves the gateway in the foreground. Press Ctrl+C, then:

   openclaw gateway install
   openclaw gateway status
   

   The first installs a LaunchAgent; the second should show the gateway listening on port 18789.

What install.ps1 does on Windows

1. Run it in PowerShell (no execution-policy change needed for this form):

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

2. Node. If missing: winget, then Chocolatey, then Scoop. If all three are absent or leave an unsupported version, the script downloads the official Node 26 zip into %LOCALAPPDATA%\OpenClaw\deps\portable-node and adds it to your user PATH.

3. OpenClaw by npm install -g, launched from a temp directory so a shell opened in C:\ still works. If Git is missing on the git install method, a user-local MinGit goes under %LOCALAPPDATA%\OpenClaw\deps\portable-git.

4. PATH and doctor. The bin directory goes on your user PATH; on upgrades it runs openclaw doctor --fix --non-interactive. A failure stops with a terminating error and leaves the window open.

5. Onboarding, as on the Mac. Then the service:

   openclaw gateway install
   openclaw gateway status --json
   

   This registers a Scheduled Task, "OpenClaw Gateway", that runs as you before login without storing a password. If task creation is denied you get a Startup-folder item instead, which only starts after you log in; run the install from an elevated terminal to get the boot task.

Where it breaks

The error text as it appears, why, and the fix, with the source.

• openclaw is not recognized on Windows, zsh: command not found: openclaw on a Mac. npm's global bin directory is not on PATH, or the terminal predates the install. Reopen the terminal; still missing: npm config get prefix and add that directory to the user PATH on Windows (no \bin), on a Mac export PATH="$(npm prefix -g)/bin:$PATH" in ~/.zshrc (docs).
• npm error spawn git ENOENT on Windows. Git is missing, so a dependency from a git URL cannot be cloned. Rerun the script so it bootstraps MinGit, or install Git for Windows and reopen PowerShell (docs). January's node.exe : npm error code ENOENT was the installer calling npm through node.exe (#3917, #11049); gone since.
• Error: Could not find a package manager (winget, choco, or scoop). Windows without winget, and the window closed before you could read it (#38054, March). Fixed since: the script downloads portable Node, and the piped form keeps the window open. If it recurs, install winget from the Store or Node from nodejs.org.
• Node 22.22.0: node:sqlite truncates TEXT at embedded NUL (nodejs/node#61954); use 24.16+/26.1+ or a build with the fix, with a v before the number on Windows. Node below the September floor. Let the installer provision Node, or answer Y to the CLI's Update NodeJS: Y/N [N]: prompt; a private runtime lands in ~/.openclaw/tools/cli-node.
• blocked because they are not covered by allowScripts. A manual npm install -g openclaw on npm 12. Add --allow-scripts=openclaw; npm approve-scripts does not work for global installs (docs).
• npm.ps1 cannot be loaded because running scripts is disabled on this system, same for openclaw.ps1. Running npm or openclaw by hand under the Restricted policy Windows client editions ship with; npm's .ps1 shims are scripts, the piped installer is not. Set-ExecutionPolicy RemoteSigned -Scope CurrentUser, or call npm.cmd and openclaw.cmd (Microsoft).
• You have not agreed to the Xcode license agreements, or the "install the command line developer tools" dialog on a Mac. npm needs git; without Command Line Tools, Apple's /usr/bin/git shim opens the dialog, and the gateway runs git at every start even with update checks off (#31527, #158099, open). xcode-select --install, finish the dialog, rerun; sudo xcodebuild -license only if full Xcode is installed.

Two more, after the install rather than during it. On Windows a Scheduled Task that reports success proves nothing, "A successful task result alone does not prove the Gateway is healthy": read openclaw gateway status --json and the gateway log, as the 2026.9.1 update taught (#137813). On a Mac an open issue has the LaunchAgent reach ready, get SIGTERM a second later and loop, while the same build is stable in the foreground (#144094, 2026.8.2 and 9.3); openclaw gateway run tells the two apart. On 2026.9.4 on Windows, read the docs' note on updating before openclaw update.

After the install: a checklist

1. openclaw --version prints 2026.10.5 or whatever you asked for.
2. openclaw doctor has nothing red; openclaw doctor --fix repairs what it can.
3. openclaw gateway status shows the gateway on port 18789, from a new terminal.
4. openclaw dashboard opens the Control UI and a typed message gets a reply.
5. The gateway is bound to 127.0.0.1, which is the default: do not forward 18789 on your router (security).
6. A first channel: OpenClaw in Telegram in five minutes. An unknown sender gets a pairing code until you approve them.
7. The machine stays on. Sleep, a closed lid and OS updates stop the gateway; the upstream FAQ says to use a VPS after disconnects (providers and hosting).
8. When something is off, openclaw triage runs read-only checks and writes a diagnosis for Claude Code, Codex or the built-in agent.

Without installing anything

Aplite runs a standard OpenClaw instance for you: one gateway in its own container on our server in Nuremberg, currently 2026.9.8, set up from the browser.

1. Create an account with email or Google; no card on the Free plan.
2. Pick an agent and press Deploy; the workspace opens about a minute later.
3. Ask. There is no Node, no PATH, no key to paste: models come through one metered account, billed in credits, 1 credit = 1¢, token usage under each reply.

What a managed instance gives up, the shell, raw openclaw.json, local models and day-one releases, is in Managed OpenClaw vs self-hosting; other hosts are in Managed OpenClaw hosting compared. Aplite is independent and not affiliated with the OpenClaw project.

FAQ

Does OpenClaw run on Windows?

Yes, three ways: the Windows Hub desktop app, which creates its own WSL distro; the PowerShell installer, a native CLI and a gateway kept alive by a Scheduled Task; or WSL2 with the Linux script. The docs recommend the Hub for desktop use.

Do I need WSL2 to install OpenClaw on Windows?

No. The PowerShell script installs a native gateway. WSL2 is the most Linux-compatible option and what the Hub uses underneath, but it is not required.

Which Node version does OpenClaw need?

Node 24.16 or newer, or 26.1 or newer; Node 26 is recommended. Node 22, 23 and 25 are unsupported since release 2026.9.3. The installer provisions a supported Node when yours is missing or too old.

Does OpenClaw work on an Intel Mac or an older macOS?

The CLI and gateway run wherever Node 24 or 26 runs, which Node builds for macOS 13.5 and newer, on Intel and Apple Silicon. The menu bar app needs macOS 26.2 Tahoe; on an older Mac use the script.

Why does PowerShell say openclaw is not recognized after the install?

The global npm bin directory is not on the PATH of that window. Open a new PowerShell first; if it still fails, run npm config get prefix and add that directory to your user PATH, then reopen.

Can I try OpenClaw without installing it?

Yes. A managed host runs the gateway for you; on Aplite that is an account, Deploy and a chat, on the Free plan without a card. The trade-offs are above.

Search for what is openclaw and the first page answers three different questions. Wikipedia tells the story: an Austrian developer, four names in two months, a star count. The official docs explain the architecture: trust boundaries, runtimes, a WebSocket protocol. Hosting companies explain it on a page that ends at their control panel. None of them says in plain words what the thing is, what the five words in every tutorial mean, and what it will and will not do. This page does, with the terms as the OpenClaw docs use them, read on 20 October 2026.

The short version

OpenClaw is an open-source program that you run on a computer you control, and that turns an AI model into an assistant you message from the chat apps you already use. The docs' own line: "an open-source AI assistant that runs on your own hardware and meets you in every chat app you already use."

Open source: the code is on GitHub under the MIT license, about 393,000 stars on 20 October 2026, stewarded by the OpenClaw Foundation, a US non-profit that sells nothing ("No subscription. No hosted tier. No token." on openclaw.ai). You run it: on a laptop, a Mac mini or a rented server, and only while that machine is on. An AI model: OpenClaw contains none. It connects to one, from Anthropic, OpenAI, Google and others through a key or a subscription login, or a local model through Ollama. The model thinks; OpenClaw is everything around it: the chat connections, the memory on disk, the tools, the schedule. The current release is 2026.10.5.

The five words

Gateway. The one process that runs all the time: it holds the connections to your chat apps, serves the browser dashboard, runs the scheduler and talks to the model, on port 18789 of the machine itself by default (architecture). When it is down, nothing answers; a message sent while your laptop sleeps waits until it wakes. That is why "where does the gateway run" is the first practical question.

Agent. A configured assistant with its own folder, default model and history. The folder, the workspace, holds plain markdown you can edit: AGENTS.md for operating instructions, SOUL.md for persona, USER.md for who you are, MEMORY.md for long-term memory (agent). A fresh install has one agent, main; more can be added, with different chats routed to each (multi-agent).

Session. A conversation. By default every direct message from every channel lands in one rolling conversation, the main session, so what you said in Telegram is known when you continue in the browser; groups get their own, and scheduled jobs start fresh ones (sessions).

Skills. Instructions for a specific job: a folder with a SKILL.md that names the skill and says when and how to use which tools (skills). Some ship with OpenClaw, some you write, some you install from ClawHub, the public registry, with openclaw skills install @owner/<slug>; the docs call a third-party skill untrusted code, to be read first.

Channels. Where messages come in and go out. Telegram and the browser's WebChat are in the core install; Discord, Slack, WhatsApp, Signal, iMessage, Matrix, Microsoft Teams, Google Chat and a dozen more are official plugins (channels). An unknown sender gets a pairing code instead of an answer until you approve them.

Four more, briefly: tools, what the agent can do (read, write, exec, browser, web_search; tools); memory, MEMORY.md plus dated notes in memory/, so it remembers only what it wrote down (memory); cron jobs, a prompt on a schedule delivered to a channel (automations); the heartbeat, a check every 30 minutes by default, silent when nothing needs attention (heartbeat).

What it does, and what it does not

It answers in your chat apps; reads and writes files in its workspace; runs commands on its machine; opens a browser; searches the web; remembers across days in files you can read; messages you first, on a schedule or when a heartbeat finds something; runs several agents and sub-agents. The docs' own examples: inbox and calendar briefings, research, reminders, browser automation, a task sent from your phone (FAQ).

What it does not do, from the same sources:

• It has no model of its own. You bring a provider key, a ChatGPT or Claude subscription login, or a local model; the model is where the bill comes from. Every path, including the free ones: OpenClaw without an API key.
• It does not run while its machine is off. Sleep, network drops and OS updates stop the gateway; the upstream FAQ says to use a VPS after disconnects (providers and hosting). The free ways of keeping it on: OpenClaw hosting for free.
• It is not locked down by default. Sandboxing and exec approvals are off until you turn them on, and one gateway is one trust domain: "a trusted single-operator assistant" (why OpenClaw, security).
• It is not a coding IDE. The docs say Claude Code or Codex is better for fast coding inside a repository; OpenClaw is the assistant around them, and can run both as its agent runtime (agent runtimes).
• It is not a hosted service. Hosting is what third parties sell, seventeen of them in October 2026; the hosted path below is one.

How it differs from a chat with a model

ChatGPT, Claude.ai and Gemini are a model behind a website: you go to it, the conversation lives there, and it touches what the vendor connected. OpenClaw inverts that: the assistant comes to the messenger you already have open, its memory is files on your disk, its hands are tools on your machine, and it acts on a schedule without being asked, which no chat tab does. The model is a setting: /new <model> switches it mid-conversation, and an agent can fail over to a second provider. The price is that you operate it: machine, keys, updates and security are yours.

Beside it in the same search: Hermes Agent by Nous Research, the same shape, self-hosted and messaging-first; and Claude Code and Codex, coding harnesses OpenClaw can drive rather than replace.

How to try it

Install it

About five minutes, by the upstream getting started page.

1. Have Node 24.16 or newer, or 26.1 or newer (the installer adds it when missing) and a model: a provider key, a Claude Code or Codex login, or Ollama.

2. Install. macOS, Linux or WSL2:

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

   Windows, in PowerShell:

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

   Desktop apps: install page.

3. Onboard. The installer opens the wizard; Quick start finds your login or key, verifies it with a real completion and opens the dashboard.

4. Keep it running. Press Ctrl+C in the foreground gateway, install it as a service, then check it:

   openclaw gateway install
   openclaw gateway status
   

   The second command should show the gateway listening on port 18789.

5. Talk to it. openclaw dashboard opens the Control UI; type a message. From your phone, Telegram is the fastest channel: OpenClaw in Telegram in five minutes.

If something fails, openclaw doctor reads the findings.

Without installing anything

Aplite runs a standard OpenClaw instance for you: one gateway in its own container on our server, currently 2026.9.8, set up from the browser. Create an account with email or Google, no card on the Free plan; pick an agent, press Deploy; about a minute later the workspace opens and you ask. No key to paste: models come through one metered account, billed in credits, 1 credit = 1¢, token usage under each reply. On top sit a markdown workspace the agent reads and writes, an approval queue for its changes, and Telegram from a bot token. What a managed instance gives up, the shell, raw openclaw.json, local models and day-one releases, is on the OpenClaw page; other hosts are in Managed OpenClaw hosting compared. Aplite is independent and not affiliated with the OpenClaw project.

FAQ

What is OpenClaw?

An open-source AI assistant you run on your own computer or server: it connects a model of your choice to your chat apps, gives it tools such as files, shell, browser and web search, keeps its memory in markdown files, and can act on a schedule.

Is OpenClaw free?

The software is free and the project sells nothing. Running it costs a machine that stays on and the model, billed per token unless it runs locally (OpenClaw hosting for free).

Do I need an API key to use OpenClaw?

Self-hosted, yes, or a substitute: a ChatGPT or Claude subscription login that onboarding reuses, or a local model through Ollama. On a managed host that holds the model account, Aplite among them, there is no key.

How does OpenClaw work?

The gateway stays running. A message from a chat app goes to an agent with its instructions, memory and skills; the agent calls the model, runs the tools the answer needs, and the reply returns through the same channel. Scheduled jobs and the heartbeat start the same loop without a message.

Who makes OpenClaw, and is it owned by OpenAI?

Peter Steinberger published it in late 2025 as Clawdbot, renamed Moltbot and then OpenClaw in January 2026 (Wikipedia); the OpenClaw Foundation, an independent non-profit funded by donations, stewards it since he joined OpenAI in February 2026. OpenAI is one donor; the docs say donors do not control the project and no lab's model is privileged in the code.

Search for openclaw telegram setup and you get a dozen walkthroughs by hosting companies, each ending at their own control panel, and the upstream documentation, which is correct but spread over four pages: setup, pairing, access control, troubleshooting. The actual job is short. A Telegram bot is a token from @BotFather; OpenClaw needs that token and then an approval for every chat that may talk to it. That is the whole thing, and it takes about five minutes whether the gateway runs on your machine or on someone else's.

Below, both paths on OpenClaw 2026.9.8: the self-hosted one with every command, checked against the upstream docs and run in a fresh container of that version, then the hosted one, where the commands become two buttons.

What you need

• A bot token. Free, from @BotFather inside Telegram. Check the handle is exactly @BotFather; there are impostors.
• A gateway that stays up. The Telegram channel long-polls Telegram's API from the gateway, so a bot only answers while the gateway runs. A laptop that sleeps stops it.
• A model. The bot token is not a model key. Self-hosted, you still need a provider or a local model (the ways without an API key); hosted, the models are included.

Self-hosted: six steps

1. Create the bot. In Telegram, open @BotFather, send /newbot, answer with a display name and a username ending in bot. BotFather replies with the token, shaped like 123456789:AAF…. Keep it private: whoever has it can run your bot until you regenerate it with /revoke.

2. Install OpenClaw if it is not there yet. On macOS, Linux or WSL2:

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

   Onboarding picks the model provider and starts the gateway; the longer version is in OpenClaw hosting for free.

3. Give OpenClaw the token. One command:

   openclaw channels add --channel telegram --token 123456789:AAF…
   

   It prints Added Telegram account "default". and writes the channel into ~/.openclaw/openclaw.json:

   "channels": {
     "telegram": {
       "enabled": true,
       "botToken": "123456789:AAF…"
     }
   }
   

   The same block typed by hand works too, as does the TELEGRAM_BOT_TOKEN environment variable for the default account; a tokenFile beats botToken, which beats the env. There is no openclaw channels login telegram step: unlike WhatsApp with its QR login, Telegram needs nothing but the token (setup docs).

4. Make sure the gateway is running. A running gateway hot-reloads the new channel. If it is not running, start it in the foreground with openclaw gateway, or as the service that openclaw gateway install set up (gateway docs):

   openclaw gateway start
   

   Then check the channel:

   openclaw channels status --probe
   

   The Telegram line reads Telegram default: enabled, configured, mode:polling, token:***; with --probe the gateway also calls Telegram with the token. A wrong token fails at gateway startup as getMe returned 401.

5. Send /start to your bot. A Telegram bot cannot write to you first. On the first message from an unknown chat, OpenClaw's default DM policy, pairing, does not answer the question; it replies with a pairing code instead: eight characters, uppercase, no 0, O, 1 or I, valid for one hour, at most three pending per bot (pairing docs).

6. Approve the chat. On the server:

   openclaw pairing list telegram
   openclaw pairing approve telegram K7QX2MPB
   

   Add --notify and the bot tells the requester they are in. From now on that chat talks to the agent; the approval lives in ~/.openclaw/state/openclaw.sqlite, and the first approved sender becomes the command owner if none was set (CLI reference). Anyone else who finds the bot gets a code and silence until you approve them, which is the point.

Two things people do next, both optional. Lock it to yourself: set dmPolicy to allowlist and put your numeric Telegram user ID in allowFrom. The ID is in the bot's pairing reply, in openclaw logs --follow as senderUserId, or at https://api.telegram.org/bot<token>/getUpdates; @username entries are not accepted (access control). Add it to a group: groups default to groupPolicy: "allowlist", so a group is silent until its chat ID (a -100… number for supergroups) is listed under channels.telegram.groups; the bot then answers only when mentioned unless requireMention is off, and reading unmentioned messages needs Privacy Mode disabled in BotFather (/setprivacy) or admin rights, then a remove and re-add. /whoami@your_bot in the group confirms both IDs.

Hosted: the same bot in three steps

Aplite runs a standard OpenClaw instance for you, so the gateway is already up and the config is a form. Same bot, same pairing rule, no server:

1. Create a workspace. Email or Google, no card on the Free plan. Pick an agent, press Deploy, and about a minute later the workspace opens with its own OpenClaw 2026.9.8 in its own container.

2. Connect the bot. Open Skills → Channels → Telegram and press Connect. Paste the token from @BotFather, press Validate: Aplite calls Telegram's getMe and shows Connected to @your_bot. Press Connect again; the workspace restarts to pick up the bot, about a minute. Under Show advanced sit the same knobs the config file has: DM policy (pairing by default), group policy (allowlist), chat IDs, require @-mention.

Skills → Channels → Telegram: the card is one Connect button away from a bot.

Paste the token, Validate, Connect: the dialog does what channels add does.

1. Say /start, then Approve. Send /start to the bot from your phone. A pairing request appears on the same Telegram card within seconds, with the sender's name and the code; press Approve. Codes expire after one hour, so if you were slow, send /start again. Only chats you approve can talk to your agent.

The advanced block: pairing, allowlist, open or disabled for DMs; allowlist for groups; the mention rule.

That is the setup. The bot answers in Telegram with the same agent that answers in the browser, routines can deliver to it, and the toggle on the Telegram card pauses the channel without dropping the token. One honest line about the Free plan: the workspace sleeps after 6 hours idle and wakes on your next visit, and while it sleeps the bot does not answer, so a bot that answers at any hour needs Pro or Power. Aplite is an independent service and is not affiliated with the OpenClaw project.

If the bot stays silent

• Nothing at all after /start. It is waiting for approval. Run openclaw pairing list telegram, approve the code, or approve on the Telegram card. A code older than an hour is gone; send /start again (channel troubleshooting).
• getMe returned 401 or deleteWebhook 401 Unauthorized at startup. The token is wrong or revoked. Copy it again from BotFather and update botToken, tokenFile or TELEGRAM_BOT_TOKEN (Telegram troubleshooting).
• Answers in DMs, silent in a group. The group is not in channels.telegram.groups, or Privacy Mode is on while you expect unmentioned messages: /setprivacy → Disable, then remove and re-add the bot. Hosted, the same two settings are in the advanced block.
• Worked yesterday, silent today. Self-hosted: openclaw gateway status should say Runtime: running; a laptop that slept or a VPS that rebooted without openclaw gateway install stops polling. Hosted on Free: the workspace is asleep; open it once.

FAQ

Does OpenClaw need a server to work in Telegram?

Yes. The gateway polls Telegram for new messages, so something has to keep running: your machine while it is awake, a VPS, or a managed instance. A message sent while the gateway is down is not answered until it is back.

Can more than one person use the bot?

Yes. Each person sends /start, gets their own pairing code, and you approve each one; an unapproved sender never reaches the agent. For a shared group, list its chat ID under groups or switch the group policy to open.

Do I need an API key for the Telegram bot?

Not for Telegram: the bot token from @BotFather is free. The model behind the bot is the usual question. Self-hosted, that is a provider key, a subscription login or a local model (every option); on Aplite the models are metered and included.

Why does the bot answer my question with a code?

That is pairing, OpenClaw's default DM policy: the first message from an unknown chat gets a code, not an answer, until the operator approves it with openclaw pairing approve telegram <CODE> or the Approve button. Set dmPolicy to open only if you want anyone who finds the bot to command your agent.

Search for openclaw hosting free and the first page is three kinds of result: hosts whose "free" is a trial or a first-month price, Oracle free-tier guides (several still quoting an Arm allowance Oracle has since halved), and "best hosting" lists written by the hosts. What none of them does is the accounting. An agent that works around the clock costs two things — a machine that stays on, and model tokens for every reply — and most free offers cover one of the two, for a while.

Here is that accounting as of October 2026, and the one path that costs nothing, with every command run on OpenClaw 2026.9.8. If you would rather not run a server at all, the same runtime is hosted in three steps further down.

What "free" has to cover

The server. The gateway is a Node process that has to stay up. The upstream hardware FAQ puts the floor at 1 vCPU, 1 GB of RAM and about 500 MB of disk, and recommends 2 GB or more. The floor is real: capped at 1 GB, the 2026.9.8 gateway settles at about 740 MB of resident memory a minute after start and logs memory pressure: level=warning. The smallest free cloud machines give you exactly 1 GB.

The model. Every reply is a model call, billed per token wherever the gateway runs. Three free ways: a local model through Ollama (needs a capable machine; the same FAQ warns that small quantized models truncate context and raise prompt-injection risk); OpenRouter's :free variants, limited to 20 requests a minute and 50 a day until you have bought $10 of credits, then 1,000 a day; or the Gemini API free tier, where Flash is free of charge and, in Google's words, "content used to improve our products". Hosted plans that include credits cover tokens for as long as the credits last.

The options, side by side

Prices and terms as each page read on 13 October 2026; the links go to those pages.

The $0 path, step by step

Run in a fresh container of our 2026.9.8 runtime image with a 1 GB memory cap, and checked against the upstream docs. Your install will be newer — the gateway says so the moment it starts — but the commands are the same.

1. Install. OpenClaw needs Node 24.16 or 26.1 and newer (install docs); the installer provisions it when missing. On macOS, Linux or WSL2:

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

   On Windows, in PowerShell:

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

   Then openclaw --version — ours printed OpenClaw 2026.9.8 (fc23bc8).

2. Have a free model ready. Local: install Ollama and pull a model (ollama pull gemma4 in the upstream recipe). Or get a key from OpenRouter or Google AI Studio. Pick Ollama with the server not running and onboarding stops here:

   Ollama could not be reached at http://127.0.0.1:11434.
   Start or restart the Ollama server for this address.
   If Ollama is not installed on that machine, download it at https://ollama.com/download
   

3. Run onboarding.

   openclaw onboard
   

   Choose Quick start, pick the provider, paste the key. OpenClaw verifies it with a real completion, writes ~/.openclaw/openclaw.json, starts the gateway in the foreground and opens the dashboard. Skip the provider and everything still installs, but the first message fails with:

   No API key found for provider "openrouter".
   Auth store: ~/.openclaw/agents/main/agent/openclaw-agent.sqlite
   Configure an API key (openclaw models auth paste-api-key --provider openrouter)
   | missing-provider-auth
   

   That command is the fix: openclaw models auth paste-api-key --provider openrouter.

4. Check the gateway. In a second terminal, openclaw gateway status prints:

   Gateway: bind=loopback (127.0.0.1), port=18789 (env/config)
   Dashboard: http://127.0.0.1:18789/
   Connectivity probe: ok
   Listening: 127.0.0.1:18789
   

   openclaw dashboard opens the browser with a one-time pairing token in the URL. With no provider set, the Control UI lands on Model Setup.

The Control UI on a fresh install: the gateway is up, no model is connected yet.

1. Keep it running after you log out.

   openclaw gateway install
   

   This installs a systemd user unit on Linux and WSL2, a LaunchAgent on macOS, a Scheduled Task on Windows (gateway docs). On a server, add sudo loginctl enable-linger $USER so the unit survives your SSH session, as the upstream Oracle guide does.

Where this path ends is in the same upstream FAQ: on a laptop, "sleep, network drops, and OS updates can interrupt the Gateway", and "use a VPS if you have had gateway disconnects before". On the free clouds it is the 1 GB floor and Oracle's reclaim of idle instances; on free models it is the request cap, since one agent turn with tools makes several model requests. That is when people pay — for a $6–12 VPS plus a key, or for a managed host.

Hosted: the same runtime in three steps

Aplite runs a standard OpenClaw instance for you, set up from the browser. The whole path:

1. Create an account. Email or Google, no card. The Free plan is one agent, the full workspace, budget models and 300 welcome credits that never expire.

2. Deploy. Pick the agents and a model for each, press Deploy. About a minute later the workspace opens with its own OpenClaw instance — currently 2026.9.8 — in its own container.

A Free workspace: a standard OpenClaw runtime in its own container.

1. Ask. Nothing to install and no key to paste: models come through one metered account, 1 credit is 1¢, spent only when the agent calls a model, and the cost sits under every reply.

The agent answers in the browser with no key pasted anywhere; the line under the reply is what it cost.

Settings → Cost: the welcome credits, and what the week spent.

On top of the runtime sit a markdown workspace the agent reads and writes, an approval queue, and Telegram delivery from a bot token; scheduled routines are on Pro ($20/mo). What a managed instance gives up — the shell, raw openclaw.json, local models, day-one releases — is listed on the OpenClaw page. Aplite is an independent service and is not affiliated with the OpenClaw project.

FAQ

Is OpenClaw itself free?

Yes: open-source software under the MIT license, and the install costs nothing. What costs money is the machine it runs on and the model it calls, which is what "free hosting" offers differ on.

Can I host OpenClaw for free forever?

On your own machine with a local model through Ollama, yes — as long as the machine stays on. Oracle's Always Free tier has no end date but reclaims idle instances and is often out of capacity; Runware charges nothing for hosting and bills tokens; Aplite's Free plan has no end date and welcome credits that do not expire. The rest are trials or credits.

Do I need a credit card for free OpenClaw hosting?

Not for your own machine, Runware, RapidClaw, Ampere during its beta, or Aplite Free. Oracle takes a card and a phone number at sign-up without charging; Google Cloud requires a payment method; Hostinger and xCloud are paid from day one.

Do I need an API key?

Self-hosting, yes — unless the model runs locally with Ollama; the key is where the token bill goes. On Aplite there is no key: models run through a metered account and usage is billed in credits.

Originally published on aplite.ai. Try it free at aplite.ai.