AgentsOpenClaw

Get a shareable link for what OpenClaw makes

Add the host0 skill to your Gateway and message your agent from any channel. It runs one deploy through exec and replies with a URL that only you can open until you share it. We've read OpenClaw's docs but haven't run this ourselves, so treat it as what should happen.

Looking for somewhere to run OpenClaw? host0 can't do that: it serves files and shared records, and no long-running process of yours runs on it. The first question below points to what OpenClaw's docs suggest instead.

Get started

Free plan, no card. Pro is $9 a month.

Who can open what you publish

It doesn't matter which channel you asked from: a new deploy is private to the host0 account whose key is on the Gateway. Being in the group where it was requested gives nobody access. Reply in the thread to change that: "make it a link anyone can open" or "invite sam@acme.com".

Private: the default. You and people you invite by email, each signing in as themselves.
Anyone with the link: no sign-in, unlisted and never indexed.
Public on the web: no sign-in, and search engines may index it.

Access

No one yet. Add an address and watch the visitor's window.

Who's visiting

p7wz3nc9ha.host0.appSomeone signed out

This app is private — sign in to check your access.

Continue with Googleyou@company.comContinue

Nothing is served until they sign in, and only people you've invited get past that.

Search engines won't index itx-robots-tag: noindex, nofollow, noarchive

Ways to connect OpenClaw

Route	What it can do	Set it up
Skill: Install it on the machine running the Gateway, not on your phone. The first command puts it in the active workspace, the second in `~/.openclaw/skills` for every local agent. The setup prompt above also works when pasted into a chat, but it only loads the skill for that conversation. Best for: Anything you'll publish more than once	What it can do: Deploys, Manages	`openclaw skills install skills-sh:host0-ai/skills/host0` ~/.openclaw/skills/host0 `curl -fsSL https://host0.ai/install.sh \| bash -s -- --dir ~/.openclaw/skills/host0`
host0 CLI: The agent runs this through `exec`, with nothing installed. A normal OpenClaw install already has a newer Node than the 18.18 this needs. Without the skill, you have to tell the agent to use it. Best for: A one-off from an agent with no skills allowed	What it can do: Deploys, Manages	`npx -y @host0/cli@latest deploy ./my-site`
MCP server: Save the server, then sign in. If Settings → MCP in the Control UI offers Sign in, use that and keep the Gateway running. Otherwise OpenClaw's terminal login needs the Gateway stopped, as below. Best for: An agent with no shell that still manages apps	What it can do: Can't deploy, Manages	`openclaw mcp add host0 --url https://mcp.host0.ai/mcp --transport streamable-http --auth oauth` `openclaw gateway stop openclaw mcp login host0 openclaw gateway start openclaw mcp doctor host0 --probe`
Browser drop: Ask OpenClaw to send you the file in chat, save it, and drop it on host0. It needs no skill and no `exec`. You sign in with an email code or Google. Best for: Agents limited to messaging tools	What it can do: Deploys, Dashboard to manage	Drop a file on host0

Can host0 host OpenClaw itself?

No. The OpenClaw Gateway is a process that stays running, and host0 has nowhere to run one. If you came here for a place to run your agent, you need a server, and this page won't help with that part.

OpenClaw's own hosting docs are the place to start. They cover a Linux VPS at DigitalOcean, Hetzner, Hostinger or Oracle Cloud, browser setups on Railway and Northflank, Fly.io, Render, Azure, GCP, a Raspberry Pi, and an experimental Cloudflare template. Your own laptop counts too.

host0 is for the step after that. Wherever the agent runs, it ends up making a page, a report or a small app, and those need an address other people can open. Your Gateway can stay on loopback behind an SSH tunnel or Tailscale, as OpenClaw recommends, because visitors never reach it.

Can't OpenClaw just send me the file?

Yes, and when the reader is already in that chat, an attachment is the simpler choice. A link starts to matter when the reader isn't in the chat, when the thing is a whole site rather than one file, or when people need to write to it.

	Attachment in chat	host0 link
Who can open it	People in that conversation	People you invite by email, anyone with the link, or the whole web
What fits	One file at a time. OpenClaw's WhatsApp caps are 100 MB for documents and 16 MB for video	A folder of pages and assets, or an app. On Free, up to 3 MB a file
After a change	Send a new file, and the old one stays in the thread	Same URL, new version
People writing to it	No	Yes, with shared data from an `h0.json`
Where it lives	With the chat app you sent it through	On host0's servers, not your machine

The Control UI has a middle option: message with a local HTML path saves a snapshot as an attachment in that conversation. It has to be self-contained, because relative assets aren't bundled. Tailscale Funnel isn't a shortcut either. OpenClaw uses it to put the Gateway dashboard online behind a shared password, not your files.

There's also a reason to skip host0 altogether. OpenClaw's pitch is no hosted service in the middle, and host0 is a hosted service: whatever you publish, shared data included, leaves your hardware. If it mustn't, stick to attachments or serve it from a web server you run.

What happens when I ask from Telegram or WhatsApp?

Your message reaches the Gateway, the agent runs the deploy through exec on that host, and the URL comes back as its reply in the same thread. This should work on any channel, but we haven't tried it on one.

The command runs where exec runs: the Gateway host unless tools.exec.host sends it to a paired node. The folder and the saved key must be on that machine.
On default settings, exec on the Gateway has no approval prompt, so nothing asks before a deploy. Stricter modes do, with native approval cards where the channel has them and /approve where it doesn't.
With no key yet, the first deploy prints a link and a short code. Open the link in your own browser and approve. You don't need access to the server, and no personal details pass through the chat.
OpenClaw hides tool progress in chat by default, so the link might not arrive until the agent's turn ends, up to 90 seconds later. If so, approve it and reply that you have. The agent runs the same command again and carries on. The code lasts 10 minutes.

To be sure the skill gets used, name it. $host0 publish the report folder works for authorized senders on every channel, and /host0 is the standalone command form.

Why can't my agent run the deploy?

Most often exec isn't available to that agent or session. A skill is only instructions and grants no tools, so an agent can know exactly how to publish and still have no way to run the command.

What you see	Likely cause	What to change
It says it can't run commands	The `messaging` or `minimal` tool profile, neither of which includes `exec`	Move that agent to the `coding` profile
Fine in a DM, refused in a group	That group's `tools` or `toolsBySender` denies `exec`, or `mode: "non-main"` sandboxes groups	Publish from the DM, or allow it for your sender id
A denial, or a prompt that never arrives	`tools.exec.mode` is `deny`, `allowlist` or `ask`. With no approval surface reachable, the request falls back to deny	Approve it, or pick Always allow here, which covers that exact command in that folder
`executable could not be resolved`	In allowlist mode, `npx` isn't on the Gateway's `PATH`	Add its folder with `tools.exec.pathPrepend`
It doesn't know the skill	A non-empty `agents.entries.<id>.skills` is the final list, so anything left off stays hidden	Add `host0` to that list, then run `openclaw skills check`

Does publishing work with OpenClaw's sandbox on?

Not on the sandbox's defaults. Sandboxing is off out of the box, but once it's on, a local container gets no network, and a deploy can't happen without one. Three things need changing, and we haven't tried this combination.

Network: set agents.defaults.sandbox.docker.network. Uploads go to host0.app and everything else to host0.ai.
Runtime: the default openclaw-sandbox:trixie-slim image has curl and jq but no Node, so the skill falls back to publish.sh. OpenClaw's optional common image adds Node 24.
Key: ~/.host0/credentials on the host isn't in the container, and skills.entries.*.env only reaches the host process. Pass HOST0_API_KEY through sandbox.docker.env.

The shorter route is mode: "non-main", which sandboxes groups and leaves your main DM on the host, so you publish from the DM. tools.elevated can also run a command outside the sandbox when the operator allows it.

What changes when my agent lives in a group chat?

Everyone who can trigger that agent can use your host0 account, because the key is saved once on the Gateway host, not per sender. OpenClaw's docs say the same of the Gateway itself: approvals aren't a per-user boundary, and a shared agent suits people who already trust each other.

Do the first login in a DM. In a group, the link and code land in the room, and whoever approves decides which host0 account the agent gets.
Narrow the tool. A group's toolsBySender can deny exec for the room and allow it for your own sender id.
Group members get no access to an app just by being in the group. For a private app, invite their email addresses.

How do I check the host0 skill before installing it?

Read it. The public repository host0-ai/skills holds what gets installed, and it comes down to two files: skills/host0/SKILL.md and skills/host0/scripts/publish.sh. It's worth the few minutes, since on a default Gateway a skill's instructions turn into commands with no approval prompt.

What to look for: it talks to host0.ai and host0.app, plus the npm registry when it uses the CLI, and it saves its key in ~/.host0/credentials. It refuses to send that key to another host unless given an explicit flag.
openclaw skills verify won't help here. It asks ClawHub for its audit result, and the commands on this page don't go through ClawHub.
OpenClaw's git: install form won't find it either, because that form expects SKILL.md at the repository root. To install a copy you've read, point openclaw skills install at the local folder.
security.installPolicy runs a check of your own before ClawHub, Git, local and uploaded installs, if you want one enforced.

The same habit applies to anything on ClawHub. OpenClaw's docs say it's open by default, so anyone with a GitHub account old enough to pass its upload gate can publish there. They also say a Pass on the audit page doesn't replace your own judgment, least of all for skills that run commands or publish content.

Can a scheduled job keep one page up to date?

Yes, and it suits OpenClaw well: a cron job rebuilds a folder and deploys it, and the URL stays put. Do the first login and first publish by hand, because an unattended run has nobody to approve a link. We haven't run this.

Keep one folder. .host0/state.json inside it ties the folder to the app, so a job that builds somewhere new each run makes a new app each run.
On the default mode nothing asks. In stricter modes, a scheduled run's approval card goes to the Control UI and native apps, never to chat channels, and with none of them connected the command is denied at once.
Always allow on that card makes a standing grant for that exact job, command, folder and environment. Change any of them and it asks again.

My Gateway runs on a VPS. What's different?

The deploy itself is the same, since the VPS is the machine doing it. What changes is where things sit: the folder, .host0/state.json and the key are all on the server, not your laptop.

OpenClaw's docs tell you to treat the VPS as the source of truth and back up its state and workspace. Keep .host0/state.json in that backup. A rebuilt server without it creates a new app on its next deploy, unless you name the slug.
Signing in to host0 needs no tunnel. The link opens in your phone's browser, and the agent on the server picks up the key.
The MCP sign-in is the fiddly one, because OpenClaw's terminal login expects a browser that can reach a loopback address. Use the printed --code fallback, or Sign in under Settings → MCP over the Gateway's HTTPS or Tailscale Serve address.
The browser drop needs the file on your device, so ask the agent to send it to you first.

Do I need the host0 MCP server in OpenClaw?

Probably not. It manages apps that already exist and can't deploy, and the skill's CLI does the same management over exec. It's worth adding when you've turned exec off for an agent and still want that agent to change who can open an app.

Keep --transport streamable-http in the command. OpenClaw assumes SSE when it's missing.
Credentials are shared by default, so everyone using that agent acts as whoever signed in. Per-requester OAuth gives each sender their own account and needs gateway.publicOrigin. OpenClaw warns that its sign-in links work for whichever chat participant opens them.
MCP tools go through the same tool profile and policy as everything else, so connecting the server doesn't get around a restricted agent.

Give it a shared database, with no server code

Ask your agent for something people add to and it writes an h0.json beside the page. Visitors talk to host0, never to your Gateway, so the list keeps working while your laptop sleeps or your server restarts.

By default, everyone the app admits can read and add, and changing or removing a record takes an invite. Collaboration is poll-based, so a change shows up in the other window on its next poll. Try it: this list only lives on this page.

p7wz3nc9ha.host0.app5s

Team todos

AKAna

Ship the pricing pageAK
Book the launch demoSR

GET /.h0/data/todos → 200

p7wz3nc9ha.host0.app3s

Team todos

SRSam

Ship the pricing pageAK
Book the launch demoSR

GET /.h0/data/todos → 200

h0.json

{
  "records": {
    "todos": {
      "fields": {
        "title": {
          "type": "string",
          "required": true,
          "maxLength": 140
        },
        "done": {
          "type": "boolean",
          "default": false
        }
      }
    }
  }
}

Real apps, live on host0

Each one came from a single prompt. Open it, or paste the prompt into OpenClaw to get your own copy.

One-page report as a PDF link

A PDF report that opens in any browser, with a download button.

Built with Claude Code

Open app

Async standup log

Your remote team posts daily updates on one page instead of a meeting.

Built with Claude Code

Open app

Potluck sign-up sheet

Guests claim a dish so you don't end up with five salads.

Built with Claude Code

Open app

Undo a bad deploy, and see who opened it

Versions

Each republish of one folder adds a numbered version behind an unchanged URL. When a run goes wrong, message your agent to roll back, or restore from the dashboard. On Free: live version only. On Pro: 30 days of version history.

Analytics

Page views, countries and referrers on every app, with no cookies and nothing for you to install. On Free: total visits only. On Pro: full, 90 days.

OpenClaw and host0: questions

Does host0 run my OpenClaw agent?

No. host0 runs no agents and no server processes of yours. Your Gateway stays on your own machine or server, and host0 serves the files and shared data it publishes.

Does this cost anything on top of OpenClaw?

Free plan, no card. Pro is $9 a month. On Free, a file can be up to 3 MB and an account holds 5 apps. OpenClaw's docs say it has no paid tier of its own, and your server and model costs don't change.

OpenClaw doesn't see the skill. What do I check?

Run openclaw skills check on the Gateway host, which separates skills that were found from those that are ready and visible. If you set OPENCLAW_STATE_DIR, the shared folder is skills inside that directory rather than ~/.openclaw/skills. Existing sessions pick up new skills on their next turn unless skills.load.watch is off, in which case restart the Gateway.

How do I redeploy from another chat?

Ask the same agent again from any channel. The app is tied to the folder through .host0/state.json, not to the conversation. A different agent has its own workspace, so give it the slug, or it creates a new app.

Do the people I share with need OpenClaw?

No. They open an ordinary URL in a browser and never touch your Gateway. They sign in only if the app is private or they're changing shared data.

What can't my agent publish?

Anything that needs a process. A dev server your agent started on the Gateway host won't move to host0, so ask for static output, and move lists and forms to shared records. An archive is refused too, so have the agent extract a zip and publish the folder.

Publishing from something else?

Share what OpenClaw made

A link in one sentence, private until you say otherwise.

Get started

host0 isn't affiliated with OpenClaw or its maker. Product names and marks are used only to identify the product.

Who can open what you publish

Ways to connect OpenClaw

Can host0 host OpenClaw itself?

Can't OpenClaw just send me the file?

What happens when I ask from Telegram or WhatsApp?

Why can't my agent run the deploy?

Does publishing work with OpenClaw's sandbox on?

What changes when my agent lives in a group chat?

How do I check the host0 skill before installing it?

Can a scheduled job keep one page up to date?

My Gateway runs on a VPS. What's different?

Do I need the host0 MCP server in OpenClaw?

Give it a shared database, with no server code

Real apps, live on host0

One-page report as a PDF link

Async standup log

Potluck sign-up sheet

Undo a bad deploy, and see who opened it

Versions

Analytics

OpenClaw and host0: questions

Sources

Publishing from something else?

Share what OpenClaw made