Skip to Content

Use cases

Real jobs from start to finish, with video walkthroughs.

Set up with Clem

You'll have an agent named Support and a teammate invited to use it. You need to be an administrator, and a teammate's email.

Ask Clem in plain words, then approve the one sensitive step
A free trial may have room for only one agent VM. If there's no room, the chat says so.

Open a chat with Clem

Open Chat. In the sidebar, point at Clem and choose New chat with Clem, the + beside it.

Clem greets you: “Ask me how to do anything here, or ask me to do it for you.”

Ask for an agent

Send:

Create an OpenClaw agent named Support on the default size. Go ahead, no need to confirm first.

Clem starts creating it. When it's done, the chat says “Support is ready.”

Invite a teammate to it

Send this, with your teammate's email in place of EMAIL:

Invite EMAIL as a user and give them Chat and settings on Support.

A card shows what Clem will do, with Decline and Approve. Giving Chat and settings waits for your click.

Approve it

Read the card and choose Approve.

The card shows how it went. Your teammate gets an email invitation.

See what Clem set up

Open Users. Your teammate is listed as Invited.

Open Agents. Support is listed.

Next

Build an app and publish it

You'll have a booking page online, built by an agent named Scout. You need to be an administrator.

Scout builds a booking page, puts it online, then changes it
A free trial may have room for only one agent VM. If there's no room, the chat says so.

Ask Clem for Scout

Open Chat. In the sidebar, point at Clem and choose New chat with Clem, the + beside it.

Send:

Create an OpenClaw agent named Scout on the default size. Go ahead, no need to confirm first.

When it's done, the chat says “Scout is ready.”

Let Scout publish

In the same chat, send:

Let Scout publish web apps from its own VM through the Sunport's domain.

A card lists the permissions Clem will give Scout. Choose Approve.

The card shows how it went.

Ask Scout for the app

In the sidebar, point at Scout and choose New chat with Scout, the + beside it.

Send:

Build a booking page for my cafe, Juniper Cafe: date, time, party size, name, and a Book a table button. Put it online for anyone with the link, and send me the link.

While Scout works, the chat's header says so. Its answer has the link.

Open the page

Open the link and book a table.

The page opens without signing in, as it would for anyone.

Change it

Send:

Make the Book a table button orange. Keep the same link.

Scout answers when it's done. Reload the page to see the orange button.

Find it in Compute

Open Compute and choose Domains.

Scout's address is listed with Anyone with the address reaches it. Turn on sign-in in its menu.

Next

Use an agent's computer

You'll have a short video an agent named Scout made on its own VM. You need to be an administrator.

Scout installs ffmpeg on its own VM and makes a time-lapse
A free trial may have room for only one agent VM. If there's no room, the chat says so.

Ask Clem for Scout

Open Chat. In the sidebar, point at Clem and choose New chat with Clem, the + beside it.

Send:

Create an OpenClaw agent named Scout on the default size. Go ahead, no need to confirm first.

When it's done, the chat says “Scout is ready.”

Give Scout the job

In the sidebar, point at Scout and choose New chat with Scout, the + beside it.

Send:

Install ffmpeg. Then draw 500 frames of a sunrise over a cafe and make a 20-second time-lapse. Attach the video here.

While Scout works, the chat's header says so.

Installing may need root on Scout's VM. If Scout asks for it, open Agents, choose Scout, then Server, and approve it there.

Watch the result

When Scout answers, play the video in the chat.

Worked for above the answer says how long it took.

See Scout's computer

Open Agents, choose Scout, then Server.

Under Virtual machine, you see the VM Scout works on, with its size.

Next

Run a job across agents

You'll have research and a launch post in one chat: Scout researches, and Quill writes. You need to be an administrator on a paid Sunport, as a free trial may have room for only one agent VM.

One ask to Scout brings back research, a launch post and a poster

Ask Clem for two agents

Open Chat. In the sidebar, point at Clem and choose New chat with Clem, the + beside it.

Send:

Create two OpenClaw agents on the default size, named Scout and Quill. Go ahead, no need to confirm first.

When they're done, the chat says “Scout is ready.” and “Quill is ready.”

Let Scout message Quill

In the same chat, send:

Let Scout message Quill.

Clem answers when it's done.

Open Agents, choose Scout, then Access. Sees and messages names Quill.

Ask Scout for the whole job

Open Chat. In the sidebar, point at Scout and choose New chat with Scout, the + beside it.

Send:

Research autumn cafe trends in Portland, with sources. Then ask Quill to write a short launch post for our autumn menu from your research. Send me both here.

While Scout works, the chat's header says so. Its answer has the research and Quill's post.

See the hand-off

At the top right, choose Side panel, then Tasks.

Under This chat, choose Finished. It lists Scout's message to Quill, marked as answered. Choose it to read their conversation.

Next

An agent for each teammate

The situation

Each person on your team wants an AI assistant at work. You want to choose the models and see what it costs.

What you set up

For each person:

  • an invitation to your Sunport;
  • an agent on a VM of its own, owned by them;
  • optionally, their own Google account, connected by them.

Steps

Invite the person

Follow Invite a person.

They sign up from the email's link.

Create their agent

Follow Create an agent. Under Settings, choose them as Owner.

They must have signed up first. As the owner, they can use the agent at once.

Let them connect Google

They follow Connect Google, and choose their agent under Who can use this connection.

They pick only the services the agent needs.

What it costs

Each size on the New agent page shows its monthly price, or Included on OpenOrange capacity. Model requests cost extra, and Requests lists each one. An administrator can cap each agent's model spend with a Monthly spend limit when creating it, or later with Limit what an agent spends.

Guardrails

  • Only administrators, the owner and people given access can use the agent. To add someone, follow Let people use an agent.
  • Through Google, the agent deletes data, sends messages, invites people or shares access only after its person approves that action in Settings › Approvals, as in Decide an approval.
  • A Google connection keeps its approved permissions until the person disconnects it.

A team assistant in chat

The situation

Your team works in Slack or Telegram. You want one assistant there that anyone on the team can ask.

What you set up

  • an agent that runs on an assistant runtime, OpenClaw or Hermes;
  • its bot in your chat app;
  • access for each teammate, through their linked chat account;
  • answers in group channels, when someone mentions the bot.

Steps

Create the agent

Follow Create an agent. Under Assistants, choose OpenClaw or Hermes.

Coding agents run no chat app bots.

Let your teammates in

Follow Let people use an agent for each teammate.

Each teammate then follows Link your chat account. That needs Clem's bot on the same app: Connect Clem to chat apps. Without it, an administrator follows Link a person's chat account.

Let it answer in channels

For OpenClaw, follow Let an agent answer in groups. A Hermes agent answers in groups without this step.

The bot then answers in the channels it is in, when someone allowed mentions it.

What it costs

The agent's VM has a monthly price, or Included on OpenOrange capacity. Each answer is a model request, listed on Requests. To cap it, follow Limit what an agent spends.

Guardrails

  • The bot answers only people on its allowlist, in channels too.
  • In a channel, everyone there reads the answer.
  • Each person's message runs at their own level on the agent.
  • A linked chat account counts as its person: their limits apply, and their usage shows under them.
  • Telegram bots start in privacy mode and miss mentions in groups. Turn it off in BotFather with /setprivacy, then add the bot to the group again.

An agent for your Google account

The situation

You want your agent to read your mail, find files in Drive, or manage your calendar. You want to approve anything it sends, shares or deletes.

What you set up

  • your organization's own Google app, if you need more than Gmail labels and chosen Drive files;
  • a Google connection of your own, which your agent may use;
  • approvals for what the agent does with it.

Steps

Set up your organization's Google app, if needed

OpenOrange's Google app offers Gmail labels and the Drive files you choose.

For Gmail, Calendar, Docs and more, an administrator follows Set up your Google app.

Connect Google

Follow Connect Google.

Under Who can use this connection, choose your agent. Pick only the services it needs.

Ask your agent

Follow Chat with an agent and ask it for the task.

Your connection's Available actions lists what it can do with it.

Approve what it sends

When the agent wants to send, share or delete, it waits.

Follow Decide an approval.

What it costs

The agent's work is model requests, listed on Requests.

Guardrails

  • The agent deletes data, sends messages, invites people or shares access only after you approve that exact action.
  • A connection keeps the permissions you approved until you disconnect it.
  • Disconnect stops future access. It doesn't remove Google content already in conversations.
  • Google content the agent reads enters its conversation, and its model's provider processes it.

A coding agent for your repo

The situation

You want an agent that changes your code, runs it and pushes branches. You want it on its own computer, not your laptop.

What you set up

  • an agent on a coding runtime, on a VM of its own;
  • your repository on that VM, through OpenOrange's GitHub App;
  • the secrets your code needs.

Steps

Create the coding agent

Follow Create an agent. Under Coding agents, choose a runtime such as Claude or Codex.

Give its VM your repository

Ask the agent to follow Give a computer GitHub access on its VM. You approve OpenOrange's GitHub App, for the repositories you choose.

Or, from your own computer, clone it onto the agent's VM:

oo move-in --from-github OWNER/REPO --agent coder

coder is the agent's ID.

Give it the secrets it needs

Follow Give an agent a secret for each one.

Give it work

Follow Chat with an agent and describe the change.

What it costs

The agent's VM has a monthly price, or Included on OpenOrange capacity. Its model requests cost extra, and Requests lists each one. A Devin agent runs on its owner's Devin subscription instead, outside your Sunport's budgets and usage.

Guardrails

  • The GitHub token stays on the agent's VM. It reaches only the repositories where you installed the app.
  • A secret's value is never shown after saving. The agent's tools read it on its VM.
  • Only administrators, the owner and people given access can use the agent.
  • oo move-in is a person's command. An agent never runs it.

Research your team keeps

The situation

Your team researches the same topics again and again. You want what one person found to be there for the next, with its sources.

What you set up

  • Megabrain, your Sunport's shared memory;
  • a research agent with web search;
  • sources and packets your team can read.

Steps

Turn Megabrain on

Follow Turn Megabrain on.

An administrator follows Check a service's cost for Web search. Its health shows whether agents can search the web.

Create the research agent

Follow Create an agent, then Give an agent instructions.

In its instructions, ask it to keep its sources and publish what it found in Megabrain.

Say who reads it

Items are Owner only unless chosen otherwise. Ask the agent to keep them for Everyone, or Owner's teams.

Publish a packet shows the same choice when you publish yourself.

Find it later

Follow Search Megabrain.

What it costs

Each answer is a model request, listed on Requests. Web search is a service, charged per use. Search by meaning runs on your Sunport by default. Hosted meaning is billed as model use.

Guardrails

  • Each item's audience decides who reads it.
  • A connected agent given Megabrain reads what its owner's teams and everyone share. Its vendor gets what it reads.
  • A published packet never changes. Its owner supersedes it with a new one.
  • A withdrawn item stays readable by its reference, but leaves search.

An app on its own address

The situation

You want a small app, such as a guestbook or an internal dashboard. You want your team to open it at its own address.

What you set up

  • an agent that builds and runs the app on its VM;
  • a name under your Sunport's domain that leads to the app's port;
  • OpenOrange sign-in in front of it, unless it must be public.

Steps

Ask it to build and run the app

Follow Chat with an agent. Describe the app and ask the agent to run it as a service.

If it needs root on its VM, it asks. Follow Give an agent VM access.

Give the app a domain

An administrator follows Publish an app, on the agent's VM and the app's port.

Keep Ask for OpenOrange sign-in on, unless the app must stay public, such as a webhook.

What it costs

The agent's VM has a monthly price, or Included on OpenOrange capacity. Building the app is model requests, listed on Requests.

Guardrails

  • With OpenOrange sign-in, only administrators and the people who may use the agent reach the app.
  • The app's own login, if it has one, stays behind the sign-in.
  • Without it, anyone with the address reaches the app.
  • A name on an edge gateway doesn't pass your Sunport, so it has no OpenOrange sign-in.