Cloud assistants
A cloud assistant is a Chat assistant that runs at a model provider instead of on an Edge machine. A workspace owner or admin connects the provider with the workspace's own API key, then adds named assistants that members pick in Chat.
OpenAI is the provider you can connect today. A provider is added once it passes One Horizon's tests.
Cloud assistants are turned on one workspace at a time. If Settings → Assistants has no Cloud providers section, they aren't on for your workspace yet.
Before you connect a provider
Connecting a provider changes where conversations go and who pays for them.
- What is sent. Messages, earlier conversation context, and tool results are sent to the provider when someone uses one of its assistants.
- Who bills. The provider bills your workspace's account with that provider. Usage isn't part of your One Horizon plan, and a ChatGPT or Claude subscription doesn't cover it.
- Where the key lives. The key stays on One Horizon's servers. Members never see it, and it's never sent to a member's computer or written into a conversation.
Connect a provider
Only workspace owners and admins can open Settings → Assistants.
- Under Cloud providers, select Connect on the provider's row.
- Paste the key into API key and read What connecting means.
- Select Connect. The key is checked with the provider before it is saved.
Nothing is saved when the provider refuses the key or can't be reached. The message under the field says which of the two happened.
A saved key is never shown again, in full or in part. To change it, replace it.
After a first connection, the dialog continues to Add assistant, so members have something to pick in Chat. Select Not now to add one later.
Add an assistant
Select Add assistant on a connected provider's row, or at the top of the Assistants section.
- Under Runs on, choose Cloud provider. This choice appears only when the workspace has both a connected provider and an Edge machine.
- Choose the Provider and the Model. The list holds the models One Horizon has tested with that provider. A model your key can't use stays in the list, disabled, and reads Your key doesn't have access to this model.
- Give it a Name. Members see this name in Chat.
- Select Add. The assistant's details page opens.
An assistant's provider and model can't be changed after you add it. To use another model, add a new assistant. You can still change its name, its capabilities, Enabled, and Default for Chat.
An assistant always answers with the model it was added with. If the provider can't serve that model, the answer fails. One Horizon never moves the conversation to another model, assistant, or provider.
Capabilities
Cloud assistants use the same Capabilities card as Edge assistants. A new cloud assistant starts with the same two entries: One Horizon read tools · Built-in and Web search · Built-in. Connecting a provider adds nothing else. Agent capabilities covers how to add, require, disable, or remove an entry.
Turn off or remove an assistant
Switch off Enabled to stop new chats from using an assistant. Switch it back on to offer it again.
To remove an assistant for good, open its options and choose Remove. Members can no longer pick it. Conversations that used it stay readable.
Replace a key, disconnect, or reconnect
Open the options on the provider's row under Cloud providers.
Replace key
Paste a new key and select Replace key. If the provider refuses the new key, your current key stays in use. Assistants, the default, and conversations don't change.
Disconnect
Disconnecting deletes the stored key. The provider's assistants stay in your list, marked Provider disconnected, and can't answer until you reconnect. Saved conversations stay readable. Other providers aren't affected.
Reconnect
Select Reconnect on a disconnected provider's row and enter a key. Its assistants work again, with the same names and capabilities. The default assistant stays the default.
Pick a cloud assistant in Chat
Chat needs a connected computer, whichever assistant you pick. Open One Horizon on your computer before you send a message.
In the assistant picker, a cloud assistant carries a Cloud marker, and its second line names the model and the provider. While one is selected:
- The picker button shows the assistant's name followed by Cloud.
- A line under the composer says where the message runs and who is billed, for example Runs at OpenAI. OpenAI bills this workspace's account.
- Each answer names the assistant, model, and provider that wrote it.
Admins also get Manage assistants at the foot of the picker.
Switching provider inside a conversation
When you pick an assistant whose provider hasn't received this conversation yet, a notice sits above the composer until you send or change the selection:
Team research runs at OpenAI. Sending shares the earlier messages in this conversation with OpenAI.
It doesn't appear in a new, empty chat, or when the provider has already received the conversation.
Private chats
Private chats never use cloud assistants. In a private chat, cloud assistants are listed but can't be picked. One that works in other chats reads Not available in private chats.
When an assistant can't answer
When the assistant you selected becomes unavailable, it stays selected and disabled. Nothing is picked for you. A notice above the composer gives the reason, and Send stays off until you pick another assistant.
| What happened | Member reads | Admin also gets |
|---|---|---|
| Assistant turned off | Team research is turned off. Pick another assistant to continue. | Open assistant |
| Provider disconnected | Team research can't answer because OpenAI is disconnected. Pick another assistant, or ask a workspace admin to reconnect it. | Reconnect OpenAI |
| Provider no longer offers the model | OpenAI no longer offers the model Team research uses. Pick another assistant. | Open assistant |
| Assistant removed | Team research was removed. Pick another assistant to continue. | None |
A failure after you send shows under the answer. Text that already arrived stays visible.
| What happened | Member reads | Admin also gets |
|---|---|---|
| Provider refused the key | OpenAI didn't accept this workspace's key, so the answer stopped. Ask a workspace admin to replace it. | Replace key |
| Provider account out of credit | This workspace's OpenAI account is out of credit. Ask a workspace admin to add credit at OpenAI. | Open provider settings |
| Provider is limiting requests | OpenAI is limiting requests from this workspace right now. Wait a moment, then retry. | None |
| Provider unreachable | OpenAI couldn't be reached. Try again in a moment. | None |
A refused key also shows in Settings as Key not accepted on the provider's row, with Replace key as its main action. Credit is added in your account at the provider, not in One Horizon.
Retry a failed answer
Retry with Team research answers the same message again with the assistant that failed, whichever assistant is selected now. Your message appears in the conversation once, however often you retry.
If that assistant can no longer be used, the button is off and Retry isn't available: gives the reason. Pick another assistant and send your message again.
Actions you already approved
An answer can stop after you approved an action, such as creating a task. Under the failed answer, each action you approved for that message gets a line, earlier retries included:
- Create task was approved and ran before the answer stopped. One Horizon saw the action finish.
- Send update was approved and may have run. The action reported a failure, or no result arrived. It may have stopped before it started, or finished and lost its reply. Check before you rely on it.
The five most recent actions are listed, then a count of the rest. The list ends with Retry asks you again before running anything.
In a retried answer, every approval request carries a warning above its buttons:
Already approved for this message, in an answer that stopped:
- Create task (ran)
- Send update (may have run)
If this is the same action, approving runs it a second time.
One Horizon doesn't try to decide whether the new request is the same action. Select Decline to run nothing.