Hosted AI assistants

Get answers in your chats from an AI assistant that DonutChat runs for you with your own OpenAI or Anthropic API key. Once it’s set up, @mention it and it replies in the thread under your message.

What you need

Approval. You request access in the app.
An OpenAI or Anthropic API key. Use an API key from your provider account, not a ChatGPT or Claude subscription sign-in and not a bot token. Your provider bills you for the model usage.

Create the assistant

Open My Bots. On iPhone and Android, go to Profile > Tools & content > My Bots. In the desktop app, go to Settings > App > My bots.
Select + (desktop app: Add a bot).
Select Create an AI assistant.
Enter a Name and a Username. You’ll @mention the assistant by its username. See the name and username rules.
Select Create and continue (Android: Continue).
On iPhone and Android, tap Continue setup on the Bot created! screen. The desktop app opens the setup dialog for you.

Creating the assistant doesn’t start it or use your API key.

Request access

Approval covers all your hosted assistants, so you need to request it only once.

On the assistant’s setup screen, Set up your assistant, go to 1. Hosting access.
In Why do you want personal AI bots?, describe how you’d use the assistant.
Tap Request access.

To check the decision, open the assistant from My Bots, tap Continue hosted setup (or View status or continue setup), then tap Check status.

Write 1 to 1,000 characters, and don’t include passwords or keys.
The DonutChat team reviews each request. The app doesn’t notify you of the decision, so check back.
Approval sets how many assistants you can have. Saved drafts count toward that number; deleted setups don’t. Android (My Bots > ⋮ > Hosting approval) and the desktop app (Hosting access) show how many you’ve used.
If your request is declined or your access is revoked, you can ask again with Request reconsideration (desktop app: Send request). Your access isn’t restored until the new request is approved.
If your access is revoked, your assistants are paused. After you’re approved again, resume them yourself.

Set up the assistant

After your request is approved, continue on the same setup screen. Copy your API key first, then paste it and finish setup without switching to another app. If you leave the app or the setup screen, the key is cleared and setup goes back to the provider step. The desktop app does this whenever its window loses focus.

Under 2. Connect a provider, choose OpenAI or Anthropic in Provider.
Paste your key into OpenAI API key or Anthropic API key.
Tap Check connection.
Under 3. Choose a model, choose a Model. Search models filters the list.
In What should your assistant help with?, write its instructions, for example “Summarize discussions and highlight next steps.”
Optionally, open Optional tools and turn on the tools you want.
Tap Review assistant.
Check the summary under 4. Review and start, then tap Start assistant.

Under 2. Connect a model provider, choose the provider in Provider. It’s listed as openai or anthropic.
Paste your key into OpenAI API key or Anthropic API key.
Tap Check connection.
Tap Choose a model.
Under 3. Choose a model, choose a Model. Search models filters the list.
In How should your assistant help?, write its instructions.
Optionally, open Optional tools & permissions and select the tools you want.
Tap Review setup.
Check the summary under 4. Review and start, then tap Start assistant.

When the assistant’s status shows Ready, it can answer. If it doesn’t show Ready yet, refresh the screen (iPhone: Check status; Android: Refresh; desktop app: Refresh status).

Check connection lists the models your key can use. It doesn’t save the key or run a model, and a successful check doesn’t guarantee that later requests will succeed.
The model list shows only models that DonutChat offers for hosting and that your key can use. If yours is missing, try another provider or key.
Instructions can be up to 16,000 characters and go to your provider with every request. Never put passwords or API keys in them.
The optional tools are Task planning, Private workspace and Durable wait, and the app describes each one. You choose them when you start the assistant; to change them later, set it up again.
If you see Hosted Hermes capacity is full. Try again after a slot becomes available., DonutChat can’t host more assistants right now. Try again later.

Add it to a conversation

Your assistant answers only in conversations it belongs to.

Its own conversation: on iPhone and Android, open the assistant from My Bots and tap Open conversation. In the desktop app, click Open conversation next to the assistant in My bots. You get a conversation with just you and the assistant.
A group you own or manage: add the assistant like any other bot. See Add a bot to a chat.

Even in its own conversation, plain messages don’t start the assistant. Ask it with a mention or Run personal bot.

Ask your assistant

Mention it

In a conversation the assistant belongs to, @mention its username in a message or in a thread reply. You can pick it from the @ suggestions. Mentions work from every app, including the web app, although some setup screens still say that only Run personal bot starts the assistant.

In a new message, the assistant answers in the thread under your message, based on that message alone.
In a thread, it answers in the same thread. It also reads your earlier messages in that thread and its own earlier replies. If the thread is long, the oldest messages are left out.
Write the username as a separate word, such as @research_helper, not inside another word or an email address.
Only your mentions start your assistant. When other people mention it, nothing happens.

Use Run personal bot

In the iPhone, Android and desktop apps, you can also run your assistant on one of your own messages without mentioning it.

Long-press one of your own text messages in the chat. In the desktop app, hover over the message and click More actions (⋯), or right-click it.
Choose Run personal bot.
Choose a ready assistant in Personal bot. On Android, tap its name instead.
Turn on I agree to share this message with the selected bot’s AI provider.
Select Share message and run.

The reply appears in that message’s thread. Only the message text and your instructions are shared, with no attachments or other messages. The assistant must already be in the conversation, and running the same message again doesn’t start a second request.

If you see Run not confirmed…, the assistant may not be Ready or not in this conversation, or DonutChat may be busy. Check the message’s thread before you try again.

Replies and notices

The reply appears in the thread at once, marked as in progress, and the answer replaces it when it’s ready. It’s posted under the assistant’s name with a BOT label, and everyone in the conversation can see it. See Bot status and Response activity.

If the assistant can’t answer, it posts one of these notices instead:

Notice	What it means
I’m still working on your earlier request. Mention me again after I reply.	Your assistant is still answering an earlier request, or DonutChat is already running as many requests as it allows for you or for everyone. Wait for the reply, then mention it again.
Hosted assistants are paused right now. Mention me again later.	DonutChat has paused hosted assistants. Try again later.
I couldn’t reply to this message: the request to your AI model provider failed or timed out. Check your API key, its billing/credit balance and model access, then mention me again.	The request to your provider failed or took too long, or the answer was longer than 16,000 characters. Check your key, your provider billing and the key’s access to the model. If you asked for a long answer, ask for a shorter one.
I stopped before replying because this bot or conversation changed. Mention me again if you still need an answer.	The request was stopped before it finished (see below).
I couldn’t finish this reply. Mention me again to retry.	The request ended without an answer.

The first two notices appear only when you mention the assistant. Run personal bot shows Run not confirmed… instead.

There’s no Stop button. A request stops before it finishes if you pause or change the assistant, switch off its bot’s Active setting, turn on an event webhook for its bot, delete the message you asked in, leave the conversation, remove the assistant from it, or lose hosting access.

Pause, change or delete

Open the assistant’s settings:

iPhone: open the assistant from My Bots, then tap Model & connection > Manage hosted assistant.
Android: open the assistant from My Bots, then tap Model & connection > Set up or manage Hosted Hermes.
Desktop app: in My bots, open Manage bot under the assistant and click Assistant settings. Use the buttons in the dialog that opens. The Pause button under Manage bot switches off the whole bot instead, and the assistant doesn’t answer until you click Activate there.

Task	iPhone	Android	Desktop
Pause	Manage assistant > Pause assistant	Advanced · pause, resume or remove > Pause	Pause
Resume	Manage assistant > Resume assistant	Advanced · pause, resume or remove > Resume	Resume current setup
Change the key, model, instructions or tools	Pause, then open Change model, connection or behavior	Pause, then repeat the setup steps	Change API key or model
Delete the setup	Danger zone > Delete hosted configuration, then Delete	Advanced · pause, resume or remove > Delete configuration, then Delete	Remove assistant > Delete configuration…, then Delete configuration

Pause stops new requests and any request in progress.
Resume restarts the assistant with the same key, model and settings. Your access must still be approved.
While its bot has an event webhook turned on, the assistant doesn’t answer, and you can’t set it up or resume it. Turn the webhook off to use the assistant again.
Any change means running setup again with your API key, even if the key stays the same. In the desktop app, Change API key or model pauses the assistant for you.
Deleting the setup removes the assistant’s settings and your saved API key. The bot itself stays, with its token and chats; to remove it too, delete the bot.

What your assistant can access

Each request sends your provider your instructions and the text described in Mention it. It runs on your API key, under your provider account.
Other people’s messages are never sent, even the first message of a thread if someone else wrote it. Neither are attachments or anything from other conversations.
Each request starts fresh. The assistant doesn’t remember other threads or conversations.
Other people can’t use your assistant or your API key, because only your requests start it. They can read its replies.
If you delete your DonutChat account, your assistants’ settings and saved API keys are removed and any request in progress stops.

Limits

Item	Limit
Providers	OpenAI and Anthropic API keys
Assistants	The number approved for you
Access request reason	1–1,000 characters
Instructions	16,000 characters
Request	16,000 characters. A longer message doesn’t start the assistant. In a thread, older messages are left out to fit.
Reply	16,000 characters
Requests at a time	1 per assistant. DonutChat also limits how many run at once for each person and overall.
Attachments	Not sent to the assistant