HUBAL › User manual › AI Assistant
AI Assistant
The HUBAL system is integrated with an advanced LLM-based assistant, named Agent Brain, which can analyze project data in real time.
Activation
Click the Assistant icon () in the sidebar — the chatbot panel will slide out from the right side.
What can the Assistant do?
The full list — split into the chat and the voice chat, with who can use what — is in the section What the assistants can do.
Saving to the Drive and transcribing handwritten notes
The assistant can save a Word file (DOCX) for you in your private Drive — e.g. a handwritten note transcribed from a photo, a transcript of a conversation or a longer text from the chat. This is different from a note in the widget: you get a regular file that you can open, download and forward.
In the assistant chat (also in the mobile app) type or dictate the content, or attach a photo of handwritten notes with the attachment button.
Use the words "to the Drive", "to my private space" or "to the file manager", e.g. "Transcribe this note and save it to the Drive in my private space as Meeting minutes". If you don't give a file name, the assistant picks one.
The assistant shows a card with a single step (saving the file) — click Run. Nothing is saved without it.
The assistant replies that the file is in Drive → Private drive and gives its name. Open the Drive icon in the sidebar, go to the private drive — the file has a preview, download and the same options as any other file.
What the assistant recreates from handwritten notes. It does not just transcribe the words; it tries to reproduce the layout of the page:
- a hand-drawn table → a real Word table with the same rows and columns,
- underlined words and sentences → underlined in the document (a single word inside a sentence too),
- titles → headings, bullets and numbering → lists,
- crossed-out text is skipped, and fragments that cannot be read are not guessed — they are marked [nieczytelne] (unreadable).
Shoot straight from above, in good light, the whole page. With a poor photo the table may come out incomplete — then the file shows [nieczytelne] where the assistant was not sure. Check the document before forwarding it.
By default a Word document is always created — like every document in HUBAL. You get a plain text file (TXT or Markdown) only when you explicitly ask for one. A document for the project canvas is a separate request ("make a document on the canvas of project X") — the assistant never saves one text in two places at once. Saving to the Drive works in the chat and in the mobile app; the voice chat does not do it yet.
Connect your own AI assistant — Claude, ChatGPT, Cursor
If you work in Claude, ChatGPT or Cursor every day, you do not have to give them up. Connect them to HUBAL using the MCP standard and your assistant gets the same tools as the assistant in the app: tasks, projects, deadlines, mail, procedures, finances — always on real data and always within your permissions. Available on the Start and PRO plans.
In the app open Settings → Integrations and copy the MCP server address (e.g. https://hubal-project.com/mcp).
Claude: Settings → Connectors → Add connector and paste the address. ChatGPT: Settings → Connectors → developer mode → Create and paste the address. Cursor: MCP settings → add server with this address. No key is needed.
The assistant redirects you to a HUBAL page: you sign in to your account, see who is asking for access and to which organisation, and click Allow. An administrator can instead connect the assistant as a virtual agent (a separate identity in the register; PRO plan).
You ask your assistant about tasks, deadlines, mail or procedures and it fetches the data from HUBAL. Every write (a new task, a financial entry, an e-mail) arrives as a proposal to approve in the Assistant panel in the app — nothing is saved without your click.
Settings → Integrations → Connected assistants shows what is connected and when it last worked; Disconnect takes effect immediately.
Exactly what you see: taking away your role or your access to finances applies immediately to the connected assistant too, because the server computes permissions on every call. The data the assistant asks for goes to its provider (Anthropic, OpenAI, Cursor) under your agreement with that provider — HUBAL is not a party to it.
So that a loop in someone else's assistant cannot eat the pool, call limits apply: per connection and per organisation (higher on PRO). AI tools called by the assistant use the organisation's credits just like in the chat. On the PRO plan an administrator can additionally create a virtual agent (its own name in the change history) and issue an access key for tools that cannot sign in (scripts, agent platforms).
Writing emails with the assistant
The assistant can draft an email sent from your own mailbox — but it never sends it on its own. Sending happens only after you click "Send".
Step by step:
— e.g. "write an email to John Smith asking him to send the signed contract".
— it searches your address book (built automatically from all your correspondence) and the organization memory. If several people match, it shows a list to choose from; if it finds no one, it asks you for the address.
— with the recipient, subject and the full message body to review.
— every email sent by the assistant automatically gets a short footer noting it was prepared by the HUBAL assistant (you see it in the preview). It goes below your standard signature from Mail settings.
— you click "Send" (the email leaves from your mailbox and lands in the "Sent" folder) or "Reject" (nothing happens).
The AI Assistant respects privacy settings. It does not process data from confidential areas without appropriate permissions. More details in the Collaboration and Security section.