Feature

AI chat

Chat with AI on your own, or bring AI into a conversation with teammates by @mentioning it.

TL;DR

Start a solo chat with AI, or invite teammates and pull AI in with @Tana when you want a response. The AI reads and searches Tana, creates and updates documents, connects to GitHub, Slack, Linear, Jira, and HubSpot, and generates images. Reference any document with @, invoke skills with /, and speak instead of typing. Every AI change goes through proposals you review before it applies.

Solo chat or team chat

Every chat in Tana is either a solo conversation with the AI or a conversation with teammates. The two work the same way, with one difference:

  • Solo chat: You and the AI. The AI responds to every message.
  • Team chat: You, teammates, and optionally the AI. The AI stays quiet by default. Mention @Tana when you want a response, so human conversation is not interrupted.

In both cases, the AI has the same capabilities: reading your content, creating and updating documents, invoking skills, and connecting to your integrations and custom MCP servers.

Starting a chat

There are several ways to start a conversation with AI in Tana:

  • Calendar: The chat bar on the Calendar takes what you are looking at as its context, the visible range or the meeting you have open.
  • Space home: The “Ask Tana or start a chat…” bar on every space home. Type a message and press enter to create a new chat in that space, addressed to its default agent.
  • Home and Work: The Ask or capture… bar at the bottom of Home and Work starts a chat with Tana.
  • Meetings: Every meeting has two chats. Agents is yours alone, and Tana, or your own Claude or ChatGPT, answers every message there. Meeting is read by everyone in the meeting, guests included, and an @Tana question asked there is answered privately to you until you share it. See Chat in the call.
  • Quick chat: Press Cmd+LCtrl+L on the desktop app to open a new chat from anywhere. Press Cmd+Shift+LCtrl+Shift+L to open it with voice dictation active.
  • Create menu: Select “Chat” from the create menu.
  • Documents: Every text document has a chat bar at its foot, “Ask or change this document…”. The document goes along as context, so “add a section about X” comes back as a proposal on it.
  • Send to: The Tana tile on an item’s Send to menu starts a private chat with the item attached.

Your chats are listed on the Chats tab of Work.

Composing a message

Mentions and skills

Reference documents and invoke AI capabilities directly from the message input:

  • Pick which agent answers from the picker in the compose bar. Chats run against the built-in Tana agent unless you or the space set another. See Agents.
  • Type @ to search and reference any document in Tana. With nothing typed yet, it lists your Recent documents first. The AI receives the referenced document as context. Mentioning @Tana triggers an AI response in chats with other people.
  • Type / to search and attach a skill. Picking one runs it straight away, using whatever you have typed as its input, or its own defaults when you leave the message empty. If what you typed is an instruction about the skill itself, the AI answers that instead of running it. The same / also accesses emoji and formatting commands. What you type after @ or / stays in the line and narrows the menu.
  • Type : to open the emoji picker (about 250 common emojis searchable by shortcode). Press : again to close the picker without inserting anything.

Reactions and replies

Hover a message, or right-click it (long-press on a phone), to react with an emoji or to Reply. A reply quotes the message in yours; click the quote to jump to the original, and a reply directly under its original shows a connecting line instead. Press your own reaction to take it back. Tana sometimes reacts to the message it’s answering, and in a busy chat its answer quotes that message.

Attachments

Add images, audio, video, PDFs, and any other ordinary file, such as a Word, Excel, or PowerPoint file or a zip, by dragging it into the chat, pasting it, or clicking the attach button. Files that can run code, such as programs and installers, are refused. Each attachment shows upload progress while it is sending, and Send waits until uploads finish so nothing is silently dropped. Failed uploads can be removed and retried. A sent file shows as a card with Download. In a document, the /File command adds a file as a card the same way.

What the AI can read depends on the kind of file:

  • Images get AI-generated descriptions automatically.
  • PDFs, audio, and video can be summarized and asked about, and the AI can embed media inline in its replies.
  • Text files of any kind, such as plain text, Markdown, CSV, JSON, code, or logs, are read in full.
  • Word, Excel, and PowerPoint files are not read. They stay attached for people to download.

Voice input

The chat toolbar has two voice modes:

  • Dictation: Click the microphone button to speak your message. Words appear in the input as you speak. Supports English, Norwegian, and many other languages.
  • Audio recording: Click the voice button to record an audio clip, preview it, and send as an attachment. The audio is automatically transcribed.

What AI can do

In Tana

The AI has full access to your content (within your permissions). It can:

  • Read: Documents, types, fields, meeting transcripts and participants, screen capture screenshots, calendar events, skill definitions, previous chats, tables, attached PDFs, and uploaded audio and video
  • Search: Keyword search, semantic search, relationships between documents
  • Create: Docs, tasks, typed entries, types with fields, skills, artifacts, AI-generated images, and calendar events
  • Update: Titles, content, fields, calendar events, space configuration, table data, and type definitions
  • Delete: Any document
  • Link: Build a shareable web link to a document or a meeting, for pasting into another tool or sending to a teammate

Ask for a link when you want one that works outside the chat, in an email, a CRM record, or an issue tracker. The link opens Tana in the browser and the person opening it has to be signed in to your organization, so it is not a public link. Inside the chat you do not need one: documents the AI mentions are already clickable.

When the AI creates a calendar event it needs a title and start time; end time, timezone, location, description, and participants are optional. Passing participants restricts the event to that list with you as admin. It can propose changes to an existing event’s title, description, time, timezone, location, mode, and existing tagline or summary. For a synced event, title and time changes require two-way calendar sync and the connected calendar’s organizer. Attendees and meeting links remain calendar-sourced. Use pinItem to pin documents to an event.

Updates and deletes and creates all appear as proposals for your review. AI never changes anything without your approval.

Connected tools

When you connect integrations, the AI can act on them from chat:

  • GitHub: Search and create issues, review and merge pull requests, read code and files, comment on issues and PRs
  • Slack: Send messages to channels or people, list channels and users
  • Outlook: Search and read your mail, draft a reply, and send a message once you accept the action that proposes it
  • Linear: Search, create, and update issues
  • Jira: Search, create, and update issues, with meeting screenshots embedded as visual context
  • HubSpot: Search contacts, companies, and deals, update properties and lifecycle stages, create deals, and log notes to CRM records
  • Custom MCP servers: Use the tools from any remote MCP server you connect, surfaced through the linked skill for that server
  • Coding tools: Hand off the current document or chat as a self-contained prompt to Claude Code, Cursor, Codex, GitHub Copilot, or Lovable. Triggered by phrasings like “create a PR” or “send this to Cursor”; when more than one tool is enabled, you pick.

Configure integrations in Settings. See Integrations.

Managing access

The AI can also propose changes to a document’s access: restricting it to specific participants with editor or admin roles, opening it back up to everyone in its space or event, and toggling the Guest link (anyone with the link can view). Access changes go through the same proposal approval flow as the rest of the AI’s edits, and the proposal preview shows a clear access callout before you accept.

Default content language

Every organization picks a default content language in Settings, Organization, Language, from 33 languages. The AI writes its replies, generated artifacts, outcomes, summaries, and document titles in that language no matter which language the conversation is happening in. Meeting transcription stays multilingual (auto-detected per speaker) so a mixed-language call still transcribes everyone correctly. The Tana interface itself is English.

Usage limits

Your monthly AI credits are the one limit you can manage. When you have used 80 percent of them, an amber callout appears at the top of chat (“You’ve used N% of your monthly AI credits”) and a notification lands in Work. When your credits run out, chat shows the same style of callout instead of a reply, with a “Resets in...” countdown and, where billing is available, a Top up button that opens Billing settings. Separate safety limits also protect against runaway usage (for example, a looping automation); if one trips, the callout shows a reset countdown but no Top up button, because a purchase cannot clear a safety limit. A reply cut off mid-stream ends with a Usage limit reached callout.

Web search and image generation

Models that support web search can look up current information when your question goes beyond Tana. Results include citation links to sources. Web search availability varies by model.

The AI can also generate images from text descriptions. Generated images become image documents owned by the chat they were created in, so an image made in an “Only me” chat stays private.

When a reply includes a Mermaid diagram, Tana renders it inline once the diagram is complete, so you never see a half-drawn render. A malformed diagram shows a compact notice instead, and the AI corrects it on its next turn.

Search results in chat

Ask the AI something that is really a search, “the open decisions from this quarter”, “feedback filed per customer this month”, and the answer comes back as a result card rather than prose. Collapsed it shows a type badge and a count. Opened it is a table of the matching documents, with the attribute columns that type carries. Show more pages through larger sets without reloading the card.

The count reads how many documents matched, not how many are drawn. Where the table shows fewer, a note says so, so a total you take from the rows is a floor rather than the answer.

  • Save it. The card carries its own Save search action, so a search the AI ran becomes a saved search document you can pin, exactly like one you built yourself.
  • Searches attached to a type. A type can carry the searches worth running on it, listed on its panel. The AI can run one of those, or attach one to the conversation.
  • An existing saved search wins. Where a saved search in your organization already answers the question, the AI shows you that one instead of composing a new query.

Reviewing changes

Every AI-generated document, edit, or deletion appears as a proposal you review before it takes effect. Proposals show up as inline cards on the message that produced them, so you can see what the AI did and decide on it in place. Each pending proposal shows a Lives in callout for where the new document will live, alongside an access callout for who will see it, so the consequences are visible before you accept. Accept individually, reject, or use Ask for changes... at the bottom of the proposal to have the AI revise it before you accept. Once decided, a card stays on its message and reads Saved or Replaced.

When the AI assigns a task to you while acting on your behalf, you do not get a notification for your own work. Assigning to someone else still notifies them, attributed to you.

Writing to a connected service works the same way, as an action you approve before anything leaves Tana.

See Proposals and actions for the full review workflow.

Choosing a model

Tana supports models from OpenAI, Anthropic, and Google. Switch models at any time during a conversation. The change takes effect on the next message.

Models are grouped by cost:

  • Cheap: Smaller, faster models for simple tasks
  • Moderate: Balanced cost and capability
  • Expensive: Most capable models for complex reasoning

Some models support reasoning with configurable effort levels (low, medium, high). Some support web search. Both capabilities are indicated in the model selector.

Following an agent’s work

While an agent works, one line above its answer says what it is doing; press it to see each step. The line folds away once the answer starts. A reply that ends without any words keeps a “Worked for” line with the time and its steps.

Clarifying questions

The AI may ask clarifying questions before proceeding with a task. The question and its options sit on the AI’s message, so the message box stays free and an earlier question stays answerable.

  • A short choice is a row of pills. Tap one to answer.
  • A longer or multi-pick question is a card. Pick your options, or use its last row, Something else…, to answer in your own words, then press Send, or Skip the question.

Replying to the message with text answers it too.

Side panels

The chat toolbar opens side panels for:

  • Outcomes: Pending proposals and accepted outcomes from this conversation
  • In common: Documents shared between chat participants. Only visible when the chat has multiple participants.

Agents

Beyond the default Tana assistant, you can build custom agents with their own capabilities, prompt, voice, and schedules. Click the pencil button at the top of the sidebar, describe what you want, and Tana scaffolds the agent for you. You can also ask the AI to find one by what it does, for example “find the agent that does customer research”, and it searches your own agents by name and purpose. See Agents for the full guide.

Reporting issues

If the AI produces an incorrect or unhelpful response, open the chat’s ⋯ menu and choose Send feedback, then describe what went wrong. The conversation goes along unless you switch it off, and you get a reference to quote to support.

When an AI run fails outright, the chat displays a red callout with a one-line, actionable reason: the model rejected the request (quoting its reason, for example an attached image being too large), the response timed out, or a generic “Something went wrong” fallback. A run you stopped shows Cancelled. Transient problems are retried automatically, falling back to another model rather than surfacing as an error.

AI chat - Tana Learn