Chats

​

What is a chat?

A chat is a conversation with your Chief of staff, the agent that runs your workspace. You ask in plain language, “which prompts lost citations this month?” or “write a comparison page for our pricing”, and it does one of three things: answers from your tracked data with the numbers behind it, hands the work to the project whose crew does that kind of work, or reports what was done. Each chat is a session you can reopen and continue. The Chief of staff commits to the work, never to an outcome. What an answer engine does with a page is not ours to promise, so it tells you what was done and what it changed, not what it will achieve. Attensira guarantees the volume of work, not visibility. Chats are scoped to one Workspace. The agent answers from that workspace’s Prompts, Topics, Models, Answers and citations, and cannot read another workspace’s data. You start a chat from the box on Home, or with Cmd-N (Ctrl-N) from anywhere. Your recent chats are listed in the sidebar, pinned ones first, and All chats opens the full list.

​

The box

The same box starts a chat on Home and replies inside one. Enter sends; Shift-Enter starts a new line. A message of several lines is sent as one message.

​

Modes

Type / at the start of a new chat to pick a mode. The menu lists the modes first. A mode holds the whole chat to a set of tools, so what you asked for is all it can do: Typing the mode works too: /plan rewrite the pricing page starts a Plan chat about rewriting the pricing page. The mode shows as a chip in the box; Backspace in an empty box removes it. Modes are chosen when a chat starts, so a reply inside a chat does not offer them.

​

Adding things

The + button offers three things:

• Add a link puts the address into your message, adding https:// if you left it off. The agent reads it as part of what you wrote.
• Add file and Add from a connector (also @ in the box) add a chip for a file, a Topic, or a connected integration.

The / menu also lists your skills under the modes.

​

Dictation

The microphone in the box dictates: press it, speak, and the words are added to the box as you go; press it again to stop. Nothing is sent until you send it. Dictation uses your browser’s own speech recognition, so your browser’s speech service does the transcription, not Attensira. Where the browser has none, the microphone is shown switched off with “Dictation isn’t available in this browser”. If the microphone is blocked for the site, missing, or the speech service cannot be reached, the box says which.

​

Output

On Home, an Output pill under the box says what you would like back: Let us pick, Article, Page change, Report, Social post or Image. A chat started from Home does not use it yet, so say the format in your message as well: “as an article”, “as a LinkedIn post”. When the Chief of staff hands that work to a project, it passes the form on, and the crew delivers it that way.

​

When a chat cannot start

What you typed always stays in the box:

• If the credit balance is spent, the box says You’re out of budget · Top up, which opens Billing. See Credits.
• If the chat could not be started for any other reason, you see Couldn’t start. Try again.

Other screens sometimes offer a question for you to ask. It is put in the box and waits there: nothing is sent until you send it.

​

There is no model to pick

Attensira picks the LLM that runs each turn of a chat. There is no model or effort picker in the app: choosing the right model for a question is our job, not a decision you should have to make before asking it. A short lookup runs on a quick model and a hard analysis on a capable one, and the credits a turn spends follow what it actually used.

​

An agent model is not a Model

Two different things get called “model”, so Attensira keeps them apart:

• A Model is a surface you track: ChatGPT, Perplexity, Claude, Gemini, Google AI. It is the thing being measured. Each has a stable channel id that survives the underlying model being swapped out.
• An agent model is the LLM that answers you in a chat. It changes how the agent reasons and writes, and nothing about your data.

Every agent model reads the same stored runs, so a different one would not give you a second opinion on your numbers. If a number looks wrong, check n and the tracked flags.

​

What can the agent reach?

The agent reads the same measurements the dashboard shows: your Prompts and the Topics they sit in, the Models tracked for the workspace, the Answers those models returned, the Mentions and Citations inside them, and the Sources those citations roll up into. It can also act: adding prompts, adding a competitor, creating or running Automations, drafting work for your approval, and handing work to a project. It cannot reach anything you have not tracked. If a Model is not configured for the workspace it was never queried, so there is no data to summarise; the agent should say that rather than report zero. If a Prompt does not exist, no answer about it exists either.

​

When the agent asks you something

Some work cannot be done well without something only you know: who a piece is for, which page, which of two directions. The agent then stops and asks, before it starts rather than halfway through.

• At most three questions, each with up to four one-click options. You can always type your own answer instead.
• It asks everything it needs at once, never one question at a time, and does not ask what a tool can find out or where a sensible default exists.
• The chat waits for your answer, for up to 7 days, then carries on from where it stopped with your answers in hand.

Asking for permission is separate. When the agent wants to do something your approval mode says needs a yes, you get an approval card, not a question. Automations never stop to ask questions: nobody is in front of a scheduled run to answer.

​

Handing work to the agents

Ask for work rather than an answer, “write an article on X, fix the meta descriptions on the pricing pages, pitch two directories”, and the Chief of staff hands each piece to the project whose crew does it, one task per line, instead of doing it inside the chat. A card shows where each line went. See Delegation. When a chat turns out to be about ongoing work, something needed every week or a channel to build, the Chief of staff may offer a Make this a project card: a name, what it is for, the crew it needs, and what wakes it up. Nothing is created until you press Create project; Not now dismisses it.

​

Asking a good question

​

Costs and limits

Asking the agent spends Credits, because each question runs a model and may run tools on your behalf. A long analytical question costs more than a short lookup. A chat that triggers an Automation run spends credits for that run as well, and work handed to a project is metered in that project’s own session. Some questions take long enough that the agent returns a running status and finishes in the background; reopen the session to read the completed answer. The same behaviour is exposed over MCP through ask and get_session. When a chat stops to ask you something, it counts in the Inbox badge in the top bar and is listed under Needs you on Home until you answer. See Inbox.