Everything Argus does,
and exactly how to reach it.

Type /tutorial and press Enter. It is the first entry in the palette, so / then Enter also gets you there.

It is offered automatically the first time Argus runs, before any key has been set.

• Enter on the empty bar turns the page — the only place an empty bar means anything
• /tutorial back goes back one; from page one it leaves
• /tutorial 7 jumps straight to a page
• /tutorial exit leaves without closing the bar, so you can try something

The tutorial never calls a model. It runs entirely in the bar, which means it works at the one moment it matters most — a fresh install, no API key, someone wondering what they just installed.

It only ever describes. Nothing in the tutorial moves your mouse or types for you — demonstrating Agent Mode by taking control would be doing the exact thing the reader has not yet agreed to.

Press Alt + ` and just type a question. Anything that reads like one — a question mark, an opening question word, or a verb aimed at Argus rather than the machine — is answered rather than performed.

The bar opens beside your pointer, where you are already looking. Near an edge it flips to the other side of the cursor rather than covering the thing you asked about. Drag it anywhere and it stays put.

If a question reads like an order and gets routed to the agent by mistake, prefix it with ask.

• Errors you cannot parse. A stack trace, a Windows error code, a red squiggle in an IDE — without retyping any of it into a chat window
• Dense screens. A spreadsheet, a dashboard, a legal PDF: "which of these numbers is the one that changed?"
• Another language. An interface, a document, a menu — it reads what is rendered, so nothing has to be selectable
• Follow-ups. The screenshot stays attached to the conversation, so "and the one below it?" works without re-capturing

The screenshot lives in RAM and is dropped when you dismiss the bar. Saved conversations keep the text, never the image.

Type what you want done. Acting is the default: anything that is not recognisably a question goes to the agent, so there is no magic word to remember.

Click the Talk chip to force it, or prefix with agent when the wording is ambiguous.

• Fiddly sequences you know but hate. Six clicks buried in a settings tree you have to rediscover every time
• Apps with no API. Legacy desktop software, an intranet portal, anything where scripting was never an option
• Reading something out of a UI. Ask it to check or summarise and the answer comes back — not a description of the clicking

Click, type or scroll and the agent stands down between steps and gives you the machine back. Go quiet for three seconds and it resumes where it left off. The frame turns from amber to grey so you always know who is driving.

Hold Esc to stop it outright, mid-movement.

• Fourteen actions per task. A confused agent cannot grind forever
• It notices when it is going in circles — the same action twice earns a correction, three times ends the task honestly rather than burning the remaining steps
• Two minutes of you working abandons the task, because the screen it planned against is gone by then
• Follow-ups work: the last few tasks and how each turned out are carried into the next one, so "open it in Edge instead" means something

Start your request with teach me, then click the chip to Agent so it points at your screen rather than writing out steps.

Left on Talk, the same words give you a numbered written walkthrough instead — useful when you want something you can keep.

• Learning software you will use again. Being shown where a control is teaches you where it is; being done for teaches you nothing
• Helping someone remotely. They keep the mouse the whole time
• Anyone wary of handing over control. This is the mode to start with — the worst it can do is point at the wrong button

Teach Mode never touches your mouse or keyboard. The ghost cursor is a drawing on a click-through overlay; every real click is yours. Nothing advances on a timer, so a step stays up for as long as you need.

Press Space or Enter to skip ahead a step, Esc to end the lesson.

Click the microphone button in the bar. Click it again to stop. The bar turns red and the mic pulses while it is live.

"um, open youtube and uh search for mister beast" becomes "Open YouTube and search for MrBeast".

Fillers, false starts and misheard brand names are cleaned up as part of transcribing, so you get what you meant rather than a literal transcript.

• Never a wake word. Nothing listens in the background — recording only ever starts from that click
• Esc discards a recording, and one stops itself after 60 seconds so a forgotten mic cannot stay open
• Talk Mode sends straight away; Agent and Teach put the text in the box and wait for Enter, because speech misheard by one word should not move your mouse
• Argus never speaks back. This is dictation, not a conversation

You do not have to type /recall. While memory is on, a question that could only be about the past is answered from the timeline automatically.

• The dialog you dismissed too fast. The single most common reason to want this
• A tab you closed — the price, the order number, the phone number you did not write down
• Standup notes. "What have I been doing" produces a short chronological list of apps and pages
• Where did I see that? When you know you read something in the last ten minutes but not where

• Every frame is dated. The answer says when it saw the thing — "3m ago, in the Chrome window" — so you can go back and check
• It quotes, it does not paraphrase. Error codes and prices come back character for character, and "it is not in the last 10 minutes" is a valid answer. A plausible invented error code is the one failure you could not detect
• The live screen is always included, so "is that still up" and "what changed" have a now to compare against

Run an Agent task. When it succeeds, the result offers to keep it.

Typing the workflow's name on its own in Agent Mode replays it too.

• The same setup every morning. Open the four apps, land on the right tab
• Anything you have already watched succeed once. A replay is a shortcut, not a scheduler
• Staying inside the free tier. A replay costs nothing and finishes in seconds instead of a minute

• It records the agent's actions, not yours. Watching you would mean a keylogger running all day, in an app whose whole promise is that it is not watching
• Only successful actions are kept, and a replay stops at the first failure rather than typing the rest into whatever is now on screen
• It refuses to run on a differently shaped screen. 1080p to 4K is fine; 16:9 to ultrawide moves every control, and a blind click would land on whatever now occupies that spot

• Gemini Flash — free tier, 20 requests a day, no card. The default
• Claude Sonnet, Opus or Haiku — paid; reads cluttered screens and small text best
• GPT-5 and GPT-5 mini — paid
• Ollama — free and offline, running on your own machine

• The menu marks which rows are ready and which need a key before you pick one, rather than letting you find out a request later
• An ambiguous name is refused, not guessed. /aimodel claude is fine; /aimodel flash matches two models and says so
• Agent and Teach Mode run on Gemini regardless: they make a dozen quick "which control next" calls per task, which is a different job from one answer worth waiting for

Get a free Gemini key at aistudio.google.com/apikey — no card required.

• You never have to say which provider a key belongs to. sk-ant-…, sk-… and AIza… each route themselves, and Talk Mode switches with them
• Keys stack rather than replace. One over quota hands off to the next automatically
• Quota is per Google Cloud project, so a second key only buys headroom if it comes from a different project
• Paste a bare key into the bar without /key and it is still recognised — never sent to the model, never written to the transcript

• Text is saved; screenshots never are. A resumed conversation remembers what was said about a screen, not the screen
• /new clears what the agent remembers too, so a follow-up cannot quietly continue a task you thought you had cleared away

If Windows refuses to give up the combination, Argus says so and keeps the one that works rather than leaving you with no way in.

• The obvious pick, Win + `, is already Windows Terminal's — a hook can see the key without being able to stop the other app receiving it, so it would open a terminal too
• Ctrl + Space is avoided as well: it toggles IME language input and loses to the OS
• Argus also lives in the system tray — click the icon to open the bar

• demo when you are screen-recording, or showing someone what the agent is doing
• instant when you have stopped watching and just want the task finished