Instructions: the complete Supercomputer guide (desktop)

This is the desktop version of the Supercomputer manual. Phone and tablet visitors are automatically shown the mobile version instead, because the two layouts differ enough that a single guide would mislead half its readers.

The Supercomputer is the main working surface of AIR. You reach it from the home page with Open Supercomputer, or directly at /supercomputer. Every other part of the product — projects, files, voices, avatars, exports, skills, approvals — either begins there or is one click away from it.

The core idea is that you describe an outcome in ordinary language and AIR works out which engine, which specialist and which tools are required to finish it. Nothing below is mandatory. Every control exists for the moment you want to override an automatic decision, not because the product needs configuring before it works.

The map of the screen

On desktop the Supercomputer is a three-part layout.

The left sidebar is your navigation and history: account summary, credits, cloud usage, then Projects, AI Tools, My Files, View Supers, and a collapsed Advanced section. It stays open on wide screens; the ☰ button toggles it when you want more room.

The centre column is the conversation. When it is empty you see the entry state: the animated wireframe sculpture, a greeting with your name, a rotating line of example outcomes, and a refresh icon that shuffles those examples.

The floating header sits above the conversation and carries the model picker in the top right corner. It floats so the picker is reachable no matter how far you have scrolled.

Below the conversation is the composer, with the Live Web globe, microphone and send button on its right edge, the plus (+) and Apps buttons on its left, and a black strip underneath holding Agents, Skills and Settings.

And everywhere, ⌘K (Ctrl+K on Windows and Linux) opens the command palette — a single search box over every screen, feature and recent conversation in AIR. If you learn one shortcut, learn that one.

The model picker (top right corner)

Click the pill in the top right of the floating header to open the engine list. While the picker is closed it shows only the engine name; credit estimates appear once it is open, which keeps the header quiet during normal work.

Air Max is the default and the first entry, and it is not a single provider model. It is AIR's orchestration layer: the message is understood first (what is the real goal, what is missing, what would 'done' look like), then planned into ordered steps, then executed across whichever models fit each step, then reviewed by a critic pass before it reaches you. That is why Air Max can be slower than a raw model on a genuinely hard request and faster on a trivial one — it sizes the work to the question.

Below Air Max are the direct engines. The visible set covers the common cases: Gemini 3.6 Flash for speed, Gemini 3.1 Pro for long structured reasoning, GPT-5.5 Pro for careful writing and analysis, Claude Opus 5 for the hardest reasoning and longest documents, Claude Sonnet 5 as the balanced daily driver, and Gemini 3.1 Flash Lite for very cheap short turns.

Click More models to expand the rest: the GPT-5.6 family (Sol, Terra, Luna), GPT-5.4 Mini, Gemini 2.5 Flash and Gemini 2.5 Pro. They are collapsed because choosing between them rarely changes a single answer — it matters over hundreds.

Each row carries a one-line blurb and, with the picker open, an estimated credit range for the message you have typed. The range moves with your message length because cost follows tokens in and out.

Practical advice: stay on Air Max. Switch deliberately when you are comparing engines, when you want a specific model's writing voice, or when you want to force the cheapest possible turn. Your selection persists across messages and sessions.

Response modes and thinking depth (Settings on the black strip)

The gear icon on the black strip opens the controls that shape the answer rather than pick the engine.

Response mode has four settings. Auto sizes the answer to your message and is the default. Flash returns a single short line and is the cheapest, fastest option. Standard returns a balanced, well-structured answer. Deep Research returns a long, expert-level answer and is the only mode that runs the iterative research loop — several rounds of searching, reading, scraping and cross-checking rather than one pass.

Thinking depth is separate: Low (quick and direct, least reasoning, cheapest), Medium (balanced, the default), High (deepest step-by-step reasoning, and more expensive because it generates more internal tokens). Because the two controls are independent you can ask for a short answer that was thought about hard, or a long answer produced quickly.

The same panel holds the Air Max controls — tone, explanation depth, and whether citations are included by default. These are enforced on the server for every message, not merely suggested to the model, so 'always cite sources' or 'plain language, no jargon' genuinely holds until you change it.

Memory lives here too. You can review what AIR has remembered, scope a fact to this Super or to all Supers, and delete anything you do not want kept. Recall is semantic: only the facts relevant to the current message are injected, not the entire history.

Agents — the specialists (black strip, first button)

Click Agents on the black strip to open the specialist panel.

There are six specialists and one coordinator. The Planner Agent turns a goal into an ordered plan. The Research Agent gathers and verifies external facts. The Analyst Agent handles numbers, comparisons and trade-offs. The Writer Agent produces the finished prose, deck or document. The Outreach Agent writes anything addressed to another person — email, DMs, pitches, replies. The Critic & Verification Agent checks the work against the original goal and marks what is unsupported. The Workflow Coordinator chains several specialists when one is not enough.

Two things matter here. First, you normally do not need to choose: AIR detects the right specialist from your message, and the reply's routing card shows which one ran. Second, when you do choose manually, the specialist's instructions apply to every engine, not only to Air Max — Research Agent with Claude Sonnet 5 selected gives you Claude working under the research specialist's rules.

A selected agent stays active (the button highlights) until cleared. Clear it when you change topic, or you will get a research-shaped reply to a writing request.

Skills — packaged research playbooks (black strip, toolbox icon)

The toolbox icon opens Skills: pre-written playbooks that define the shape of the output so you do not have to describe a methodology in your prompt.

The eleven built-ins are Competitor Analysis, Market Landscape Mapping, Company Intelligence Reports, Customer & User Research, Feature Gap Analysis, Go-to-Market Research, Pricing Analysis, Technology & Stack Research, Trend & Opportunity Research, Strategic Recommendations, and Interview me.

Interview me is the underrated one. Rather than answering immediately, AIR asks targeted questions first and then runs the work against your real situation. It usually beats a long, carefully engineered prompt.

Each skill has a question-mark button that expands a one-line explanation before you commit. An active skill highlights green and applies to your next message; Clear in the panel header switches it off.

Your own skills live at /skills, where you can record a sequence you liked, save it, and schedule it to run on a rhythm. Additional community-style skills and workflows can be installed from /marketplace.

Apps and attachments (left of the composer)

The four-blocks Apps button is how you get a real file instead of a chat reply: Create PDF, Create Word doc, Create Google Doc, Create document, Create PowerPoint, Create Google Slides, Create Excel sheet, Create Sheet, Create CSV. Choose the format, give the topic, and you receive a finished downloadable file — not an outline, not instructions for making one yourself.

The honest ceilings: roughly 90 pages for PDF and Word, about 90 slides for a deck, around 6,000 rows for a spreadsheet, about 10,000 rows for CSV. Ask beyond that and AIR produces the largest coherent version rather than failing quietly.

The plus (+) button attaches files, and on desktop you can also drag files straight onto the composer. AIR reads PDFs, Word documents, spreadsheets, decks and images and answers about their contents. Any reply built on an attachment or the web ends with clickable sources so you can verify it.

Live Web, voice input, and stopping a reply

The globe inside the composer toggles Live Web. On, it turns blue and reads Live, and replies draw on fresh sources from across the internet rather than training data alone. Use it for anything time-sensitive: pricing, news, availability, current versions. AIR also detects time-sensitive questions and reaches for live data on its own, so leaving it off is not an error.

The microphone is voice input — click, speak, click again. It writes into the composer so you can edit before sending. That is a different thing from the Voice Generator in AI Tools, which produces spoken audio files with ElevenLabs voices.

While a reply generates, the send button becomes a black square; click it to stop. Stopping keeps what has already been produced, and you are not billed for output you never received. Failed steps, failed 4K renders and failed translations refund automatically.

Inside a reply: what the cards mean

Replies come with structure, and every part of it is clickable.

The routing card names the engine and specialist that actually handled the message. It is there so that when a reply disappoints you, you know exactly what to change.

The reasoning trace expands into the steps AIR took: what it understood, what it planned, what it searched, what it verified.

Sources appear at the end of any reply that used the web or your files — real links, not citations in name only.

Next actions are the chips under a reply: one-click follow-ups such as converting the answer into a document, going deeper on one section, or producing the email version. They are derived from that specific answer, so they differ every time.

Goal cards appear when your message described an outcome rather than asked a question. They show milestones, recommended connectors, and an estimate of credits and time to completion. Accept, edit or ignore.

Image results carry a Retry in 4K button that reruns the render through the full high-resolution pipeline. It is charged once, and refunded if the render fails.

The left sidebar — Projects, Tools, Files, Supers

Projects opens the goal-driven workspace: each project has its own tasks, its own memory, a timeline of what happened, a health card that surfaces what has stalled, and a suggested next step. An empty Projects screen offers one-tap starter goals rather than a blank form.

AI Tools opens the drawer of generators: image generation and the Image Intelligence pipeline, Voice Generator, AIR Reels and avatar video, video translation, talking photos, sound effects and the rest.

My Files holds everything you uploaded and everything AIR produced. View Supers is searchable conversation history.

Advanced is collapsed by default and contains Brand Brain (your tone, audience and product facts, applied to every generation), Power tools, connectors, and developer controls such as rate limits.

Right-click — or hold for three seconds — on any conversation in the history list to rename it, add it to a project, export it, or delete it. The hold is deliberately long so scrolling never triggers it.

⌘K and the All features hub

⌘K (Ctrl+K) opens the command palette. Type a few letters of any screen, feature or recent conversation and press Enter. It is the fastest path through the product on desktop, and it covers everything.

If you would rather see everything laid out, the All features hub at /menu is the same set of destinations as a single grouped page with a search box: Create, Work, Intelligence, Trust and Account. It is linked from the Supercomputer sidebar and from the site's hamburger menu, so nothing in AIR requires a keyboard shortcut to reach.

The full destination list: Workspace, Projects, Mission Mode, Jobs, Activity, Library, Templates, Marketplace, Skills, Approvals, Connectors, Usage, Digest, Capabilities, Status, Observability, Privacy Center, Trust Center, Referrals, Account and Blog.

A quick tour of the less obvious ones. Mission Mode runs a long multi-step goal end to end with a live plan surface. Jobs is the unified queue for anything running in the background — long documents, renders, scheduled skills — and it keeps working when you close the tab. Approvals is where actions with real side effects wait for your explicit yes, with an option to always approve a specific kind. Connectors shows what is linked and what is genuinely available. Usage shows credits spent by type, with history. Digest is your weekly report: what finished, hours saved, week-over-week movement. Status shows live model health and latency. Privacy Center is where you export or delete your data. Trust Center is the public-facing summary of how AIR handles it.

A first hour that actually teaches you the product

Send one plain sentence on defaults, then open the routing card and the reasoning trace to see what AIR actually did.

Click a Next action chip instead of typing a new message. This is the habit that changes how the product feels — most of the value is in the follow-up rather than the first answer.

Run the same topic through Apps → Create PDF, so the difference between an answer and a deliverable is concrete.

Open Skills, choose Interview me, and answer its questions. Compare the result with your first attempt.

Press ⌘K and jump to Projects. Turn what you were doing into a project — that is the point at which AIR stops behaving like a chat and starts behaving like a workspace with memory.

Honest limits

Live Web reads public pages and cannot see anything behind a login you have not connected. Deep Research is slower by design, and it deliberately does not trigger for straightforward tool requests such as 'make me a voiceover', because research there would only spend credits for no benefit.

Credits are charged for delivered work. Failures refund, stopped generations are not billed for output you never received, and everything is itemised at /usage.

Free plans use generous daily allowances rather than one monthly wall: the daily cap is a share of the monthly allowance, so one heavy day cannot consume your month. When you reach it, AIR says so plainly and invites you back tomorrow.