Every leading model, one conversation
Claude, GPT, Gemini and more work in the same session. One drafts, another checks.
Learn more →
Bring in your current setup, gain instant value
Gmail
Google Calendar
Google Docs & Sheets
Microsoft 365
QuickBooks
Xero
Stripe
Square
Shopify
HubSpot
Mailchimp
Klaviyo
Calendly
Google connects right inside Luminair. Most others connect through the Claude or ChatGPT account you already have.
On Team+, if a tool you use isn’t connected yet, we build the connection within 7 days.
GmailGoogle CalendarGoogle Docs & SheetsMicrosoft 365QuickBooksXeroStripeSquareShopifyHubSpotMailchimpKlaviyoCalendly
Google connects right inside Luminair. Most others connect through the Claude or ChatGPT account you already have.
On Team+, if a tool you use isn’t connected yet, we build the connection within 7 days.
Put each AI on
the work it does best.
Luminair brings Claude, GPT, Gemini and more into one platform, with the same conversation experience you’re used to. They all know your business, so there is no memory loss.
Research
Writing
Coding
Reasoning
Private
Triage
3 in 4 of your queries cost more than they need to. We fixed that.
With Luminair auto routing, you can save up to 41% on token costs.
Progressive disclosure
With Journal, information is organized in layers: a one-line summary, an outline, then the full doc. Each message carries a short digest of what matches it. The AI opens an outline or a single section only when the work needs it.
Model auto-select
Luminair picks the model for each message. A built-in judge sizes the message as light, standard or heavy without calling a model. Light work goes to a lower-cost model you have connected; hard work keeps a top model.
Hard cap
Luminair limits what rides along with each message. A rule file over 24KB is left out, not sent. The working summary of a long session (objective, recent turns, files touched) is cut to a fixed size, so the AI does not re-read the whole project to catch up.
75%of queries automatically switched to a lower-cost model than the default.
Average across 120,000+ real queries. Queries with a hand-picked model are left out.
41%saved on token price when Luminair picks the model.
Each query is sized first, then sent to a model that fits
Light36% of queries$1.59 per 1M tokens. Mostly Sonnet 5 and GPT-5.6 Luna.
Standard57% of queries$2.55 per 1M tokens. Mostly Sonnet 5, Opus 5.5 when needed.
Heavy7% of queries$4.39 per 1M tokens. A top model is kept.
Same 120,000+ queries. Prices are each provider’s published price per million input tokens.
Auto-model, but you stay in control. Luminair picks the model for each message. Change it any time.
Pick for one messageChoose a model before you send. Only that message uses it.
Pick for a whole tabChoose a model for a tab. Every message there uses it.
Turn auto offSwitch it off and Luminair stops picking models.
Connects with the apps your business already runs on
WooCommerceIntercomGmailAirtableZoho CRMFathomXeroHarvestAsanaTrelloZoho Booksmonday.comZoomWooCommerceIntercomGmailAirtableZoho CRMFathomXeroHarvestAsanaTrelloZoho Booksmonday.comZoom
Tools to manage them that you won’t find anywhere else
Run several models on one job, give them all the same memory, queue and branch their work, and check in from anywhere.
All your business information centralized
Keep your prices, tone and rules in one Journal. Every AI and every teammate has them instantly, with nothing to re-explain.
Learn more →
Runs work side by side
Queue requests while a session works. Branch out, and independent ones run in parallel.
Learn more →Goes where you go
Check progress, answer a question, or queue the next request from your phone.
Learn more →Keep going, wherever you are.
A session keeps working after you close the laptop. Open Luminair on your phone or tablet to check progress, answer a question it is waiting on, or queue the next request.
See how Mobile Relay works →
You stay in charge
An app on your own computer
Luminair runs on your Mac, Windows or Linux machine. You choose which folders and accounts each session can use.
Your own AI accounts
Sign in with the AI plans you already pay for, and switch models whenever you like.
You control the models for every query
Pin a model to any query, or let Luminair pick. Every Flow step names its model too.
See where the tokens go
Usage is tracked per model and per session, so cost is never a surprise.
Grow faster. Decide better.
Hire when it actually makes sense.
Luminair gives solo founders and small teams the leverage of a much bigger company.
Your entire business in control, one simple view.
13 features in three groups, then Luminair, which watches over all of them. Skim the headlines, then open any one for the full story.
Saved layouts
Branches
Luminair
Journal, Flows, Team
/luminair
- Journal
- Actions
- Flows
- Team
Active sessions
Active Sessions
Project folders
Folders
- Acme Web App
- API refactor16h
- Fix checkout bug4h
- Landing page copy1h
- Onboarding flow9h
- Marketing Site
- Rate limiting2h
- Q4 Launch
- Cache layer6h
- Signup funnel12h
Session + JournalLanding page copyOpus 5.5 Brand guidelinesJournal · Folder
This is close, small tweaks and it's done.
Tightened it up. Same structure, cleaner lines throughout.
Read it back to me like a customer seeing it cold.
Cold read: it leads with the outcome, not the feature. Someone skimming gets the point in the first line.
Does the tone hold in a deck?
Mostly. I'd drop the contractions for a deck version but keep the structure.
How we sound
Warm, plain, a little dry. We sound like the person behind the counter.
Always
Prices include tax. Name the baker. Say when it came out of the oven.
Price list 2026Wholesale terms
Models
Ask anything, pick any model
Opus 5.5GPT-5.5GeminiHigh
Branch out
Branch out
4 of 4 agents working · 0/11 requests done
4Sub-agents
11Requests
Queue
Queue3
- Also check this holds up on mobile width.
- Add a test for the empty state too.
- One more: write a one-line changelog entry for this.
The workspace
Sessions that live in the cloud, organised by project, forkable, and side by side.
01The workspace
Active Session
See what is working, what is waiting and what is stuck.
Active Sessions lists every session that is doing something right now, each with a status dot and a label. You see what is working, what is queued and what has stopped, without opening a single one.
Every running job, in one glance.
Activity, not open windows
A session appears the moment it starts working and stays until you have seen the result. Close its tab and it keeps going in the background.
A dot and a label for every state
Orange and pulsing means thinking. Green means a message is queued. Red says Error or Stopped in words, so you read the label and not just the color.
Stuck runs show up fast
A run that breaks turns red and says Error, so you find a crash or a rate limit in seconds instead of discovering it an hour later.
One list for everything in flight.
Five sessions across three projects: some thinking, one queued, one stopped. Active Sessions puts them in one list in the sidebar, so you check on all of them without hunting through tabs.
- Lands in the list the moment it starts working
- Stays until you have seen the result
- Closing the tab does not stop the run
Step away without losing track.
One switch on the Active Sessions bar pauses every queue. Replies already in flight finish, manual sends still work, and the pause holds across restarts until you switch it back.
- Pause all queues from one switch
- Press Stop to keep what was written
- Finished runs come back to the list
02The workspace
Project Folders
One folder per client or project, each with its own memory.
A folder in Luminair is not a bookmark. It is a workspace: its own sessions, its own abilities, its own journal notes, its own default model. Open it and everything about that project comes back with it.
Every project keeps its own head.
The folder is the unit of work
Add a folder and Luminair treats it as a project: sessions grouped under it, file tools scoped to it, history kept with it.
Settings travel with the project
Default model, intensity, abilities and connectors are remembered per folder, so a marketing project and a codebase never share the wrong defaults.
Open where you left off
Reopen a folder and its sessions come back in the order you left them, on any machine you sign in from.
Two projects, two brains, no bleed
A research folder can run a cheap fast model with web-focused abilities. A production repo can pin high intensity and a strict "never touch main" rule. Neither one leaks into the other, and you never re-set them by hand.
- Per-folder default model and intensity
- Per-folder abilities and journal notes
- Per-folder session list, kept in order
Drop a folder in and go
Drag a folder onto the window, or open one from the sidebar. Luminair reads it on your machine with real file tools, not an upload into a sandbox.
- Drag and drop, or open from the sidebar
- Real read and write on your own disk
- Nothing uploaded to make it work
Landing page copyDark mode passAuth rewrite
This is close, small tweaks and it's done.
Tightened it up. Same structure, cleaner lines throughout.
Read it back to me like a customer.
03The workspace
Split & Tab View
See two chats side by side, like browser tabs.
Tabs across the top, split panes below. Put two sessions side by side, drag the divider, double-click to reset. The arrangement is yours and Luminair keeps it.
See two answers at once.
Two, three, or four panes
Split as far as the window allows. Every pane is a full session, not a preview.
Tabs that mean something
Tabs are named by session, grouped by folder, and survive a restart. You always know which thread is which.
Drag to resize, double-click to reset
Move a divider to give a pane more room. Double-click it to snap back to even. Auto-tidy evens everything out in one click.
Watch one, write in the other
A long run does not have to hold your window hostage. Let it stream in the left pane while you plan the next thing on the right, in the same project, with the same files open.
- Every pane is live at once
- Collapse the sidebar for more width
- Auto-tidy layout in one click
Compare before you commit
Two branches of the same question, side by side, same scroll. It is the fastest way to tell a good answer from a confident one.
- Pair with Open Branches for A/B answers
- Pair with Multi-model to compare engines
- Save the arrangement as a Layout
BuildSaved just now⋯
Sidebar2 panesEvenly sized
04The workspace
Layouts
Save your screen setup and bring it back in one click.
Set up the panes and tabs you want for a kind of work, save it as a Layout, and switch back to it any time. Research, build, review: three arrangements, one keystroke apart.
Your arrangement, saved.
Save the current arrangement
Panes, splits, sizes, tabs and which sessions are open. Named, saved, one click away.
Switch between them
A research layout with two reference panes. A build layout with one wide pane and a sidebar. Move between them without dragging anything.
Auto-tidy
One click evens out the panes and resets the sidebar when the window has drifted out of shape.
A layout per kind of work
Reviewing is not building. Give each mode its own arrangement and stop rebuilding the same three-pane setup from scratch every morning.
- Name and reorder your layouts
- Restores which sessions were open
- Works with collapsed or expanded sidebar
It follows your account
Layouts are part of your setup, not a local preference file. New laptop, same arrangement.
- Synced with the rest of your settings
- No dotfiles to symlink
- Nothing to export or import
Throughput
Hand over a list, route each request to the right model, and lose nothing on the way.
Queue4Auto-send
- Session modelAlso check this holds up on mobile width.
- GPT-5.5Add a test for the empty state too.
- Opus 5.5Rewrite the onboarding email copy.
- Session modelWrite a one-line changelog entry for this.
05Throughput
Queue
Line up tasks, reorder them, and let them run while you do something else.
Every session has a queue. Type while it is busy and your message waits in line, then runs when its turn comes. Drag items into a new order, pause one or all of them, send one now, or give a single item its own model.
Queue it. Walk away.
Nothing gets lost
Fire the next thought without waiting on the last. It goes in the list, in order, and it will run.
Reorder, pause, play
Drag an item up or down the list. Pause one, or hold the whole queue with Auto-send off, then let it run again when you are ready.
A model per item
Each queued message runs on the session model or on one you pick for it, so a quick check and a heavy rewrite do not have to share a model.
Controls for every way you work
Enter sends. Cmd+Enter jumps the queue, Ctrl+Enter injects your message into the task already running, and Option+Enter parks it paused for later. One switch on the Active Sessions bar pauses every queue at once.
- Auto-send on or off, per session
- Pause all queues with one switch
- Send any item now
A queue you can steer from anywhere
Add three items from your phone on the way home. They are waiting at your desk, in order, ready to run. Same list, every device, no export step.
- Cloud-backed, not machine-bound
- Pauses itself after an error
- Holds when a usage limit hits
06Throughput
Branch Out
Turn a queue into a team of agents working at once.
Branch Out turns a queue into parallel work. A master agent groups your requests, gives each group its own agent on its own branch, then merges the results and hands them to you. Your main branch is never touched.
One queue. Many agents.
A list becomes a fleet
Once a session has more than three queued requests, a Branch out button appears. Press it and the requests are worked in parallel instead of one after another.
Agents that cannot collide
Each agent works in its own isolated copy of the repo, on its own git branch, then merges back like a pull request.
Best of several
Type //branchout x3 and one request is solved three different ways in parallel, then a judge ranks them. Anywhere from two to five ways.
Queue the whole list, then step away
Write every task as its own message, press Branch Out, and leave. Parallel agents clear the list while you are gone. Research questions are answered straight into the chat, and code lands on a side branch for you to review.
- Diff, merge or discard from Branches
- A token budget can cap the whole fleet
- Main is never touched
Guard rails built in
A restore point is created before anything starts. Sub-agents cannot switch branches, push or reach the network. A conflict that cannot be resolved is reported, never forced.
- Restore point before every run
- Hard timeouts on every agent
- Requests with images stay in the queue
07Throughput
Multi-model
Claude, Codex, Gemini, Kimi and more, chosen per chat.
Claude, GPT and Codex, Gemini, Kimi, DeepSeek, Grok, Qwen, GLM, and local models on your own machine. One window, one interface, one set of abilities. Choose per session, and switch mid-thread.
Pick a model by task, not by logo.
Every major model, one window
One integration, 500+ models. Frontier models, open-weight models, and local models through Ollama or LM Studio.
Per session, not per app
Reasoning model on the hard thread, a fast cheap one on the scratch thread. Both open at the same time.
Your key, your cost
Bring your own account. The key stays on your device and you control what gets spent.
Switch mid-thread without losing the thread
Not happy with the answer? Change the engine and ask again in the same session, with the same context. No copy-paste into another app, no re-explaining the project.
- Same transcript across engines
- Same abilities, same folder, same files
- Compare in a split pane, side by side
Abilities route work to the right engine
Type //perplexity for live research, //image to generate, //gemini for a second opinion. The session stays where it is and the right model does the piece it is best at.
- //perplexity, //image, //gemini and more
- Set a default model per folder
- Intensity dial: how hard it should think
Memory & reach
A memory every model reads, reusable instructions, your voice, and your phone driving your desktop.
Brand guidelines
How we sound, what we promise, and what we never say. Every model reads this before it writes for us.
Voice
Warm, plain, a little dry. We sound like the person behind the counter, never like an advert.
Always
Prices include tax. Name the baker. Say when it came out of the oven.
Never
No discounts in writing. No exclamation marks. No delivery promises we cannot keep.
Price list 2026Wholesale termsRefund policy
08Memory & reach
Journal
A notebook your AI reads before it answers.
The Journal is a vault of small notes, one fact each, injected into every message automatically. Rules you set once, project facts, things you told it last month. It is memory that does not depend on the model remembering.
A notebook it actually reads.
Four layers, right scope
Global rules apply everywhere. Folder notes apply to one project. Session notes apply to one thread. Engine notes apply to one model.
Injected, not hoped for
Relevant notes are added to the context on every message. There is no "please remember" and no silent forgetting.
One fact per note
Small, named, editable files. You can read them, correct them and delete the wrong ones yourself.
Rules that survive the session
Tell it once how you want replies written, which branch it may push to, what your stack is. It is there in the next session, and the one on your phone.
- Global rules, set once
- Per-folder project facts
- Per-session and per-engine notes
You own the vault
Notes are plain files in a folder you control. Nothing is locked in a provider's memory store, and nothing is lost when you switch engines.
- Readable, editable, deletable
- Travels between models
- Synced across every device
09Memory & reach
Second Brain
Every document linked to the ones it depends on.
Journal docs link to each other. Your brand guidelines point to your price list, your price list points to your wholesale terms, and the graph draws the whole business as one connected map instead of a pile of files.
Everything you know, connected.
Link as you write
Write [[Price list]] inside a doc and the two are linked. Links are plain Markdown wikilinks, so they also work in Obsidian or any editor.
See what points back
Open a doc and see which other docs link to it, so you know what leans on a price or a policy before you change it.
Pull a doc into any chat
Type [[ in the composer and pick a doc. It goes into your message, whichever model you are talking to.
Docs that explain each other
A quote template links to the rate card. The rate card links to the terms. Open one and the rest are one click away.
- Plain Markdown files on your disk
- Readable in Obsidian or any editor
- One folder per client or project
A map, not a filing cabinet
The graph shows how your notes hang together: which docs everything else leans on, and which sit alone with no links at all.
- Every doc is a dot, every link a line
- See which docs the rest lean on
- Spot notes that link to nothing
//
//perplexityLive web research with citations
//deep-researchFan out sources, verify, write it up
//imageGenerate an image in this session
//briefCustom · your house style, three bullets max
//ship-checkCustom · verify, then deploy, then confirm live
Built-in and your own, side by side. Search by name or by what it does.
10Memory & reach
Abilities
Save an instruction once, reuse it with a short command.
Abilities are the // commands that change how a session behaves: how it writes, what it is allowed to touch. Use the built-in ones or write your own and they are there on every device.
Say it once. It sticks.
Built-in abilities
Route a message to a different engine, generate an image, run live research, change how hard it thinks. All from the message box.
Write your own
Turn a paragraph you keep retyping into a named ability. Call it by name from any session.
Browse and search
Every / command and // ability in one searchable palette, so you never have to remember the exact name.
Reusable instructions, not a bot store
An ability is a saved way of working: the instruction, the intensity, the tools it may use. Yours, private, and available everywhere you sign in.
- Named and callable from any session
- Carries its intensity with it
- Scoped globally or to one folder
The palette knows what exists
Hit the shortcuts panel and search. Every command and ability with what it does, so a new one is discoverable the day you write it.
- Search by name or by what it does
- Built-in and custom side by side
- Insert straight into the message box
Listening…
You saidum so can you like, pull the retry path out and, uh, make it idempotent
Sent after cleanupPull the retry path out and make it idempotent.
Clean up dictationFocus lock while talkingMute the spoken reply
11Memory & reach
Speech to Text
Talk instead of typing.
Hold the globe key or open live conversation and just speak. Luminair transcribes, cleans up the dictation, sends it, and can read the reply back. Focus lock keeps the mic on your session while you keep working.
Talk to it, hands free.
Live conversation
Open the overlay and talk. It listens, transcribes, sends, and speaks the reply back. Mute the voice any time.
Clean up dictation
Filler words, false starts and stray punctuation are cleaned before the message is sent, so the model reads a sentence not a stream.
Focus lock while talking
The mic stays bound to the session you are in even if you click elsewhere, so a long dictation does not land in the wrong pane.
Push to talk, or leave it open
Use the Fn / globe key for a quick line, or open the overlay for a real back and forth. Type mid-conversation without breaking it when you need to paste something exact.
- Fn / globe key for quick dictation
- Overlay for continuous conversation
- Type-bar for exact text mid-flow
It goes into the same session
Dictation is not a separate mode with its own history. It lands in the session you are in, synced like anything else, ready to read on your phone.
- Same transcript, same sync
- Works with abilities and the queue
- Mute the spoken reply, keep the text
In sync12Memory & reach
Mobile Relay
Send work to your computer from your phone.
Relay sends a request from your phone to your Mac and runs it there, with real file tools on your real machine. You get the output on the phone. The desktop mirrors the whole run so nothing happens off the record.
Your desktop's power, from your phone.
Real tools, remotely
A relayed request runs on your desktop, so it can read and write your actual files. Not a cloud sandbox with no access to your work.
No Max-tier gate
Remote control of your own machine is a feature, not a paid research preview. It works on whatever account you already have.
The desktop mirrors it
Everything relayed from the phone appears on the desktop as it happens, in full, so nothing runs invisibly.
Fifteen seconds, from a meeting
Pull out your phone, read where the run got to, send one correction, put it away. The run never stopped and you never opened a laptop.
- Steer, stop or queue from the phone
- Live output on the phone as it runs
- Handoff back to the desk mid-run
Link a device once
Pair your phone to the Mac from settings and it stays linked. Connected devices are listed, and you can unlink any of them.
- Link the Luminair app on your phone
- Connected devices listed in settings
- Unlink from either end
13Memory & reach
Remote
Your phone is the remote for your Mac.
Turn your phone sideways and it becomes a remote for the sessions open on your Mac. Tap a session to aim at it, hold the mic, speak, and let go. Your words go straight into that session.
Point, speak, it runs.
Aim at any session
Layout draws your Mac window as it is, split panes and all, with each session by name. Tap one and your Mac brings it to the front.
Voice only, by design
Hold the mic, say what you want, release to send. An idle session runs it now. A busy one puts it in its queue.
Wi-Fi first
On the same Wi-Fi the phone talks to your Mac directly, which is faster. Anywhere else it goes through the cloud and does the same thing.
Two views of the same Mac
List shows every open session with its status, folder and queue. Layout mirrors the split panes of the window and shows names only, never the conversation.
- List: status, folder and queue per session
- Layout: the split panes of your Mac window
- Hold a pane to speak into it directly
Open, reorder and close from the couch
Search finds any session on your Mac and opens it as a new pane. Hold a session name to drag it into a new order, or to close its tab.
- Search every session on the Mac
- Drag to reorder within a pane
- Closing a tab does not stop its run
The intelligence behind it all
Luminair
It watches every request and every run, so nothing slips.
Turn it on with one switch and it tracks each request to done, checks the finished work, picks the model for every message and watches what you spend.
Tracks every request to done
Type //track and the request becomes a row in its ledger. Each row moves through open, working, done, tested and confirmed. A summary that claims success does not clear the row, so a run that stops halfway shows as half done.
Checks finished work with fresh eyes
When a tracked item is done, a fresh AI with no memory of the session traces it through the code and passes or fails it. Tested needs evidence, and only you can mark an item confirmed.
Notices when something stalls
It runs a quiet check every fifteen minutes. An item with no progress for twenty minutes is flagged as possibly stalled. Background jobs join the ledger on their own.
Picks the model for each message
A built-in judge sizes every message as light, standard or heavy, then routing picks the model and how hard it thinks. Research and important code can go to specialist models you choose. Pin a model on a message and it leaves it alone.
Watches the spend
The Spend tab shows today's dollars, tokens and turns, the burn rate per hour and the last seven days, split by model, project and session. As your budget runs low, Smart conserve suggests a cheaper model for heavy tasks, or switches for you if you allow it. It never touches a running task.
Trims the waste
Lean replies cut filler and keep code, commands and numbers exact. Surgical edits change a file with small patches, not a full rewrite. Old turns leave a short summary behind.
Finds anything and answers questions
Press Cmd+K, or Ctrl+K on Windows, and type a word you remember. Matching sessions and Journal docs appear as you type. Ask how something in Luminair works and it answers from the app's own help, and says so when it does not know.
Does as much as you decide
One switch turns it on or off. Under it, each part has its own switch, so you can run the full ledger on a production project and almost nothing on a scratch session.
Try it on your own machine.
Mac, Windows and Linux, with a phone app that keeps every session in your pocket. Bring the AI accounts you already pay for.