Tasks and projects
Unlimited nested subtasks, timed reminders and repeat rules, plus a dedicated project management window. Delegate any task to an agent and progress writes itself back.
Local-first · macOS
It listens, it talks,and it does the work
Cortex Desktop isn't another chat box. Speak to it and it answers. Let it click and type its way through the apps on your Mac. When one worker isn't enough, it dispatches a squad of subagents in parallel. What you get back is a finished deck, report, or document — produced entirely on your own machine.
- Cortex is built in — works out of the box
- Or drive your local Claude Code and Codex
- 37 providers · 1000+ models
macOS · Apple Silicon and Intel · auto-updates
What it does
From one sentence to something you can actually send
These are real entry points in the app. Pick a scenario, add a line of your own, and it runs.
Everyday work
- Slide decks「Build a 10-slide launch keynote — strong opening, clear call to action at the end」
- Deep research「Map the 2026 AI coding assistant market: key players, positioning, business models」
- Contract review「Review this contract from the supplier's side — flag risk level and edits clause by clause」
- Incident notice「The platform was down from 23:05 to 01:10 last night — draft a customer-facing statement」
- Company knowledge「What's our travel reimbursement policy? Cite the source」
- Writing「Turn these meeting notes into a full project retrospective, keep the conclusions intact」
- Automation「Every weekday at 9am, summarize AI news and push it to my team chat」
Design & ideas
- Design mockups「Design a SaaS landing page — dark theme, sharp value proposition above the fold」
- Slide decks「A 6-slide product intro deck, pixel-game styling, bold color palette」
- Loop workflows「Fetch competitor updates → summarize → judge importance → notify me if it matters」
Examples taken straight from the scenario shortcuts inside the app.
Voice assistant
Skip the typing — just say it
One floating orb. It sends when you stop talking, reopens the mic when it finishes speaking, and keeps the conversation going.
Wake it with a word
No button, no shortcut to memorize. Wake-word detection runs entirely on device — ambient sound never leaves your Mac.
Local transcription, streaming as you speak
SenseVoiceSmall plus a streaming recognizer, covering Chinese, English, Japanese, Korean and Cantonese. Text appears while you talk; audio is never written to disk or uploaded.
It speaks as it writes — cut in anytime
No waiting for the full answer. Start talking mid-sentence and it stops to listen. Its own echo and your keyboard won't trigger a false interrupt.
What you say actually gets done
Voice runs through exactly the same session pipeline as typing. It can drive your computer, create scheduled jobs, check your calendar, and call MCP tools and skills — not just chat.
Session workspace
You can see every step it takes
Most AI apps give you a spinner. This one opens up the middle.
Process timeline
What it's thinking, which tool it called, which subagent it dispatched — laid out step by step, so you always know where it is.
Artifacts live in the chat
Generated pages, charts, SVGs and documents render as cards right there. Edits update in place and get versioned, all inside a sandbox, and they survive a restart.
Split the pane to compare
Drag a session to the top, bottom, left or right of any pane to split it — no limit on how many sit side by side. Compare answers without switching tabs.
Goal whiteboard
For long jobs, the goal and progress get written onto a whiteboard in the right rail, updated as you go, and saved with the session.
41s elapsed · 9 steps · 3 files read · 6 tool calls
- ThinkingLet me look at the project layout and its dependencies first
- readpackage.json
- bashrg --files -g '*.ts' | head -40
- Repo explorerMap the boundary between data layer and state
- writedocs/architecture.md
View the full run
Output
What it hands back is a finished file, not a wall of text
No copy-pasting into another app to format it yourself. Every one of these exports to something you can use directly.
One sentence to a deck you can present
Describe the topic and it drafts an outline first. Approve it and the deck is generated from one of 35+ built-in templates, editable page by page.
- Drag, snap-align, swap images and reorder layers directly on the canvas, with full undo
- Speaker notes written for you, dual-screen presenter mode included
- Export to editable PPTX, plus PDF, PNG and a web version
Subagents
Too much for one worker? Send a squad
The lead agent decides how to split the job and who to hand each piece to, then runs them in parallel.
45+ built-in roles
Repository explorer, code reviewer, security auditor, fact checker, UX researcher and more. Roles marked read-only can't touch your files.
Or define your own
Name it, give it an avatar, write its brief — and pin down which tools it may use, which model it runs on, and how hard it should think.
Let it pick for you
Turn on automatic roles and it chooses the right specialist for the task type, so you don't have to decide up front.
Delegated work stays visible
Each subtask shows as a live card in the conversation. Open one to read its full execution log.
You can also call people out by hand: @-mention several roles and they answer in the order you named them.
Hands-on
It doesn't just advise — it goes and does it
Plenty of things have no API, only a user interface. So let it click like a person would.
Every action on your computer needs your approval
Clicks, keystrokes, drags, shortcuts, clipboard writes — anything that actually changes your machine requires a per-action confirmation, even when the permission mode is set to automatic. That line does not move.
Computer use
It moves the mouse and types for you, driving any app on your Mac — including the old ones that never shipped an API.
Browser
An embedded browser in the right rail lets it navigate, click, fill forms, screenshot and read pages. The automation runtime ships inside the app — nothing to install first.
Terminal and files
Built-in terminal, file tree and Git status bar. Code opens in Monaco, spreadsheets are editable down to cells and formulas, and PDF, DOCX and PPTX all preview inline.
It can read anything you throw at it
23 document formats parsed, including legacy Office, RTF, EPUB and ODF. Headings, tables and lists keep their structure, so it doesn't misread tabular files.
Needs your approval
Let Cortex control Keynote?
- Action
- Click
- Target
- Menu item “Export as…”
- Position
- (1284, 96)
This will really click on your screen. Nothing happens until you approve.
DenyAllow
Unattended
Away from your desk? It keeps working
Hand off the repetitive parts and let the results come to you.
Say it once, the job is scheduled
"Summarize AI news for me every weekday at 9am" — it creates the job on the spot, tagged as AI-created in the list. Creating, editing, deleting or running it all need your approval first.
Loop visual workflows
26 node types to build a pipeline: schedule or webhook triggers, AI steps, conditional branches, HTTP requests, sub-workflows and loops — plus a human-approval node that waits for your sign-off.
Results find you
Push to WeCom, DingTalk, Lark or email the moment a run finishes, so you're not going back to check.
Nothing is lost if the window was closed
Run results, workflow announcements, task reminders and available updates all land in the notification center and stay for 60 days.
Catch it in passing
Don't let it interrupt what you're doing
All four of these live outside the main window — no need to switch back.
Quick replyDouble-tap ⌘
Inside any app — chat, email — double-tap ⌘ or hold Fn to speak, and a non-focus-stealing bubble offers draft replies. Hit enter and it types straight into the field. It learns your tone, and pulls open slots from your calendar when you're setting a time.
Flash notes⌘⇧N
Catch an idea the moment it lands. A separate overlay window — it never interrupts the current session or leaks into its context.
Spotlight barGlobal hotkey
One hotkey brings up an input box. Throw a thought at it without hunting for the window.
System-wide dictationGlobal hotkey
Dictate into any app, push-to-talk or toggle mode. Transcripts can be auto-corrected by AI for homophones and proper nouns.
Models
Bring your own models — keys never leave the machine
The built-in catalog covers 37 providers and over 1000 models. Five ways in; take whichever suits you.
37
providers
1000+
models
5
ways to connect
Already paying for Claude, ChatGPT or Copilot? Sign in once with OAuth and use it — no separate API credit required.
- Anthropic
- OpenAI Codex
- GitHub Copilot
- xAI
- OpenRouter
- Kimi
Once it's connected
Keys go into the system keychain
The config file holds only a reference name. Nothing is written in plaintext, and the real key is decrypted at the moment of the request.
Capabilities detected, not guessed
Image input, context length and reasoning tiers are resolved automatically. When they can't be determined, they're marked unknown rather than assumed.
It won't silently swap your model
If the provider you chose can't be resolved, it fails loudly and stops. The model shown is the model that runs.
More
Everything else is in the box too
Not add-ons. Installed with the app.
Memory that dreams
During idle time and overnight it replays recent conversations, distilling your profile, preferences and project context into every future chat. Memory is stored as three markdown files you can open and edit yourself.
Company knowledge base
Connect a remote knowledge base and answers come back with citation chips — click one to read the original in the right rail.
System calendar
Ask whether you're free Wednesday afternoon, or have it put the review on your calendar — it always confirms before writing.
Skills and extension market
The skill library is a two-pane file tree: templates and scripts preview and edit in place. Plugins and extensions install from the market and sync across machines.
MCP
Connect your own MCP servers over stdio or HTTP, and a dozen-plus built-in capability domains are handed to the agent out of the box.
Desktop pet
Over 4000 to choose from. It shifts posture with the agent's state — thinking, running a tool, waiting on your approval, or failed — so a glance tells you where things stand.
Branches, archive and search
Regenerating creates a branch instead of overwriting the original, so you can page between them. Sessions archive and full-text search.
Local-first, and we mean it literally
Your data sits on your disk
Chats, notes, tasks, automations and memory all live in ~/.cortex-desktop/ — open it in Finder, they're readable, editable files.
Credentials in the system keychain
Encrypted by macOS. The config file only ever holds a reference name; nothing lands in plaintext.
Voice stays on device
Wake-word detection and local speech recognition both run on your machine. Audio is never written to disk or uploaded.
You hold the permissions
Three execution modes: plan, ask, automatic. In plan mode it physically has no tool for writing files or running commands.
Get started
Install it and give it one job to do
Free to download. Connect one of your own model providers in settings and you're running.
5,058
Downloads
3,135
People using it
macOS · in-app auto-updates
Bring your own model provider: sign in with an existing subscription, or paste an API key.
Read the changelogReach the author
Something wrong? Come find me
A one-person product — no support ticket queue. If it feels off, misses a feature, or breaks, just say so.
