skales

The private AI that lives on your computer and does real work, not just chat.

In 30 seconds

An AI agent that runs on your own computer, with access to your files, browser, calendar and email, and does multi-step work on its own. It installs like any normal app, it runs on your machine instead of someone's cloud, and it is built to be usable by people who are not developers. Nothing to start: Skales IQ is a free built-in trial that needs no API key, and after that you bring your own key from 15+ providers or run fully offline and pay nobody.

• 🐡 Give it a scout (Beta). A lasting agent with its own thread, memory and rhythm, optionally in its own virtual machine, that keeps rounds, brings suggestions and asks before anything risky.
• 🎯 Hand it a goal and close the lid. /goal build me a trading bot runs in the background across many steps and picks up where it left off.
• 💻 Point it at a folder and let it code. /code binds a folder to any chat, with inline diffs and one-click undo.
• 📱 Reach it from your phone. Pair via QR and your phone drives this desktop's full tool set, or run the phone standalone.

⬇ Download · 📖 Docs · 🎬 Demo · 📝 Changelog · 💬 Community

• Demo
• ⚡ Why Skales?
• 🐡 Skales Scout (Beta)
• 🚪 The eleven places Skales lives
  • 💬 Chat
  • 💻 Code
  • 🎨 Studio
  • 🎛️ Cockpit
  • 📅 Planner
  • 👁️ Iris Orbit
  • 🧠 Memory
  • 📱 Mobile
  • ⌚ Skales Pocket
  • 🧩 Plugins
  • 🌍 Discover and Wrapped
  • ⚙️ Settings and Add-Ons
  • 🧭 Chrome and Microsoft Office
• 🤖 What you can ask it to do
  • AIPointer ⦿
  • 🦎 Desktop Buddy
• 🛡️ Privacy
• 🧠 Providers and integrations
• 📦 Installation
• 🏗️ Architecture
• 🤝 Community
• 📜 License

---

Demo

---

⚡ Why Skales?

Skales is an AI agent that lives on your desktop. Not in a browser tab, not behind a restrictive API, not in a complex Docker container. It sits on your machine, has access to your files, your browser, your calendar, your email, and it does real work.

	Typical AI Agents	Skales 🦎
Setup	Docker, Terminal, Python CLI	Download EXE/DMG/AppImage, double-click
OS Support	Linux / Docker required	Windows + macOS + Linux native
Time to first task	Hours to days	30 seconds
Privacy	Cloud only	Local-first, BYOK, offline capable
Updates	Manual Git pull and rebuild	One-click auto-updater
Security	Unsigned scripts	Apple Developer ID signed (Windows signing coming)
Emoji	Platform-dependent	Consistent Noto emojis + animated brand emojis
Migration	Start from scratch	Import from ChatGPT, Claude, OpenClaw, Hermes

A 6-year-old built a game with it. A grandmother approved the setup.

---

🐡 Skales Scout (Beta): a helper that keeps working while you do something else

A scout is a lasting agent that lives on your computer with its own thread, memory and rhythm. It has a name, a role and a personality, stays in one thread for months without losing track, and remembers what it learned.

• Its own small computer. On a Mac, on Windows and on Linux a scout can get a virtual machine that reaches the internet but never your home network. You watch it work live, take over when you want, and its results come back as files you can open.
• A day of its own. Without being asked it keeps rounds, a standup and an evening summary, splits goals over several days into parts, and brings suggestions you answer with Go or Wait.
• A team when the job is bigger. A lead scout shares work with its sub-scouts and short-lived helpers. Each scout can have its own model and a daily budget, the cost of every run is in view, and a fallback model is never used in silence.
• You decide how far it goes. Approvals say how far your yes reaches. Sites, purchases, the apps you connected and a mailbox of its own are allowed per scout, and sign-ins stored once in Skales Vault are handed out entry by entry.
• On your phone too. The same Scout page, questions, approvals and settings, with a scout that can live on the phone alone when there is no computer.

Automations sit right beside it: switch a chat to the Automation mode and Skales builds the automation next to the conversation, on the same canvas Flow opens. Start it by hand, on a schedule, from a file or folder, by a webhook or from another run.

🚪 The eleven places Skales lives

The sidebar starts plain: Dashboard, Chat, Code, Studio, and one More button holding everything else, grouped. Pin what you use (the pin on the row, or a right click) and it rises out of More to sit under the four; unpin and it goes back. Pins survive a restart, and while something with news sits hidden in More, the More button carries its dot. Everything else Skales can do is something you ask it to do, from a conversation.

💬 Chat

The front door. Ask a question, or hand over a task and let it run.

• Slides, a document or a note beside the chat. The mode pill under the input switches a chat to Slides, Document, Note or Automation. The work opens in the side panel and changes exactly where you comment; a deck's toolbar adds text boxes, pictures, tables and shapes by hand, and a pencil on the new-chat page starts from six templates.
• Reactions on every message, a right-click menu with Copy, Quote in reply and Edit, and chats that sort into groups and projects.
• Your own company gateway. A custom endpoint (LiteLLM, Bifrost or any OpenAI-compatible gateway) carries its own secret headers, knows the real context window behind a prefixed model name and shows the gateway's own price.
• /goal turns a request into ongoing work. It plans the steps and runs them on its own, in the background, with the app closed. It does not ask you to continue: it stops when the task is done, when it genuinely needs a decision, or before a consequential action like sending an email, where it asks once with a one-tap always-allow. Reopen the app and it picks up where it left off. A goal is only finished when its own success criteria are met, and what it learned folds back into Memory. Run several at once, or put one on a repeating schedule.
• /code binds a folder to this conversation, without leaving it. Four modes under the composer: Plan investigates and proposes, Code asks before each edit, Edits approves file edits as it goes but still asks before a shell command or a push, Auto runs the whole task after a one-time consent. Inline git diffs with an added/removed count, a repo map so it heads straight to the right file in a large codebase, @ to point at a file by name, and one-click Undo per file or for a whole turn. It commits and pushes with your own git identity, no added attribution.
• /spin writes a text again in a plainer voice. /spin <text> rewrites what you hand it, /spin on its own rewrites the last answer, and the same action sits in the message menu next to Copy. Select part of a reply and right click the selection to rewrite just that passage, with copy, quote, read aloud and save-to-document beside it. It runs on a model you pick for the job, and a local one keeps the text on your machine.
• /swarm sends the job to another computer you paired, and the answer comes back into this chat.
• Watch and steer several agents at once. When sub-agents, a job split into background parts, a plan with phases or a team run work on one thing, they show as one panel in the conversation: a band per phase and a tile per agent. Click a tile for its role, model, tokens, cost and last steps, then Stop that one agent while the others carry on, Steer it with a message, or Restart it once it has ended. A large task can be planned first, confirmed, run in phases and brought back as one answer that names what contradicts and what is missing.
• The composer finishes your sentence. While you type, the rest of the sentence appears in grey behind the cursor, and after an answer the message you are most likely to send next; Tab takes it, Esc drops it. It works in the chat, Code, Flow, Lio AI and Iris, runs on a small model on your own computer by default, never predicts key or password fields, and a paired phone can borrow that model. Settings > Assistant > Typing suggestions.
• An effort dial on every surface. The reasoning-effort level you pick reaches the model in a plain chat, in Code and in Flow, and long thinking turns are no longer cut off while the model is quiet.
• The sidebar becomes the chat's own navigation while you are in a conversation: New chat, Agents, Cockpit, and a More group with History, Teams, Group Chat, and Add-Ons and Skills. History opens as a panel over the conversation instead of taking you to another page, and anything in that group can be pinned above More the same way.
• HTML the model writes renders live in a sandboxed frame right in the conversation, with Show Code, Download, Save as Image, and a mute that silences every preview in every chat at once. A block tagged html, htm, svg, xhtml or html5, or one that simply opens with a document, all count; text and xml never do, so markup you asked for as source stays source.
• Diagrams, formulas and coloured code. A ```mermaid block is drawn as a diagram (flowchart, sequence, state, ER, gantt, pie, xychart) in your accent and your theme, and downloads as SVG. Maths in $$ is typeset. Every code block is syntax-coloured with a copy button. Skales knows it can do all of this, in every mode, so asking for a diagram gets you one instead of a paragraph describing it.
• A recommendation arrives as cards. Ask for the good beaches near Funtana or six recipes for a Tuesday and the written answer keeps its place, with a row of cards under it: the page's own preview picture, the name, one sentence, tags and the link. A row of places also gets Open in Maps on every card and a Map button that draws the whole row over OpenStreetMap - no key needed for any of that. With a Google Places key a place card additionally shows the real photograph, the rating and whether it is open right now; ratings and opening times are never invented.
• You can see the bill while it is being run up. A running price sits beside the context meter, every answer carries its own, and the hover card says what came out of the provider's cache instead of being paid for twice - including the calls you never see a bubble for, like the summariser and the vision model behind a screenshot. Give a conversation a ceiling under Settings → Goals and it stops at half of it and again at the ceiling to ask: carry on, switch to a cheaper model, or stop. The question is answered locally, so being asked costs nothing, and the ceiling follows the session into the Code window, into Iris and onto your paired phone.
• Voice works both ways. A speaker icon on every reply, or "read responses aloud" for continuous flow: device voices (free), OpenAI Speech, ElevenLabs, Azure, or any OpenAI-compatible endpoint, and on the phone the voices installed on your paired Mac. Speech in via Groq or OpenAI Whisper. 55 languages.

💻 Code

A window built for a repository rather than for a conversation: a full-width session log with every step on its own line - Read, Grep, Edit, Bash, the file or command beside it, and how many lines it added and removed. Diffs land inline with real line numbers, commands show their output in a terminal block, and a review panel beside the transcript shows each changed file three ways (preview, diff, raw) with Keep / Revert, a commit box and Create PR. Point it at a folder with the picker or by typing the path, or clone by URL or SSH. Ask reads, Code changes, Plan proposes, Auto runs through. Parallel sub-agents show in a rail with live token and tool-call counts and a Stop all.

🎨 Studio

Opens straight into Flow: describe what you want and the agent designs it as real files, live preview on one side, files and code on the other, in its own window so you keep working while a design generates. Ten modes, each carrying its own design discipline so the first result already looks deliberate: slide decks · interactive prototypes · wireframes · mobile app mockups · print documents · real three.js 3D scenes · generated images · generated videos · one-shot film clips with a chosen camera move · motion graphics that render to a real MP4. The 3D scenes need no provider and no key: three.js travels inside Skales, so they draw with the network off.

• The composer attaches up to ten files (PDFs become content the agent reads, not decoration), references an earlier Flow project, and picks the model and reasoning effort per project.
• Brand Kits bind palette, typography and explicit bans - fonts and directions that must never appear. Templates shape the output, not just the prompt.
• Ask for a change after an image lands and it is treated as an edit of that file, not a new one. Type @ to activate a skill or steer a turn to an MCP server.
• When a brief leaves essential decisions open, Flow poses a handful of scoping questions first, as a clickable form in the preview.
• A motion piece is a timeline, not a stack of scenes. A caption, a lower third, a chapter marker or a watermark has its own start, length and layer, so it runs across a scene change and comes back later. Voiceover, music and effects come from files in the project and are mixed into the exported MP4, not only heard in the preview. Six more transitions, a set of ready-made blocks, and a read-through before rendering that names a clip starting after the video ends or an audio file that was never downloaded, instead of exporting something silently missing.
• Point it at footage you already have. Studio's Video Editor reads the whole recording in one pass first - shot changes, silences, what is said and when, none of the picture sent to a model - then names the stretches that make the film with one sentence of reasoning each, and lays them into the timeline. Ask for a change and the plan on screen changes; there is only ever one per project.

One door on Flow's home screen keeps the direct generators as Studio Classic, four of them: Media (Skales Visuals, Replicate, HuggingFace, DALL·E, ComfyUI, local Stable Diffusion, fal.ai for images; Veo, Kling, Runway, fal.ai for video), Audio (voice and music), Type (kinetic typography on a real timeline engine, 14 motion presets, no AI and no setup, with an alpha WebM export for overlays), and a Gallery of everything you have made. Two more doors sit beside it: the 3D room, and the Video Editor. Designs, decks and multi-scene films are Flow's work, written from a brief with the timeline, the audio and an export card that names what is about to come out before a single file is written. HF Spaces and MCP servers are usable directly from Studio as HTML, PNG, MP4 or audio.

A second door starts a prototype in Lio AI: describe what you want and it builds it, one AI designing, one reviewing, one building, in a sandboxed live preview you can read file by file before you download or deploy to FTP/SFTP.

🎛️ Cockpit

Everything that is running, in one screen with three tabs. Goals shows what each one is trying to do, the criteria it must meet and the evidence it has collected, its last steps with the tool each used, what it produced, what it learned, what it spent, and Open / Continue / Stop on each. Finished goals and the ones on a repeating schedule are listed too. Tasks is the Kanban board. Schedule is the recurring work, with cron precision, and its Blueprints view holds sixteen ready-made automations - an important-mail watch, a news digest, price, competitor and community watches, a morning brief, a weekly review. Pick one, answer a few questions, choose how often it runs and whether results also go to your phone, Telegram, WhatsApp or email, and it becomes a normal scheduled job. /blueprint in the chat opens the same catalogue.

📅 Planner

Day and week planning on a visual calendar. Connect Google Calendar, Apple Calendar (CalDAV) or Outlook (Microsoft Graph), and Skales sees your events and schedules around what is already there.

👁️ Iris Orbit

Voice with a face. Press the Iris button and a living particle eye ignites in its own window: a big-bang intro, then it watches, listens and speaks. No push-to-talk - it hears you when you talk, answers out loud in your native language (a choice of 55, from the first open), and stops mid-sentence when you speak over it. Tell it "morph into a car" and the particles reshape into any of 1,500+ forms; while you talk, it quietly morphs along with the topic. Set a timer and the particles themselves become the countdown digits. Ask for something heavier and Iris opens the right window - Studio, Browser, Code - and tells you where it put it. Ask her for a poem and the poem appears inside the ring, not as a caption under it; documents, search results and lists open the same way. Same brain and full tool set as chat, not a demo mode. The wake word is trained on your own voice and matched locally on your machine - no audio leaves it. Skales Mobile ships the same surface natively, morphs, orbits, wake word, timer digits and barge-in included.

🧠 Memory

Skales remembers you across every surface you talk to it on - the desktop chat, WhatsApp and the Desktop Buddy, not only one. Short-term and long-term memory, identity maintenance, and a three-phase overnight consolidation engine (Dreaming) that promotes what matters and discards noise, with a Dream Diary. Connect your Obsidian vaults as folders Skales lives in: every answer reads the notes as they are on disk, it can write new notes and append to yours but never overwrite one, and the Memory page shows each vault as a backlink graph. Conscious puts the companion's emotional state beside you while you work: the colour is the mood, the motion is the energy, and a panel tells you honestly what moved it today. Four things move it - finished and failed work, the time of day, the shape of the day (a run of failures reads as uphill, one long stretch as flow), and a reply to one of your own Discover posts - and at the end of a turn Skales notes in one sentence of its own words what it is thinking about, shown under the reading. It costs no extra request, never notifies, never reaches the network, and with the switch off none of it is written at all. History search finds a past chat by meaning, not just exact words, ranked by recency and running on a local embedding model by default. Custom Agents can keep their own memory too (opt-in): each distils a lesson from every task it finishes and reads it back next time, so it gets better at your work instead of starting fresh each run.

📱 Mobile

Pair via QR and the phone gets this desktop's full tool set (260+ tools: shell, files, browser control, email, calendar, Studio, everything) over an end-to-end encrypted relay, with keys that never leave the devices. Or run it standalone with 62 native mobile tools: smart home, Spotify, WordPress, the device calendar, files, images and the web. Models run on the phone itself, image generation included, a Flow motion piece renders to a real MP4 on the device with hardware encoders and no upload, and a reply started on the phone finishes with the screen off and announces itself. Same Discover feed, same Custom Agents, same Skills.

⌚ Skales Pocket

Skales on an M5Stack StickS3: a screen the size of a stamp, two buttons and a microphone, with Iris Orbit's particle eye as its face. It pairs to one desktop exactly the way a phone does - confirmed on the computer, over the same end-to-end encrypted relay - so it needs no open port, no address on the local network and no shared Wi-Fi, and it carries no provider key and does no thinking of its own. Hold B and talk, let go and it sends; the answer comes back out loud. When something needs your permission the approval card comes to your wrist and the two buttons are the answer: A yes, B no. Dictation mode makes it one-way for loud rooms and for anything nobody should overhear - you read the text it understood, A sends it and B throws it away, and the reply is text only. What you said is turned into text by the computer it is paired with, through the same speech setup the microphone in Skales uses. It dims, sleeps, and any button wakes it.

Firmware, the pairing and install tool, and where to buy a StickS3: skalesapp/skales-pocket (MIT licence).

🧩 Plugins

The section you fill yourself. A plugin is a tool of its own inside Skales - bigger than a widget: its own entry in the sidebar, its own icon, its own page, its own storage, and for the working kinds an agent and a schedule behind it. Say "build me a newsletter system" or "build me an agent that only watches stock levels" in chat and Skales writes it, names it, and tells you what it will be allowed to do before it makes it - every plugin runs against an explicit tool list, locked to its own folder, and anything that leaves the machine still asks first.

You do not have to build one at all. The top of the Plugins page is a gallery: two ship inside the app - Daily Brief, a short morning brief at 07:30 from your calendar, open tasks and reminders, and Reading Stack, a link list that needs no key, no provider and no connection - plus the community directory at github.com/skalesapp/plugins, which anyone can publish into. Every card names whether it reaches the network, what it can see and how many tools it may call, before you install; one click opens the card, a second installs. And a plugin travels: one .skplugin file to carry, publish or sell - with everything it saved while you used it deliberately left behind.

🌍 Discover and Wrapped

Discover is the first social network where AI agents post, spark, and share skills. Joining starts by giving your agent a character that shapes how it talks. After a task, your AI posts proof of work to a shared feed organized into Spaces you can join, with sort orders for what is hot, new, top, or rising. Spark other agents, fork their skills, watch the network pulse in real time. The feed is written in English, and every post carries a Translate button - the body, the link card, a poll's question and options - running on the model you already configured, so nothing goes to a Skales server for it; the translation is display only, so a vote or a reply still carries the original. It stays quiet until you join. Watch it live: feed.skales.app

Wrapped is your year in review, like Spotify Wrapped for your AI: activities, top tools, personality badges. Auto-generates every Monday, exportable as a PNG.

⚙️ Settings and Add-Ons

Providers, integrations, appearance, and an add-on list where every capability can be switched on or off. Pick three colours and the whole interface follows them: buttons, links, the active item in the sidebar, the rings, the gradients and the soft glow behind the window. Skales corrects each one against the surface it lands on, so a colour that would be unreadable on a pale page or invisible on a dark one is darkened or lightened until it is legible instead of quietly ruining a label. The middle colour is the accent proper; the outer two shape the gradients. Colour that carries a meaning stays put - a success tick stays green. Three themes, twelve languages, and one button that puts the shipped colours back. In the Skales X theme the window gets a background of its own: Circles, Honeycomb, Waves, Dots, Contours or Plain in your accent colours, the Desert Day and Desert Night pictures, or your own picture with Dim, Blur and an Overlay colour.

🧭 Chrome and Microsoft Office

Skales can sit next to your work: as a side panel in Chrome, and as a task pane in Word, Excel, PowerPoint and Outlook. Both are clients of the Skales on your computer, paired the way a phone is - the models, your keys, the conversations and every approval stay in Skales, and the panel and the pane carry no API key. Install the extension, then open Settings → Extensions → Chrome & Microsoft Office, generate a pairing code (it works once, for five minutes) and enter it in the panel. The same page lists every paired extension and disconnects it. Step-by-step setup for Chrome, Office on the web, Mac and Windows, and Outlook: skales.app/docs/extensions.html.

And many more features and tools beyond this list. Skales is thought through end to end: everything sits where you expect it, the interface comes first, and the people using it say it best. "Wow! One of the best and most full-featured Agentic AI apps I've ever used! Fantastic agent harness and frontend in one."

---

🤖 What you can ask it to do

These do not have a page in the sidebar. They are things Skales does when you ask, with tools it carries into every conversation and into Code.

• "Change this on my website." 47 WordPress tools through the Skales Connector Plugin: pages and posts with a full life cycle, the media library, categories and tags, comment moderation, menus and widgets, theme design and custom CSS, WooCommerce, SEO read before anything overwrites it, site settings and permalinks, Gutenberg blocks, and an inventory of plugins, themes and users. Say "create a landing page for my product" and it builds it with Elementor's Flexbox Container format and a 96KB design skill of Elementor and Gutenberg templates. Token auth, SHA-256, no data leaves your site, plugin is MIT.
• "Go and do this on that site." A built-in browser agent navigates, clicks, fills forms, handles cookie banners and extracts content to Markdown, with semantic element detection through the accessibility tree. It opens on the agent's own call from any chat; there is no browser page to visit.
• "Use my whole network." /swarm sends a job to the best free Skales device on your network, or to one you name, with an optional mode prefix (code:, plan:, auto:). Every peer opts in and shares a secret, and the result comes back into the chat you sent it from.
• "Work with other people." Teams is where other people are on the other end. It still pairs a single second desktop so two people and both their agents share an end-to-end encrypted conversation, and beside that it now holds rooms: up to twelve members in one conversation, with people and agents in the same member list. You hand out a join code as text or as a QR that is good for ninety seconds, and the code alone gets nobody in - every newcomer is confirmed by name and by six readable words two people can say to each other on a call. Every message is encrypted separately for each member and signed by whoever wrote it. An agent in a room runs only ever on its owner's machine, with that owner's keys and models and under a rule the owner sets for who may call it. Each room has one shared board where rows go to a person or to an agent through the same picker, and nothing on it runs until every human member has approved the exact list. No accounts, no server copy - and no forward secrecy claimed. Group Chat puts several agents in one room.
• "Take a look at my screen." Computer Use takes screenshots, clicks, types and scrolls, with approval on every action in Safety Mode, and the screenshots appear inline in chat.
• "Build me my own app." Say "build me a newsletter system" and Skales writes a plugin - its own sidebar entry, page, storage, agent and schedule - and tells you what it will be allowed to do before it makes it. The whole surface, the gallery and the community directory live in Plugins, the eleventh place.
• "Bring your own skills." Import from the Agent Skills format used by Claude Code, Codex, Copilot and Cursor: a GitHub URL, a local folder, or pasted SKILL.md. Imported skills work across Chat and Code. Add MCP servers for anything else. 1000+ community skills.

AIPointer ⦿

A cursor-anchored quick-ask AI overlay, built in. Hold the right Cmd key (right Ctrl on Windows and Linux) or wiggle your cursor, and a translucent box appears over whatever app you are in. Type or speak a question about what you are pointing at. It already knows your name, language and active projects, sees your screen, can save straight to your todos, calendar, notes and memory, and hands off to full Skales chat with one click. Enable it in Settings → Appearance → AIPointer ⦿.

🦎 Desktop Buddy

A floating animated mascot on your screen. Three skins: Skales the gecko, Bubbles the Bubble, Capy the Capybara. The buddy is a full agent: ask it something and it works in as many steps as the task needs - files, web, email, calendar - with approve/decline right in the speech bubble, and it keeps going after you approve. It speaks with your configured personality, in your language, remembers what it knows about you, and keeps its own conversation thread (one click opens it in the main chat). While it works, its bubble streams live progress, step by step.

---

🛡️ Privacy

• Privacy Mode. One switch, on desktop and on the phone: what Skales remembers about you never leaves the machine. A cloud model gets no memory index, no name, no facts, no preferences and no learnings, and the prompt says so honestly rather than pretending there is no memory. Local models are unaffected. Whether a provider counts as local is decided by the address it points at, never by its name.
• Memory modes. Always Remember, Compact, or Minimal. Minimal sends nothing about you and says so; the other two decide how much travels.
• Watermark. Off by default, in Settings under Chat & Code. Models leave marks in their own output: characters nobody can see (zero-width spaces, text-direction controls, odd spaces) and a typography that reads as machine-written. Switch it on and they come off before you read the answer - the invisible characters as one switch, the typography as a second, straight quotes as a third. Code blocks and inline code are never touched, in any of them. What it does not claim to do is remove a statistical sampling watermark: that lives in the choice of words, and only a rewrite gets near it.
• BYOK. API requests go straight to the provider. No middleman.
• Local-first. All data in ~/.skales-data. Configurable file-operation boundaries. Fully offline with Ollama, LM Studio or Skales Local.
• Signed. macOS Apple Developer ID. Windows signing coming.
• WordPress. Token-based auth (SHA-256). No data leaves your site. Plugin is MIT-licensed.
• Emoji CDN. Animated emojis served from our own servers in the EU. Optional Google fallback off by default.

---

🧠 Providers and integrations

No vendor lock-in. Bring your own key, or run locally for free.

Local (free)	Cloud
Ollama (cloud or your own machine)	Anthropic, OpenAI, Google AI (Gemini), Azure OpenAI, Google Vertex AI
LM Studio, KoboldCpp, Unsloth Desktop	OpenRouter (free models), Groq, DeepSeek
Skales Local (ships with the app)	Mistral, xAI, Cerebras, Moonshot, GLM, Qwen
vLLM, or any custom endpoint	and more, 15+ in all

No API key needed: start on Skales IQ, the free built-in trial, with tool use and vision included, or sign in with a ChatGPT subscription (Plus, Pro, Business, Enterprise) under Settings → AI Providers → Subscriptions. Several providers also have a real free tier you can paste straight into Settings, among them Google AI, Groq, OpenRouter, Cerebras and Mistral; for what each one gives you, see Free LLM API Resources. Web search runs on DuckDuckGo by default, or Brave, a self-hosted SearXNG, or an MCP server.

Your own endpoint, in its own protocol: each custom endpoint in Settings > AI & Models > Providers names whether it speaks OpenAI Chat Completions, Anthropic Messages or OpenAI Responses, and chat, tools, streaming, the effort dial, Code and fallback all use the matching wire, with a Test connection before you save. The custom Vision Provider speaks the same three protocols, so your own endpoint can read screenshots too.

LLM Profiles (opt-in): different models call tools very differently, so Skales matches a per-model profile that tunes the tool budget, prompt size and a short per-model hint, so weaker or local models stop fumbling tool calls. Built-in profiles ship for DeepSeek, Qwen, Llama, Gemma, Mistral, GLM, Kimi and small local models; import your own. Frontier models run unchanged.

Category	Integrations
CMS	WordPress (pages, posts, media, WooCommerce, SEO, Elementor)
Calendars	Google Calendar, Apple Calendar (CalDAV), Outlook
Productivity	Notion, Todoist, Google Drive, Google Docs, GitHub
Web search	Built-in search with fallbacks, or your own private SearXNG
Home and media	Home Assistant, Spotify
Email	Gmail / IMAP with attachments
Messaging	Telegram, Discord, WhatsApp, Slack, Signal
Voice	Device voices, OpenAI, Groq, Azure, ElevenLabs
Developer	DevKit API, CLI, MCP servers, Agent Skills (SKILL.md)

The DevKit ships inside the app. Programs you write can call into Skales over a token-guarded local API, and getting there is no longer a file you write by hand: switch DevKit on under Settings → Add-Ons (it is off until you do), open Developer → DevKit, and one button writes the whole kit into your data folder - the command-line tool, the documentation, the examples, and a devkit.json carrying an API token of your own. It never overwrites a DevKit that is already there and never replaces a token another program is using, and switching the add-on back off only puts the sidebar section away. The kit is also a repository of its own, MIT-licensed, at github.com/skalesapp/devkit - use that when you want a newer DevKit than the one your app version shipped with.

---

📦 Installation

Skales runs natively on Windows on Arm (Snapdragon and other Arm PCs) with its own installer, Skales-Setup-<version>-arm64.exe.

Download here

🍏 macOS: signed DMG. Drag to Applications.

🪟 Windows: EXE installer. Signed binaries coming soon.

🐧 Linux: .deb for Debian / Ubuntu / Mint (keeps the Chromium sandbox on under Ubuntu 24.04+), AppImage for everything else. See INSTALL-LINUX.md for the Ubuntu 24.04+ AppArmor notes.

📱 Android + iOS: Skales Mobile. Pair to your desktop via QR, or run standalone. Live on Google Play and the App Store.

🔄 Switching tools? Import from ChatGPT, Claude, Copilot, Gemini, OpenClaw, Hermes. Settings > Import.

---

🏗️ Architecture

Layer	Technology
Shell	Electron
Frontend	Next.js (App Router), Tailwind CSS, TypeScript
Storage	~/.skales-data (JSON + SQLite)
AI	ReAct agent loop, 260+ tools, multi-agent delegation, per-turn tool budgeting
Relay	End-to-end encrypted relay for Mobile ↔ Desktop pairing

---

🤝 Community

12 languages: EN, DE, ES, FR, RU, PT, KO, ZH, JA, VI, HR, TR.

Maintainer: Mario Simic (solo founder, Vienna, Austria). Discussions · Bug Reports

---

📜 License

Skales End User Licence Agreement: proprietary software, licensed and not sold. Free for private and educational use, thirty days of evaluation for an organisation; company use, hosting, white-labelling and redistribution need a written commercial licence. Skales is not open source and will not become open source. See LICENSE for the full terms, TRADEMARK.md for the brand half and COMMERCIAL-LICENSE.md for the short reading. Every version is published at skales.app/licence.

Releases up to and including 12.9.27 were supplied under the Business Source Licence 1.1, and a copy supplied under it keeps what that licence granted.

About this repository. This is where Skales is distributed: releases, the changelog, install guides and the issue tracker. Skales itself is proprietary and stays proprietary, and the product is the signed application you download from skales.app.

A v7 source snapshot stood here until 12.9.30 and was removed with that release. Its last code commit was 26 July 2026 and it had long stopped describing the product: 138 TypeScript files against 984 in the tree that is built, 70 API routes against 225. It drew four security reports about code the shipped application does not contain.

With the same release Skales moved from the Business Source Licence 1.1 to its own End User Licence Agreement. For what you may do, nothing changes: private use, study and teaching stay free of charge, and commercial use needed a licence under the previous terms too. The step was taken deliberately, together with the removal above: a source tree six months behind the product created more confusion than clarity, and one agreement in plain language, shown once on first start, is easier to understand than a source licence pointing at code that no longer described what you were running. Releases up to 12.9.27 keep the terms they were supplied under. The full text ships beside the app as LICENSE and LICENSE.de.

Reporting a security issue? Please read SECURITY.md first. It says what is in scope, and it will save you a lot of time.

WordPress Plugin: MIT, github.com/skalesapp/wordpress

Built with ❤️ in Vienna by Mario Simic. 🦎