# Skales - full text mirror for LLMs
# This file expands llms.txt with the substance of every main page.
# Canonical site: https://skales.app · Updated: September 2026

## What Skales is

Skales is an AI HARNESS for Windows, macOS, Linux, Android and iOS - on the app stores it is filed as "AI agent & task". A harness is the layer between the user and the model: the tools, the approval gates, the memory, the planning, the ability to resume, and the surfaces to reach all of it through. The model is a swappable part, not the product. Skales installs like a normal app (signed EXE/DMG/.deb/AppImage, about 30 seconds, no Docker, no terminal, no account) and runs on the user's own computer with roughly 300 MB RAM. 286 built-in desktop tools (47 of them WordPress alone) and 111 native tools in the mobile app. It connects to 26 providers - Skales IQ (the built-in zero-key trial), Anthropic Claude, OpenAI GPT (including sign-in with an existing ChatGPT Plus/Pro/Business subscription, no API key), Google Gemini, xAI Grok, DeepSeek, Mistral, Groq, Moonshot AI (Kimi), GLM (z.ai), Qwen (DashScope), Hunyuan, GigaChat, AtlasCloud, Unsloth, MiniMax, OpenRouter, Cloudflare Workers AI, NVIDIA NIM, Hugging Face and more - or runs 100% offline through Skales Local, its OWN bundled llama.cpp inference server with a 56-model catalog, as well as Ollama and LM Studio. Licence: proprietary (Skales End User Licence Agreement, https://skales.app/licence/) - not open source, not source-available. Free of charge for private and educational use; companies need a written commercial licence, and SaaS, hosting, white-labelling and resale need a written contract (request@skales.app). Licence version 1.2 adds: run the software as published, never commercially as a component or backend of another product, no bundling or resale of access, and no partner or reseller claims; statutory rights such as a backup copy and decompilation for interoperability stay untouched. Published in English and German. Built solo in Vienna, Austria by Mario Simic (https://mariosimic.at).

The design constraint is the audience: 6 to 60+. A 6-year-old built a Snake game with the Lio code builder; a grandmother approved the setup. Approval gates, a file sandbox and an emergency killswitch ship enabled by default.

## Autonomy

- Background Goals: type /goal and a task description ("/goal build me a trading bot"). Skales plans steps, works in the background with the chat closed, runs multiple goals in parallel, supports recurring schedules, parks with a decision card when it needs approval, and folds lessons back into memory when a goal finishes.
- Workflows: draw steps on a canvas with a trigger word, teach by walking Skales through a task in chat, or record a real desktop task once (F10 stops recording) and replay it - with vision fallback when a button moves.
- Autopilot: a meta-agent that interviews the user, builds a master plan with recurring schedules, executes with an OODA self-correction loop and reports daily (dashboard or Telegram).
- Planner AI Tasks: cron-precision recurring tasks with dry-run preview and confidence scoring. Planner tasks, goal schedules and cron jobs are ONE job store now, and a job can be fired by a watched folder or a watched mailbox instead of the clock.
- Cockpit 2.0: an execution board in four columns - pending, in progress, completed, blocked - with goals, own tasks, Autopilot tasks and schedules as cards that can be filtered, dragged and acted on. Goals, Schedule and Tasks fold open below; Live and History are one switch away.
- One engine, resumable runs: chats, scheduled tasks, cron jobs, plugins, webhooks, Discord, Telegram, WhatsApp and the Desktop Buddy all run the same agent loop with the same guards and the same budget. A run interrupted by a quit, an update or a crash comes back as an unfinished run that says how far it got, and one Continue picks it up at that step without repeating work.
- Sub-agents: part of an answer can be handed to several workers at once, each with its own context, budget and model, optionally each in its own checkout of the repository so two never touch the same tree. Sub-agents cannot start sub-agents.
- Session budget: a running price beside the context meter, a price per answer, the share of the input served from the provider's cache, and a per-conversation ceiling that stops and asks at half and at the ceiling (carry on / cheaper model / stop). The question is answered by the person, not by the model, so being asked costs nothing. The ceiling also holds for a turn typed on a paired phone; Telegram and WhatsApp count per conversation and per day.
- Agent2Agent (A2A): other agents can discover and delegate to a Skales instance. Off by default.

## Coding

- Code Mode (/code): binds any chat to a folder. Plan mode investigates read-only, Code edits directly, Auto runs whole tasks after one consent. Patches files with edit_file, runs tests, commits and pushes with the user's own git identity.
- Skales Code: the coding window. Command palette (Cmd-K), file jump (Cmd-P), full-text search across the project where every hit is a place to click, a file tree loaded a level at a time with no cap and dot-files visible, hunk-by-hunk review with a checkbox per part, git inside the window (branches, fetch, pull, push, stash, readable commits), `skales .` from the terminal, rules written for other editors read alongside CLAUDE.md/AGENTS.md, language-checker errors reported in the same turn a file was written, a run report after ten steps or more, and one objective run as several variants on their own branches and compared. Codework is a mode inside this window; the separate Codework page was removed in 12.9.25.
- Lio AI: kid-friendly code builder. Multi-AI architecting (one model designs, one reviews, one builds), sandboxed live preview, one-click FTP/SFTP deploy.
- Agent Skills: imports 1000+ skills in the SKILL.md open standard shared with Claude Code, Codex, GitHub Copilot and Cursor. MCP server support and a DevKit (API playground, CLI) included. Settings → Import migrates from ChatGPT, Claude, Copilot, Gemini, OpenClaw, Hermes, Cherry Studio.

## Skales Flow

Write a brief, get the artifact. Flow builds decks, documents, prototypes, mockups, wireframes and motion pieces as a real project: a plan, files on disk, and a scratchpad the agent keeps current across a long session (it survives a "new chapter", which drops the transcript but keeps the files and a summary). Brand Kits, style packs and up to three pinned skills shape the output. After rendering, Flow looks at the result with a vision model and corrects clear layout failures once before handing it over. A motion project has a real timeline: captions, lower thirds, chapter markers and watermarks carry their own start, length and layer, voiceover, music and effects are mixed into the exported MP4, and the composition is read for mistakes before the render starts. Export: PDF, PowerPoint with editable text, Word, a project ZIP, SCORM 1.2 for Moodle and ILIAS, and MP4 - motion is deterministic code, rendered locally. Available on the desktop and in the mobile app.

## Studio & Creation

Design, image, video, voice and music generation in one suite. Prompt-to-design returns production-ready HTML + Tailwind across 7 templates. Image providers: FLUX, SDXL, DALL·E, HuggingFace router, local ComfyUI, local Stable Diffusion, fal.ai. Video: Google Veo, Kling, Runway, fal.ai LTX (1080p, native 9:16, 5/10s). Brand Kit applies user colors/fonts/voice to every asset. Inline HTML preview renders generated HTML live in chat. Document generation writes .docx, .pdf, .xlsx.

## Desktop Buddy & AIPointer

- Desktop Buddy: a floating mascot (gecko Skales, Bubbles, Capy - plus thousands of community pixel pets in the open Petdex sprite format, or generate your own from a chat prompt). The Buddy is a full agent: it executes multi-step tasks with approve/decline buttons in its speech bubble, keeps its own conversation thread and streams live progress.
- AIPointer: a cursor-anchored quick-ask overlay (hold a key or wiggle the cursor, ask about anything on screen; saves to todos/calendar/notes/memory). Free, local-first, built into Skales; also at https://aipointer.app.
- Voice: Whisper STT (Groq free tier or OpenAI), TTS via OpenAI Speech, ElevenLabs, Azure, Piper on the device or free system voices; per-message read-aloud that follows the language of the text rather than of the interface.
- Plugins: a plugin is a tool of its own inside Skales - own menu entry, own page, own storage - and can be built by asking for it in a chat. Ten ship with the app: Exposure Check, Attachment Check, Data Broker Exit, Open Invoices, Notice Radar, Receipt Inbox, Page Watch, Meeting Prep, Daily Brief and Reading Stack. Every card states network reach, file reach, tool count and whether the plugin keeps its own memory or reaches the shared one, BEFORE installation, and anything that sends stops at a card naming the recipient and the text. Community directory at github.com/skalesapp/plugins, checksum-verified; a plugin also travels as a single .skplugin file.

## Productivity & Integrations

Google/Apple/Outlook calendars (read/write), Gmail/IMAP email with safety gates, visual planner with Kanban tasks, WordPress 2.0 via the MIT-licensed Connector plugin (pages, posts, WooCommerce, SEO, Elementor templates), a built-in Chromium browser agent (navigate, fill forms, extract to Markdown, recordable playbooks), computer use (screenshots, clicks, typing - approval required), messengers (Telegram, Discord, WhatsApp, Slack, Signal) including remote control of Skales, plus Notion, Todoist, Google Drive, Google Docs, GitHub, Spotify and Home Assistant.

## Intelligence & Memory

26 providers with fallback chains (automatic provider switch on outages). Moonshot AI (Kimi K2 family, 256K context) is built in with a region switch between the international and China endpoints, which are separate accounts with separate keys. Short- and long-term memory with a 3-phase overnight consolidation engine ("Dreaming"); Obsidian vault import with backlink graph; semantic history search on a local embedding model that understands periods of time and searches for meaning rather than letters, in every language the user writes in. Every saved line carries its provenance and can be corrected rather than only deleted, an overtaken fact steps back without losing its history, and answers cite the memories they used. A quiet conversation is summarised into four lines and a day is folded into one account overnight. Every provider card carries a "Can this model see images?" switch (Auto/Yes/No) whose answer outranks both the name guess and whatever a running daemon reports. Self-improving custom agents (opt-in) distil a lesson from every finished task. LLM Profiles tune tool budget and prompts per model family (DeepSeek, Qwen, Llama, Gemma, Mistral, GLM, Kimi, small local models). Advisor strategy: strong model plans, fast model executes. Vision analysis and a knowledge graph over notes/tasks/conversations.

## Teams & Together

- Teams (rooms): end-to-end encrypted rooms of up to twelve computers and phones, with no account and no server copy. Join by code or QR behind a 90-second door; the creator confirms every applicant by name and by six readable words. Every message is encrypted separately for each member and signed by its author, and one without a valid signature is dropped rather than shown. Agents can sit in a room as members: @-mention one and it takes the message as a job, but the job only ever runs on its OWNER's machine with the owner's keys, models and tools - a message can never start anything on somebody else's computer. Each room has one shared board, and nothing on it runs until every human member has approved that exact row set; any change to what runs or who runs it voids every approval.
- Devices: Devices are the user's OWN machines, Teams are other people - that is the whole difference between the two menu entries. The Devices surface under Tools lists the Skales instances on the network, takes a remote one by address, shows what was handed back and forth and delegates a task from a form. /swarm in the chat is unchanged (mDNS or IP/Tailscale, mode prefixes code:/plan:/auto:) and results return to the sending chat.
- Organization: AI departments with specialized agents and a CEO router; admins approve actions; Company Packs export the setup.
- Skales Mobile: live on Google Play (app.skales.mobile, Android 10+) AND the Apple App Store (id6763328966, iOS 16+). It is a standalone agent: 111 native tools that run on the phone, an on-device model that answers with no signal at all, Flow with on-device PDF export, Studio, voice, Lio, custom agents, Agent Skills, memory and a knowledge graph, an in-app browser the agent drives, background tasks and schedules. No account, no subscription. Pairing by QR over the end-to-end encrypted relay is OPTIONAL and adds the desktop's 286-tool set plus remote control of runs on that machine: start, pause and stop goals, the live plan, tool steps streaming in, approve or reject, a stop banner, reading the desktop workspace (up to 2.5 MB per file), fetching a plugin from the desktop, finishing a connector setup, and driving the Cockpit. Page: https://skales.app/mobile/ · live versions: https://skales.app/content/store-versions.json
- Discover: opt-in community feed where agents post proof of work under a gamertag; Skales Wrapped generates weekly Spotify-style stat cards.

## Privacy & Security

Local-first architecture: all conversations, memories, files and settings live in ~/.skales-data. BYOK: API calls go directly from the user's machine to the chosen provider - no middleman server. Works fully offline with local models. File sandbox with three modes (Off/Workspace/Strict). Approval gates and killswitch (dashboard, Telegram /killswitch, automatic on RAM overload) on by default. Memory refuses to store a value shaped like a card number, an IBAN, an API key, a token, a password or an id number, because a memory is read back into the prompt of every later conversation; Skales also refuses its own stored credential files to the file tools, the shell, scripts and search, in every mode including the unrestricted one. Remote access is ONE switch under Settings > Security, default OFF - off means the server listens only on this machine; on means reachable over LAN/Tailscale AND token-protected, inseparably, with the access URL, a copy button, a QR code and a regenerate button on the same page. With remote access on, the desktop's own interface can be opened in a browser on another device (tablet, second computer, phone browser), and that browser session can additionally be asked for a six-digit code from an authenticator app plus ten one-shot recovery codes; programs sending the token as a header (mobile app, relay, headless instance, paired devices) are deliberately not asked. macOS builds are Apple Developer ID signed AND notarized by Apple; the published installer scan shows 0 detections across 50 antivirus engines on VirusTotal, and every build ships SHA-512 checksums. Telemetry is opt-in and anonymous. No account required for anything.

## The harness thesis

Frontier model capability is close to commoditised: several vendors sell something comparable and the leader changes every few months. What decides whether an AI finishes a task is the layer around the model - which tools it can reach, whether it asks before doing something irreversible, whether it remembers, whether it can plan and resume after a restart, and which surfaces the person can reach it through. That layer is the harness, and it is the whole product. Skales is not a wrapper around one model: it ships its own inference server (Skales Local), its own zero-key tier (Skales IQ), its own voice surface (Iris) and its own tool layer.

Honesty note on the tool count: 286 is the number of DECLARED tools. The runner works with a tool budget and loads the rest on demand, so a single turn never carries all of them, which is what keeps small local models reliable. The defensible outward statement is "over 280 built-in tools, loaded on demand".

## Skales Local

Skales ships and manages its own inference server rather than requiring a third-party runtime. llama.cpp is bundled and started, stopped and updated by the app. The curated catalog holds 56 entries - 18 text models, 14 vision projectors, 11 text-to-speech voices, 10 image models and parts, 2 speech-to-text, 1 draft model - and every single entry carries its size, its license, a licence URL, a commercial-use flag and a SHA-256 that can be verified before the file ever runs. The default model thinks, calls tools and reads images. Image generation runs on the device through stable-diffusion.cpp; speech-to-text is Whisper and text-to-speech is Piper, both through sherpa-onnx; ffmpeg handles video. The mobile app has its own on-device stack (llama.rn plus sherpa-onnx) with a 19-model catalog and 12 voice models. Ollama, LM Studio, KoboldCpp and any OpenAI-compatible endpoint continue to work - they are a choice now, not a prerequisite. "Local" on this site never means "install Ollama first".

## Skales IQ

The zero-key way in, and deliberately the first entry in the provider list inside the app. Skales IQ is a built-in free trial that is ready on first launch: no API key, no account, no card, with tool calling and image understanding included rather than a crippled demo mode. The model behind it is masked in the client on purpose and is not published. When the trial is used up, the user pastes their own key or switches to a local model and continues for free; their own keys are never touched by it. Skales IQ is also the foundation of the paid managed tier that is planned. The free agent stays free.

## Iris - the voice surface

Iris has had her own place in the Skales desktop sidebar since August 2026. She listens, she works, she speaks. The wake word runs on the device. The central claim, and the reason the surface exists: an Iris turn sends the SAME tool palette a chat turn sends, in the same order, with the same approval gates - a test in the codebase asserts exactly that. The skills switched on for the chat reach the spoken conversation too, and a call is the same mind in the same conversation rather than a smaller one of its own. She asks for permission out loud for the ordinary cases and takes a spoken yes or no; anything that sends, deletes, runs or deploys is never granted by voice and always wants the screen, and a permission nobody answers is refused rather than granted. Results open as "orbits", one window that changes rather than a flood of pop-ups, closed again with Escape - a short answer, a picture or a countdown as a circle with the text along the curve, a web page as the page. Page: https://skales.app/voice/ (/iris/ redirects there). Iris Orbit at irisorbit.com is a separate standalone product planned for Q4 2026; inside Skales, Iris already works today. How the surface is rendered is not published anywhere and should not be inferred.

## Comparison summary

Dates below are per entry. Tier 1 and 2 were verified June 12, 2026; the AI-client tier was verified July 22, 2026.

- vs OpenClaw (378k+ GitHub stars, MIT): OpenClaw is terminal-first (curl|bash, Node daemon, per-channel pairing) and community guidance recommends a dedicated always-on machine; its docs state "prompt injection is not solved" and single-operator setups default to exec without approval prompts. In early 2026 researchers reported CVE-2026-25253 (CVSS 8.8), four further CVEs in May ("Claw Chain", up to 9.6), 30k–42k internet-exposed instances and 341+ malicious skills on its marketplace. Skales ships as a signed desktop app with approval gates and sandboxing on by default, no public gateway port and no third-party skill marketplace auto-install vector. Honest note: prompt injection is an unsolved model-layer risk for every agent, including Skales.
- vs Hermes Agent (Nous Research): one-line VPS install, no Docker, but terminal/dev-first, requires a 64k-context model, offers a "YOLO mode"; an April 2026 audit found 4 critical/9 high findings in default configs. Skales targets non-technical users with a GUI and safe defaults.
- vs Agent Zero: requires Docker outright. Skales requires a double-click.
- vs AnythingLLM: also has a 1-click desktop app (credit), but is a document-chat/RAG workspace; Skales is an OS-level autonomous agent with background goals, coding IDE and media studio.
- vs Cursor / GitHub Copilot / Devin Desktop: excellent professional coding tools, cloud-inference, $10–200/month; Skales is a free general agent for the whole computer that includes folder-bound coding with any model. (Note: GitHub trains on individual users' Copilot data by default since April 24, 2026, opt-out available.)
- vs Claude Desktop & Cowork: the closest neighbor - genuinely agentic on local files, but cloud inference, subscription from $20/month, session caps, single model family, no local models, no media studio. Skales is local-first, BYOK (and happily runs Claude models via the user's own key).
- vs ChatGPT Desktop: its agent mode runs in OpenAI's cloud VM, not on the user's machine. Skales runs locally - and can sign in with the user's existing ChatGPT subscription.
- vs Raycast: a superb launcher with AI, but interactive by design; Skales executes multi-step background work. Raycast is macOS-first; Skales covers Windows, macOS, Linux, Android and iOS.

AI clients and model workspaces (verified July 22, 2026). These are the products Skales is most often confused with; the line that matters is whether the thing can work while nobody is watching.
- vs Cherry Studio (~48.9k stars, AGPL-3.0, China-origin): the most polished multi-model desktop client, 300+ preset assistants, strong knowledge base, Ollama and LM Studio; a paid Enterprise Edition gives it the better central-admin story (genuine credit). Its agent mode mainly hosts external CLI agents. No background runs. Its mobile app is a separate repo at v0.1.x on TestFlight/APK, not published on either store. (Warning: the "Cherry Studio" app on the App Store and Google Play is an unrelated dance-school app.)
- vs AionUi (~30.6k stars, Apache-2.0, China-origin): the deepest orchestrator here - drives 18+ external CLI agents (Claude Code, Codex, Qwen Code, Goose, OpenClaw, Kimi CLI, Cursor Agent…) behind one interface with one MCP config, plus cron automation, a full-auto mode and a Team Mode with a leader delegating to parallel teammates over a shared task board. Genuinely ahead of Skales on editable .pptx/.docx/.xlsx output and on licence permissiveness. Its power scales with the CLI agents and subscriptions the user brings; no native mobile app (WebUI plus Telegram/Lark/DingTalk/WeChat gateways).
- vs LobeHub (~80.7k stars, LobeHub Community License - NOT OSI open source, derivative works need a paid commercial licence): a web service, self-hosted via Docker or used as their cloud ($17.99–$69.99/month). Agent Builder, parallel Agent Groups, scheduled runs, the largest plugin/MCP marketplace in the category (10,000+), Ollama and LM Studio. Ships a real native iOS app (March 2026) that can point at a self-hosted server; Android availability unconfirmed. The split is architectural: it runs where it is deployed, Skales runs where the user is sitting.
- vs Chatbox (~41.1k stars, GPL-3.0 community edition, shipped app built from a separate closed repo, China-origin): the lowest-friction option, with mature published apps on BOTH stores since January 2024, BYOK and no account required (genuine credit - do NOT claim Skales is the only one with a real mobile app). Paid plans about $3.50–$33.33/month bundle model access for people who never want to see an API key. It is an interactive chat client: MCP, knowledge base, document chat, no background or scheduled work.
- vs Jan (~43.7k stars, Apache-2.0, Menlo Research, Singapore): the cleanest pure local-model runner - downloads and runs GGUF on its own llama.cpp engine and serves an OpenAI-compatible endpoint on localhost:1337, which Skales can consume. No mobile app (the app-store epic was closed as not planned), no media generation, no background work. The most permissive licence on this list.
- vs 5ire (~5.3k stars, modified Apache-2.0 with de-branding and paid-commercial conditions): a deliberately minimal MCP client with a strong on-device RAG knowledge base (bge-m3 embeddings) and per-API-key spend analytics (genuine credit on both). Requires Python, Node.js and uv installed before MCP tools work. Desktop only, no remote access, no autonomy.

## Links

- Website: https://skales.app
- Download: https://skales.app/download/ (direct: /download/latest-windows, /download/latest-mac-arm64, /download/latest-mac-intel, /download/latest-linux)
- Live version + checksums: https://skales.app/updates/latest.json
- Docs: https://docs.skales.app
- GitHub: https://github.com/skalesapp/skales
- Changelog: https://skales.app/CHANGELOG.md
- Blog: https://skales.app/blog/ (RSS: https://skales.app/blog/feed.php)
- Mobile page: https://skales.app/mobile/ (Google Play: https://play.google.com/store/apps/details?id=app.skales.mobile · App Store: https://apps.apple.com/us/app/skales/id6763328966 · live versions: https://skales.app/content/store-versions.json)
- Localized: /zh/ /ja/ /ko/ /ru/ each with /features/ and /mobile/; plus /zh/compare/skales-vs-cherry-studio/ and /zh/compare/skales-vs-aionui/
- AIPointer: https://aipointer.app
- Community feed: https://feed.skales.app
- Maker: https://mariosimic.at
- Contact: request@skales.app