Skales

Skales is a private AI agent harness built by Mario Simic in Vienna, Austria. It installs like a normal desktop app on Windows, macOS, Linux, Android, and iOS, and positions itself as the layer between the user and any AI model — handling tools, approval gates, memory, planning, and multiple interaction surfaces. The current version is v12.9.35, released September 23, 2026, and the project ships updates weekly.

What It Is

Skales describes itself as an "AI harness" rather than a chatbot. The distinction the product draws is that frontier model capability is increasingly commoditized, and what determines whether an AI actually completes real work is the infrastructure around the model: which tools it can reach, whether it asks before doing something irreversible, whether it remembers context across sessions, and whether it can plan and resume after a restart. Skales provides that infrastructure layer, while the underlying model — Claude, GPT, Gemini, Grok, DeepSeek, Mistral, or a locally-run model — is treated as a swappable component. The app stores classify it as an "AI agent & task" app.

How the Autonomy Stack Works

The core autonomy feature is the Goal Engine. Users type /goal followed by a plain-language objective, and Skales plans the steps, runs them in the background with the chat window closed, and parks with a decision card when it needs human approval before an irreversible action. Goals survive restarts: if the app closes mid-task, it resumes from where it left off. Multiple goals can run in parallel, and recurring schedules can repeat them automatically.

Supporting the goal engine are several layers:

• Cockpit 2.0 — an execution board showing pending, in-progress, completed, and blocked goals as draggable cards
• Sub-Agents — parallel workers that each get their own context and model assignment
• Approval gates — mandatory human confirmation before sending, publishing, deleting, or spending
• Emergency killswitch — one-click stop for all AI activity, also triggerable via Telegram /killswitch
• Session budget ceiling — a per-conversation cost cap that stops and asks the user at half and at the limit

Surfaces and Interaction Modes

Skales reaches users through multiple surfaces rather than a single chat window:

• Chat with /goal, /code, /spin, and /swarm commands
• Skales Code — a full coding window with command palette, hunk-by-hunk diff review, and git integration
• Lio AI — a beginner-friendly code builder where one model designs, one reviews, and one builds
• Skales Studio / Flow — a creative suite covering design, image, video, audio, and motion, with Brand Kits and 10 output modes
• Iris — a voice surface with a wake word that runs on-device, full tool parity with chat, and orbit-style result display
• Desktop Buddy — a floating mascot (gecko, blob, or capybara) that is a full agent with approve/decline in its speech bubble
• AIPointer — a cursor-anchored overlay triggered by a key hold or cursor wiggle
• Skales Mobile — a native Android and iOS app with 131 standalone tools and optional QR-paired desktop control

Local-First Architecture and Privacy Model

All conversations, memories, files, and settings are stored in ~/.skales-data on the user's machine. API calls go directly from the user's computer to the chosen AI provider (BYOK — bring your own key), with no Skales cloud intermediary unless the user opts into Skales IQ. Skales Local ships its own bundled llama.cpp inference server with a catalog of 56 vetted models, covering text, vision, speech (Whisper in, Piper out), and image generation (stable-diffusion.cpp) — all offline, no API key required. Ollama and LM Studio remain supported as alternatives. The memory system explicitly refuses to store values shaped like card numbers, IBANs, API keys, tokens, or passwords.

Model and Provider Support

The app ships with 26 built-in providers including OpenAI, Anthropic, Google Gemini, xAI Grok, DeepSeek, Mistral, Groq, Together, MiniMax, Moonshot AI (Kimi), GLM, Qwen (DashScope), Hunyuan, GigaChat, Cloudflare Workers AI, NVIDIA NIM, Hugging Face, and OpenRouter, plus Skales IQ (the built-in zero-key trial), Skales Local, Ollama, LM Studio, KoboldCpp, and any OpenAI-compatible endpoint. Users can also sign in with an existing ChatGPT subscription without an API key. Models can be switched mid-conversation without a restart.

Update: v12.9.35 — "Roles" Release

The latest release (v12.9.35, published September 23, 2026) is named "Roles" and adds the ability to report answers from Iris, Flow, Lio, and Desktop Buddy (not just the main chat window), playable WhatsApp voice messages, per-agent model selection for sub-agents, and skill descriptions displayed alongside skill names. The v12.9 series introduced Cockpit 2.0, Skales Code as a standalone coding window, Teams rooms for up to twelve members, a plugin shelf, per-conversation spending ceilings, and a memory system that refuses to write down secrets. The GitHub repository reports 1,921 stars and the product claims over 47,100 installs worldwide as of the homepage copy.