Sep 9, 2026
skales-desktop-agent: Run Agent Skills in a Local-First Desktop AI
Skales is a 1,842-star desktop agent for Windows, macOS, Linux, Android, and iOS that imports the SKILL.md format used by Claude Code, keeps its data in a folder on your machine, and works with 15+ AI providers or fully offline.
Most agent skills are written for terminal-native coding tools, which leaves everyone who does not live in a CLI without a harness. Skales — "the private AI that lives on your computer and does real work, not just chat" — is a desktop agent that installs like a normal app and imports the same SKILL.md files your technical teammates use.
Why This Skill Matters
Skales runs on Windows, macOS, and Linux as a desktop app, with Android and iOS apps that pair to it — the README describes a personal agent with access to your files, browser, calendar, and email that executes multi-step work on its own. Two details matter for the skills ecosystem: it imports the Agent Skills format used by Claude Code, Codex, Copilot, and Cursor — from a GitHub URL, a local folder, or pasted SKILL.md — and imported skills work across both its Chat and Code surfaces. The README also points at a community directory of 1,000+ skills to import from.
On the model side there is no lock-in: start on Skales IQ, a free built-in trial that needs no API key, bring your own key from 15+ providers (Anthropic, OpenAI, Google, OpenRouter, Groq, DeepSeek, Mistral, and more), or run fully offline with Ollama, LM Studio, or the bundled Skales Local. Data lives in ~/.skales-data on your machine.
One honest caveat up front: Skales is closed source under BSL 1.1 — free for personal, educational, and non-commercial use, with a license that converts to Apache 2.0 on 2030-04-19. The repo distributes releases; the checked-in source is a frozen v7 snapshot, not the shipping code.
Installation
Download the installer from skales.app — no Docker, no terminal:
- macOS: signed DMG, drag to Applications.
- Windows: EXE installer (the README notes binary signing is coming).
- Linux:
.debfor Debian/Ubuntu/Mint or an AppImage for everything else. - Mobile: Skales Mobile on Google Play and the App Store.
To import a skill, hand the agent a GitHub URL, point it at a local folder, or paste a SKILL.md into the chat.
Real Workflow: Import a Skill and Hand Over a Goal
Import the skill you want, then delegate outcome-style work in chat:
/goal Collect this week's RSS entries from my industry blogs and draft a
summary doc in my documents folder
/goal turns a request into ongoing background work: Skales plans the steps and runs them on its own, continues with the laptop lid closed, and picks up where it left off when you reopen the app. It stops when the task is done, when it genuinely needs a decision, or before a consequential action like sending an email — which it asks about once, with a one-tap always-allow. The Cockpit screen shows every running goal with its success criteria, the evidence collected so far, and its spending.
For code work, /code binds a folder to the conversation with four modes — Plan investigates, Code asks before each edit, Edits approves file edits as it goes, Auto runs the whole task after a one-time consent — with inline diffs, one-click undo, and commits under your own git identity. Pairing your phone by QR gives it the desktop's full tool set (the README counts 260+ tools) over an end-to-end encrypted relay.
Tips
- Turn on Privacy Mode before connecting a cloud provider: your memory index, name, and preferences then never leave the machine, whatever the memory mode.
- Memory has three levels — Always Remember, Compact, Minimal — so you decide how much context travels with a request.
- Switching from OpenClaw, Hermes Agent, or ChatGPT: Settings has a migration importer.
- Connect an Obsidian vault as a folder Skales reads and appends to — it can add notes but never overwrite yours.
- Set a spending ceiling per conversation under Settings → Goals; Skales pauses at half of it and at the limit to ask how to proceed.
When Not to Use This
If you need auditable source, the license answers that: the product is a signed closed-source app, and the repo says so plainly. Commercial SaaS or competing products require written permission under BSL 1.1. And if your workflow is terminal-first CI automation, a coding agent in your editor will fit the same skills with less app around them.
See the leaderboard for more skills.