Atlas

Atlas is a local-first, privacy-focused desktop AI workspace and agent studio. Bring Your Own Key (BYOK), connect to any model provider, run MCP tools, orchestrate autonomous agent workflows, and render live interactive visual artifacts — all stored securely on your own machine.

Modern AI workflows shouldn't be trapped behind walled gardens, expensive monthly subscriptions, or opaque proprietary servers. Atlas is built from the ground up to give developers, researchers, and power users full sovereignty over their AI environment.

• 🔑 True BYOK Freedom — Bring your own API keys directly to OpenRouter, Anthropic, OpenAI, Google Gemini/GLM, or local models (Ollama, LM Studio, vLLM). Pay only for what you consume at provider rates, with zero middleman markups.
• 🔒 Privacy & Local-First Architecture — Your conversations, prompts, workspace state, and model catalogs remain on your local disk in SQLite. All API keys are encrypted in your operating system's native keychain (keytar).
• 🤖 Agentic Workflow & Extensibility — Move seamlessly between standard chat, planning, and autonomous execution modes with built-in Model Context Protocol (MCP) support, custom plugins, and safety-gated tool execution.
• 🎨 Rich Visual Artifacts — Beyond plain text: stream and interact with live sandboxed HTML/UI documents, Mermaid diagrams, React Flow node graphs, LaTeX math formulas, and code diffs.
• ⚡ Desktop-Native Craft — Engineered with a refined warm minimalism aesthetic, command palette (cmdk), deep keyboard navigation, integrated terminal, and smooth view transitions.

• OpenRouter Integration — Real-time model catalog synchronization with free-tier model discovery and parameter controls.
• Major Model Providers — Anthropic Claude, OpenAI, and Google Gemini / GLM.
• Custom & Local Endpoints — Connect any OpenAI-compatible provider, including Ollama, vLLM, LM Studio, or self-hosted inference servers.
• Reasoning & Thinking Tokens — Real-time streaming and inspectable thought traces for reasoning models (DeepSeek-R1, Claude 3.7 Sonnet, OpenAI o1/o3-mini, Gemini Thinking).

• Multiple Workspace Modes:
  • Chat Mode — Clean, conversational interface for rapid Q&A, writing, and brainstorming.
  • Plan Mode — Structured step-by-step reasoning and architectural planning before code execution.
  • Agent / Act Mode — Autonomous agent execution with tool calling, self-correction, and progress reporting.
  • Review Mode — Review diffs, verify tool actions, and inspect command execution safely.
• Safety-First Tool Execution — Local tools for file reading, grep search, glob matching, web search/fetch, and bash execution with user approval checkpoints.
• Context Injection & @mentions — Reference files, workspace directories, plugins, and MCP resources directly inside your prompt.

• First-Class MCP Client — Connect to any MCP server via Stdio or SSE to equip models with custom tools, resources, and prompt templates.
• Extensible Plugin System — Load and manage plugins with automated security audits, connector verification, and marketplace discovery.
• Integrated Terminal — Embedded terminal sessions powered by @xterm/xterm and node-pty for local debugging.

• Sandboxed Visual Documents — Inline live rendering for HTML, web applications, and UI components with full-screen expansion.
• Interactive Diagrams — Native support for Mermaid diagrams and Dagre / React Flow node-link graphs.
• Code Highlighting & Diffs — Shiki-powered syntax highlighting with copy buttons, language tags, and diff views.
• LaTeX Math Equations — Formatted mathematical notation via KaTeX.
• Token & Cost Intelligence — Built-in token lens for real-time prompt token estimation and per-session cost tracking.

• Framework: Electron + React 19 + TypeScript + Vite (electron-vite)
• AI Core: Vercel AI SDK (ai), OpenRouter Provider, Anthropic, Google, and OpenAI adapters
• State Management: Zustand
• Styling & UI: Tailwind CSS v4, Motion (Framer Motion), Radix UI primitives, Lucide icons, cmdk
• Database & Storage: better-sqlite3 (with WAL mode), OS Keychain via keytar
• Terminal & Editor: @xterm/xterm, node-pty, Shiki
• Protocols & Standards: Model Context Protocol (@modelcontextprotocol/sdk)

• Node.js >= 22.0.0
• pnpm >= 10.0.0
• Native Build Tools: Python 3 and C++ build tools (required by better-sqlite3, keytar, and node-pty).
  • macOS: xcode-select --install
  • Linux: build-essential libsecret-1-dev
  • Windows: Visual Studio C++ Build Tools

# Clone the repository
git clone https://github.com/olllayor/Atlas.git
cd Atlas

# Install dependencies and build native addons
pnpm install

# Start Atlas in development mode with HMR
pnpm dev

Note for macOS developers: pnpm dev creates a rebranded app launcher with an isolated development bundle ID (com.olllayor.atlaschat.dev) so test keychain entries don't collide with production. Set ATLAS_SKIP_DEV_LAUNCHER=1 to bypass the dev launcher.

# Type check and bundle for production
pnpm build

# Package desktop installer (DMG, ZIP, etc.)
pnpm package

Atlas is distributed as an unsigned DMG via GitHub Releases. macOS Gatekeeper may show a security prompt upon initial launch because the binary is built without an Apple Developer certificate.

1. Visit GitHub Releases.
2. Download the DMG corresponding to your architecture:
   • Apple Silicon (M1/M2/M3/M4): Atlas-*-arm64.dmg
   • Intel Macs: Atlas-*-x64.dmg
3. Drag Atlas into your /Applications folder.

If macOS displays "Atlas can't be opened because it is from an unidentified developer" or "Atlas is damaged":

Option A — System Settings (Recommended):

1. Open System Settings → Privacy & Security.
2. Scroll down to the Security section.
3. Click Open Anyway next to the Atlas notice.

Option B — Terminal:

xattr -d com.apple.quarantine /Applications/Atlas.app

Atlas is engineered with security and privacy as core design tenets:

1. Process Isolation: The renderer process executes in a sandboxed context without direct Node.js access. All OS APIs, database access, and network requests are mediated by a strictly typed preload bridge.
2. Key Security: API keys are never stored in plain text or SQLite files. They are saved directly into the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service).
3. Local Privacy: Conversations, system prompts, attachments, and settings never leave your computer, except for direct HTTPS requests made to your chosen AI providers.
4. Execution Safeguards: Terminal commands, file mutations, and tool invocations feature explicit user approval gates before running.

• Multi-Provider BYOK Engine (OpenRouter, Claude, OpenAI, Google Gemini/GLM, Ollama/Custom)
• OS Keychain Key Storage & Local SQLite Database
• Model Context Protocol (MCP) Client (Stdio & SSE)
• Reasoning / Thinking Token Streaming & Visualizer
• Local Tool Calling & Approvals (Bash, File I/O, Search, Web)
• Live Visual Documents & Artifacts (HTML sandboxes, Mermaid, React Flow diagrams, LaTeX)
• Workspace & Directory Binding with @mention context injection
• Integrated Terminal (node-pty + xterm.js)
• Custom Plugin System & Auditing
• Full-Text Conversation Search across message history
• Conversation Branching & Forking
• Local Vector Embeddings & Long-Term Memory
• Automated Multi-Provider Fallbacks & Health Routing
• Conversation Export & Import (Markdown, JSON, PDF)

Contributions from the community are welcome! Please check out CONTRIBUTING.md for development workflows, code standards, and submission guidelines.

This project is licensed under the MIT License.