Welcome to QARK
QARK is a Quick Agent Relay Kernel — a native desktop runtime where AI agents live, compose, and execute. Connect any provider, attach any document, dispatch any agent. No accounts, no cloud, no telemetry. Your keys, your data, your machine.
Start in under five minutes
Section titled “Start in under five minutes”- Install QARK — Download the installer for your platform or install via your package manager
- Add a provider key — Open Settings → Providers, paste your API key from Anthropic, OpenAI, or any supported provider
- Dispatch your first agent — Create a new conversation, type a prompt, hit Enter. Watch tokens stream, costs track, and tools execute in real time
Explore the docs
Section titled “Explore the docs” Getting Started Install, connect a provider, and dispatch your first agent.
Providers Cloud and local providers, your own API keys, zero lock-in.
Core Features Conversations, groups, tabs, Sparks, RAG, prompts, budget, command palette, MCP.
Core Concepts The mental models behind QARK — agent kernel, context strategies, @mentions, fine-grained control.
Agents Create, compose, and chain agents with typed input/output schemas.
Tools Built-in tools and MCP extensions. Type @ to activate.
Universal Overlay System-wide AI overlay with Sparks and Flows — available from any application.
Advanced Keyboard shortcuts, settings, performance tuning, multi-agent workflows, custom MCP servers.
What makes this different
Section titled “What makes this different”QARK is not a chat wrapper with a model dropdown. It is an agent kernel — every prompt dispatches a configured agent with a system prompt, tool vector, context strategy, and model assignment:
- Multi-provider, single runtime — Anthropic, OpenAI, Gemini, Groq, Together, xAI, OpenRouter, Perplexity, DeepSeek, Cohere, Ollama, LM Studio, and more. Switch mid-conversation. Compare in split view.
- Agent-as-tool — Agents call other agents. A project manager dispatches a researcher and a writer, each with their own tools and models. Recursion depth, typed schemas, execution traces.
- RAG that understands code — Drag in PDFs, DOCX, codebases. Code-aware chunking splits by syntax, not paragraphs. HyDE and step-back queries. Cross-encoder reranking. Inline citations with relevance scores.
- Universal overlay — A system-wide AI panel triggered by a global hotkey from any application. Highlight text, type an instruction, and the result pastes back at the cursor or copies to clipboard.
- Privacy by architecture — Local SQLite database. AES-256 encrypted API keys. No accounts, no cloud sync, no analytics, no telemetry. Your data never leaves your machine (except to the providers you choose).
- Per-token cost tracking — Every message shows input tokens, output tokens, thinking tokens, and cost in USD. Monthly budgets per provider with real-time warnings.