FAQ

Frequently asked questions

Short answers to the things people ask most. For depth, follow the links into the docs.

What is Koryphaios?

A native desktop AI workspace that orchestrates a team of coding agents against your real project — a manager plans, workers build in parallel, and a critic reviews before anything reaches you. See How it works.

Does it sell me AI usage / tokens?

No. You bring your own models — API keys or the agent CLIs you already have. Koryphaios is the orchestration layer, not a reseller of inference.

Which models and providers work?

50+ providers: API-key ones (OpenAI, Google direct API, Google AI Studio, Vertex AI, Anthropic, xAI, Groq, OpenRouter, Azure, Bedrock, Mistral, and OpenAI-compatible endpoints) and subscription CLIs (Claude Code, Codex, Grok, Cursor, Devin, Cline, Antigravity). You can mix them in one picker. See Providers.

Does my code leave my computer?

No — the backend runs locally and works on your files in place. The exceptions are the ones you choose: an API/CLI model necessarily sees the context you send it, and if you use model sharing your project is sent to the host running a shared CLI. Nothing else is uploaded.

Can I use it offline?

The app itself runs locally, but the models don't — you need network access to reach a hosted provider. If you point Koryphaios at a local model (Ollama, LM Studio, llama.cpp), you can work fully offline.

What platforms are supported?

Windows, macOS (Apple Silicon & Intel), and Linux (.deb, .rpm, AppImage). One note: the model-sharing sandbox can use kernel enforcement on Linux and macOS when the host supports it; on Windows it's a softer jail — details in The remote sandbox.

Is it safe to share my subscription with a friend?

Sharing your own API keys is fine. Sharing a subscription (Claude, Copilot, ChatGPT plans) is account-sharing that several providers' terms forbid and enforce. The app flags this per provider and defaults subscriptions off. See Share your models.

Can the AI break something irreversibly?

Every AI change is captured as a ghost commit, so you can rewind the workspace to any earlier state. Combined with the critic gate and per-tool permissions, mistakes are recoverable. YOLO mode removes the confirmations if you want speed over caution.

How are the notes different from Obsidian?

Same core idea (wikilinks, backlinks, a graph) but native to the workspace and — the real difference — readable and writable by your agents. It now covers the features people miss most: live Markdown preview, math (KaTeX) and Mermaid diagrams, a Dataview-style query language, aliases, ghost nodes, callouts, a plugin API, a spatial canvas/whiteboard, and full-text search that stays fast at thousands of notes (the graph renders on a GPU-friendly canvas with no node cap). The honest gaps: a mature third-party plugin ecosystem and a native mobile app. See Notes.

Is there a mobile app?

Not yet as a shipping native app. The desktop UI (including notes) is touch-friendly and responsive down to phone widths. A packaged iOS/Android build is on the roadmap — it needs its own device testing and store pipeline, so we'd rather ship it right than ship it untested.

How do updates work?

The app checks for updates and can install them in place; a banner appears when one is available, and the changelog lists what changed each release.

Is it open source? What does it cost?

The desktop app is free to download and run — you only pay your own model providers. Grab it from the download page.