A desktop AI that actually stays off the cloud
Thoth is a local-first assistant that routes your data to your own machine, not someone else's server farm.

What it does Thoth is a desktop AI assistant built around the idea that your chat history, documents, and voice transcripts belong on your hard drive, not in a SaaS database. It bundles chat, a personal knowledge graph, developer tools, design studio, workflow automation, and messaging channels into a single installable app. Models can run locally via Ollama or connect to hosted providers, but durable data stays local.
The interesting bit The project treats model routing as a first-class problem rather than an afterthought. It tracks provider identity, context limits, and whether a model can handle tools or should fall back to chat-only mode, so mixing local, hosted, and self-hosted endpoints doesn’t produce surprise failures. The knowledge graph layer is unusually explicit: 10 entity types, 67 typed relations, confidence decay, and an Obsidian-compatible export.
Key highlights
- Native desktop app with one-click installers for Windows and macOS, one-line curl install for Linux
- 30+ built-in tool modules covering shell, browser automation, Git workflows, calendar, email, and file parsing
- Developer Studio with repo inspection, diffs, commits, and an optional Docker sandbox with shadow workspace
- Designer Studio for decks, mockups, and storyboards with sandboxed runtime and export to PDF/PPTX
- Voice pipeline using local faster-whisper and Kokoro TTS with 10 voices
- No account system, no Thoth-hosted server, no telemetry; keys stored in OS credential store
Caveats
- Linux defaults to browser-first launch; native window and tray need GTK, Qt, and AppIndicator libraries installed
- Local models need generous context windows—the README warns that 4096 tokens can fail before the first chat turn, with 32768 recommended for agent mode
- Browser automation on Linux may require manual Playwright dependency installation
Verdict Worth a look if you want an all-in-one AI workspace without subscription lock-in or data egress. Probably overkill if you just need a chat wrapper for GPT-4; the value is in the integrated tooling and local persistence.