Self-hosted AI workspace that refuses to stay in its lane

What it does

Odysseus is a web-based AI workspace that runs entirely on your own machine. It bundles chat, agentic tasks, document editing, deep research, email triage, calendar sync, and note-taking into a single interface that works on desktop and mobile. You can plug in local models via vLLM, llama.cpp, or Ollama, or use remote APIs like OpenAI and OpenRouter, and a built-in hardware scanner recommends models that actually fit your available VRAM.

The interesting bit

The project treats self-hosting as a feature rather than a compromise. It includes an AI-powered email client that auto-tags and drafts replies over IMAP/SMTP, a CalDAV-aware calendar, and a blind side-by-side model comparison tool, all wrapped in a UI that openly admits to having “more jank and fun” than polished SaaS alternatives. The Cookbook feature is particularly unusual: it probes your GPU setup, scores model compatibility, and can download and serve GGUF, FP8, or AWQ weights automatically.

Key highlights

• Chat and agent tools built on opencode with MCP support, plus persistent memory via ChromaDB and fastembed
• Cookbook scans hardware and recommends VRAM-aware models, supporting vLLM, llama.cpp, and Ollama backends
• Deep Research adapted from Tongyi DeepResearch for multi-step source gathering and synthesis
• Full email client with AI triage, auto-summary, and spam detection over standard IMAP/SMTP
• Calendar, notes, tasks, and a multi-tab document editor with markdown and syntax highlighting
• Progressive web app with mobile support, 2FA, and theme editing

Caveats

• GPU passthrough in Docker requires manual host runtime configuration, and passing nvidia-smi into the container does not guarantee llama.cpp can actually see CUDA; you may need to reinstall the serve engine
• Windows users cannot locally serve vLLM or SGLang natively and must rely on Ollama or WSL2 for GPU-backed models
• The README repeatedly warns against exposing the instance directly to the public internet; it defaults to loopback binding and treats the app as an admin console because non-admin privileges still require careful review

Verdict

Worth a spin if you are tired of tab-switching between ChatGPT, your email client, and your calendar. Look elsewhere if you want a zero-config hosted service or if the phrase “GPU passthrough” makes you break out in hives.