Odysseus

Open Source

4.1 / 5

About

Odysseus is a self-hosted, open-source AI workspace created by Felix Kjellberg (PewDiePie). It bundles multi-model chat (local runtimes Ollama/vLLM/llama.cpp and cloud APIs OpenAI/OpenRouter), an autonomous agent built on opencode with full MCP support, deep research with source cross-validation, email management (IMAP/SMTP with AI triage), calendar (CalDAV sync), notes, tasks, and a hardware-aware model recommendation system (Cookbook) into a single Docker-deployable package. The Cookbook scans GPU/VRAM/RAM and recommends the best-fitting models from a 270+ model catalog with one-click download and serve. Built with FastAPI + Python backend and vanilla JavaScript frontend, it runs fully locally with zero telemetry — no accounts, no subscriptions, no data collection, and all data stays on the user's own machine.

Key Features

Multi-Model Chat — unified interface supporting local runtimes (Ollama, vLLM, llama.cpp) and cloud APIs (OpenAI, OpenRouter, GitHub Copilot)
Autonomous Agent — built on opencode with MCP support; tools include shell execution, file operations, web browsing, Python execution, and memory read/write
Cookbook Hardware-Aware Model Manager — scans GPU, VRAM, and RAM, then recommends compatible models with fit scores from 270+ catalog; one-click download and serve via optimal backend
Deep Research — multi-step web research with source cross-validation; generates structured visual reports with citations
Email Assistant — IMAP/SMTP inbox with AI triage, auto-classification, smart one-line summaries, and context-aware draft replies
Calendar & Tasks — local-first CalDAV calendar (Radicale, Nextcloud, Apple, Fastmail), todo lists, and cron-style scheduled automations
Memory & Skills — ChromaDB vector memory with hybrid retrieval (vector + keyword via fastembed ONNX), self-evolving reusable skill packs
Blind Compare — anonymous side-by-side model comparison to eliminate brand bias; identities revealed after judgment
Documents — multi-tab editor (Markdown, HTML, CSV) with syntax highlighting and AI-assisted writing
Mobile PWA — fully responsive Progressive Web App with touch gesture support
2FA Security — built-in two-factor authentication with admin/authorized/user role tiers

Use Cases

Personal AI Workspace, Local Model Chat, Autonomous Task Automation, Deep Research & Report Generation, Email Management & Triage, Calendar & Schedule Management, Model Evaluation & Comparison, Privacy-First AI Assistant, Developer Copilot, Homelab Self-Hosted AI

Pros & Cons

Pros

Unmatched feature integration — chat, agent, research, email, calendar, notes, model management in one deployable package
True zero-telemetry, local-first privacy architecture — all data stays on user hardware with no accounts required
Cookbook eliminates the hardest part of local AI — automatic hardware-matched model recommendations with one-click serve
Completely free and open-source (AGPL-3.0) — no subscriptions, no paywalls, no usage limits
MCP protocol support enables extensible agent capabilities beyond built-in tools (browser control, custom services)
One-command Docker Compose deployment — works on Windows, macOS (Apple Silicon with Metal), Linux, NAS
Self-evolving memory and skills system — ChromaDB vector memory persists across sessions and learns user preferences
Responsive PWA with mobile-first design — usable from phone via Termux; some features reportedly developed from mobile
Massive community momentum — 71K+ stars, 8K+ forks, 104+ contributors in first two weeks

Cons

High hardware requirements — full experience needs 12GB+ VRAM GPU; CPU-only mode is significantly slower and less capable
Not production-ready for teams — lacks enterprise features like multi-tenancy, audit logging, and granular RBAC
Broad security attack surface — shell + email + browser + MCP co-located in one process; sandboxing is an acknowledged gap (see THREAT_MODEL.md)
Young project with rapid iteration — stability and long-term maintainability unproven at two weeks old
Celebrity-driven hype — part of the 71K stars comes from PewDiePie's 110M subscriber base rather than technical merit alone
Built by assembling existing open-source projects (opencode, Tongyi DeepResearch, llmfit, SearXNG) — some critics call it a "Python UI on top of other projects"
Windows users need WSL — no native Windows support; Docker required for all platforms
Agent context window bloat — on 8K-context local models, tool schemas can consume significant context before user prompts

Reviews

No reviews yet

Be the first to share your experience with this agent.

Write a Review

Pricing

Open Source / Free (self-hosted)