Odysseus: A Self-Hosted AI Workspace

Architecture and Core Functionality

Odysseus functions as a personal operating system for language models, built to run entirely on local hardware without telemetry or external accounts. The backend is built on FastAPI, SQLAlchemy, and SQLite, prioritizing stability and simplicity for single-user deployment. The frontend notably avoids modern frameworks like React, utilizing plain JavaScript and CSS with no runtime dependencies in the package.json file. Retrieval and memory are handled via Chroma DB using local embeddings to ensure data privacy.

Agentic Capabilities and Integration

The workspace features an agent loop capable of executing shell commands and interacting with MCP servers, allowing it to process local files and automate tasks. Key integrated features include:

• Cookbook: A model-fitting engine that detects local VRAM capacity to recommend, download, and serve compatible open-source models.
• Deep Research: A multi-step pipeline that reads sources and synthesizes reports.
• Communication Suite: An integrated email client using IMAP and SMTP, alongside a calendar with CalDav synchronization for task and schedule management.

Licensing and Community Development

The project launched under the MIT license but transitioned to AGPL 3.0 following community feedback regarding the protection of open-source contributions. While the codebase integrates components from existing projects—specifically the model-fitting engine from llmfit, the agent loop from opencode, and deep research logic from Alibaba's Tongyi DeepResearch—it has gained significant traction due to its all-in-one installation approach. Users should note that on Apple Silicon, Docker containers cannot access the GPU, requiring a native installation to utilize hardware acceleration for model inference.