Odysseus is a self-hosted AI workspace that runs on your own machine or server. Built around chat and autonomous agents, it folds deep research, document editing, email, notes, tasks, and a calendar into a single application, so you can complete your AI work without handing data to an external cloud. You can switch between local models and API-based ones, and the project describes itself as “local-first, privacy-first, no telemetry.” It was published on GitHub in late May 2026 under the AGPL-3.0 license, and the whole stack comes up with Docker Compose.
Key Features
- Chat and agents share the same foundation: Alongside ordinary conversation, you can use autonomous agents that plan their work and call tools as they go. Tools for bash, file operations, the web, and memory are built in, and each one can be toggled on or off individually
- MCP server integration: Built-in MCP servers are registered automatically at startup. A Playwright-based browser MCP is also available once you set it up, giving you page navigation, screenshots, and visual page understanding (the steps are collected in the setup guide)
- Cookbook picks models that fit your hardware: Based on your machine’s specs, it recommends models you can run and handles everything from download to serving. According to the official site, more than 270 models are catalogued. If you already have an OpenAI-compatible endpoint such as Ollama or vLLM, you can simply connect to it instead
- Deep research and blind comparison: A deep-research mode searches the web across multiple steps, reads the sources, and writes up a full report. The Compare feature sends the same prompt to several models at once and lets you evaluate the answers side by side without knowing which model produced which
- Built-in IMAP/SMTP inbox: You can hand off inbound-mail summaries, reply drafts written in your own voice, automatic tagging, and spam triage to the AI. Gmail and Google Workspace can also be connected through your own OAuth2 client
- A working environment that extends to documents, notes, and calendar: It includes a document editor with AI-assisted editing, to-dos and reminders, scheduled agent tasks, and a CalDAV-synced calendar. A gallery for generating and editing images and two-factor authentication are included as well
How to Think About Cost
There is no paid plan or subscription. The only costs are whatever you provide yourself.
- The application: Free. It is AGPL-3.0 open source, with the code published on GitHub
- Compute: If you run local models, your own hardware, including GPU and VRAM, is the main cost. The app itself is lightweight; the heavy part is serving the models
- API models: Calling external models through an OpenAI-compatible endpoint incurs usage-based charges. You can mix these with local models
- Bundled services: Web search, a vector store, and notifications start up alongside the app, so no external service contracts are needed
Pros and Cons
✅ Pros
- Chat, agents, research, documents, email, and notes live in one app, so you no longer hop between tools for each task
- Designed to send no telemetry, so you can feed private data and work email to the AI without it leaving your environment
- You can route sensitive processing to local models and harder tasks to external APIs
- Cookbook handles model selection and serving, which lowers the barrier to getting started with local LLMs
- Starts with Docker Compose, with the surrounding services coming up together
⚠️ Cons
- It is a young project, published in late May 2026, and no tagged formal release has been distributed yet. The default
devbranch is the most current but is described as potentially unstable, with a more stablemainbranch maintained separately - The high-priority items on the official roadmap include “bug squashing,” “verifying fresh installs on each OS,” and “auditing whether integrations actually work,” so maturity varies relative to the breadth of features
- Installation requires Docker or a Python 3.11+ environment. On Apple Silicon Macs, Docker cannot access the Metal GPU, so native execution is needed if you want GPU acceleration
- Because of AGPL-3.0, offering a modified version to others over a network triggers source-disclosure obligations. If you plan to embed it in internal systems or redistribute it, a legal review is warranted
- In a network-exposed configuration, maintaining authentication and blocking ports is the operator’s responsibility
Comparison with Similar Services
| Criteria | Odysseus | Open WebUI | LibreChat | AnythingLLM |
|---|---|---|---|---|
| License | AGPL-3.0 | Custom license (BSD-3 based) | MIT | MIT |
| Focus | All-in-one workspace | Chat UI for local/API models | Multi-provider chat | Local-first agents and document use |
| Email, notes, calendar | Built in (IMAP/SMTP, CalDAV) | Not included by default | Not included by default | Not included by default |
| Local model serving support | Cookbook handles recommendation, download, and serve | Centered on Ollama integration | Connects to external endpoints | Connects to various providers |
| Maturity | Published May 2026, pre-release | Established mainstay | Established mainstay | Established mainstay |
This comparison is based on each project’s official repository as of August 2026.
Who It’s For
- People who don’t want to hand work email or personal notes to an external cloud, but still want AI assistance
- People who want to try local LLMs but are stuck at the starting line, unsure which models their machine can run
- People currently juggling chat, research, document writing, and email processing across separate services and want to consolidate them
- People who want to build an AI environment on their own hardware without adding more subscriptions
Summary
Odysseus starts from self-hosted AI chat and grows into an ambitious workspace that takes on agents, research, documents, email, and notes. Its strengths are the sheer breadth of what one app covers and the on-ramp to local models that Cookbook provides, making it a strong option for anyone who wants to keep data in their own hands. At the same time, it is only recently public, and the author’s own roadmap places bug fixing and install verification at the top of the list. The sensible first step is to spin it up with Docker on a test machine and see how it fits.