PewDiePie has not spent 2026 trying to launch another chatbot with a celebrity name attached to it. He has been building the opposite: a free, self-hosted AI workspace that keeps its data on hardware you control.
The project is called Odysseus. Felix Kjellberg released it publicly on May 31, and the code has kept moving since then. The interesting story now is not simply that a famous YouTuber made an AI app. It is that his personal experiment is turning into a large open-source project concerned with the unglamorous parts of local AI: model compatibility, memory, email, security, hardware detection, installation bugs, and whether the whole thing starts correctly on somebody else's computer.
Odysseus running in the browser. Image source: the official Odysseus GitHub repository.
What Odysseus actually is
Kjellberg describes Odysseus as a self-hosted alternative to the web interfaces offered by ChatGPT and Claude, but the current project reaches much further than a chat screen. It combines chat and agents with files, shell access, MCP servers, skills, and memory. It also includes deep research, model comparison, document editing, email, notes, tasks, a calendar, image tools, web search, and two-factor authentication.
The Cookbook may be its most practical feature for people who are curious about local AI but do not want to memorize every model format and serving backend. It scans the computer, recommends models that should fit the available hardware, downloads them, and helps serve them inside Odysseus. That tackles one of the most frustrating parts of running AI locally: knowing whether a particular quantized model will work on your GPU or unified memory before wasting hours downloading it.
Odysseus can also connect to API models, so self-hosting the workspace does not force every user to run the language model locally. The important distinction is that the workspace, conversations, files, and integrations remain under the user's control.
Why he built it
In the launch video, Kjellberg said the project grew out of his attempts to self-host AI. He liked the models he could run at home but found the surrounding experience incomplete. Local chat interfaces did not automatically give him the memory, research, agents, integrations, and personal workflow he had become used to elsewhere. He started building those missing pieces for himself, initially as a joke, then kept going when the result became useful.
His argument for local AI is straightforward. An assistant gets more useful as it learns about your work, documents, preferences, calendar, and past conversations. That same context also makes it unusually personal. Keeping the workspace on your own system reduces the amount of that context entrusted to a platform provider.
That does not make Odysseus magically private in every configuration. If you connect it to a cloud model or external service, data needed for that request may still leave your machine. Self-hosting gives you control, but privacy still depends on which providers and integrations you enable.
What is new in Odysseus right now
The latest work is less about adding another flashy panel and more about making the existing system reliable across real machines.
Apple Silicon and MLX support
One of the biggest recent additions is first-class MLX support for Apple Silicon. Odysseus can use Apple's MLX framework through mlx-lm or oMLX, expose those models through an OpenAI-compatible endpoint, and surface suitable MLX models in the Cookbook on compatible Macs. The change was tested end to end on an M4 Max with a Qwen model and working tool calls.
This matters because Docker on macOS cannot directly use the Metal GPU. A local AI workspace may technically run in a container yet feel painfully slow if the model falls back to the CPU. Native Apple Silicon support lets Odysseus use the hardware people actually bought the Mac for.
Better support for smaller local models
The project's roadmap admits that agent mode currently carries too much prompt and context overhead for small models. Tool schemas, memories, skills, documents, and instructions can consume a large chunk of a 4K, 8K, or 16K context window before the user asks anything.
The proposed work is sensible: slimmer prompts, smaller default tool sets, smarter tool selection, and clearer behavior for constrained models. Recent changes have also focused on more reliable memory writing and recall, while an open pull request adds parsing for the tool-call format emitted by Qwen-family models.
Email is becoming a real part of the workspace
Kjellberg's launch demonstration included an AI-assisted email client that can identify urgent messages, summarize a mailbox, prepare replies, and support reminders. July's repository activity shows developers hardening that feature rather than leaving it as a demo. Recent fixes cover Google OAuth settings in Docker, OAuth-based IMAP and SMTP connection tests, mailbox identity checks during reconnection, and protection against leaking OAuth tokens through configuration responses.
This is exactly where privacy-first AI becomes difficult. Reading local documents is one thing. Handling a live inbox requires authentication, network calls, account isolation, secure token storage, and careful testing.
Security and retrieval cleanup
The team has added CodeQL scanning and continued closing server-side request forgery risks. A current security fix addresses a DNS-rebinding gap in outbound API integrations. Another recent change prevents document indexing from swallowing hidden folders and junk directories such as .git, node_modules, and __pycache__. That should reduce indexing time and keep irrelevant code or metadata out of retrieval results.
The roadmap also calls for a prompt-injection audit across skills, notes, documents, fetched pages, and memories. That is worth watching. An agent with access to files, email, and shell commands is far more useful than a chat-only model, but mistakes have a larger blast radius too.
Less glamorous bugs are now the main work
At the time of this review, open work included fixes for Windows virtual-environment activation, failed startup caused by a missing Python type import, Markdown code blocks disappearing when a user edited a message, theme handling on the login page, concurrent file-write collisions, and model downloads that leave orphaned processes on Windows.
That list is not a criticism. It is what a fast-growing application looks like after thousands of people try it on different operating systems and hardware. The project roadmap is unusually honest: squash bugs, test fresh installs, audit integrations, improve error logs, and stop pretending every backend behaves the same everywhere.
It is no longer only PewDiePie's personal project
Kjellberg still has the largest number of recorded contributions, but Odysseus now has a broad contributor list and thousands of issues and pull requests in its history. The repository moved from his pewdiepie-archdaemon account to the odysseus-dev organization in July. Its default dev branch receives new work first, while main is intended to be more curated.
That changes how the project should be described. PewDiePie started and publicly launched Odysseus, but many of the current fixes and features are community contributions reviewed and merged into the project. Saying that he personally wrote every new change would be inaccurate.
Should you try it?
Odysseus is worth exploring if you already care about local models, self-hosting, or owning the workspace around your AI. It may also be useful if you want one interface for local and API models instead of scattering conversations, documents, searches, and automations across several subscriptions.
It is not yet the obvious choice for someone who wants a polished, zero-maintenance ChatGPT replacement. The default branch moves quickly, the open issue count is high, and the roadmap still asks for fresh-install testing across Linux, Windows, WSL, and macOS. The README points stability-minded users toward main, while dev receives the newest changes first.
The larger idea is more compelling than the celebrity angle. Odysseus treats AI as a personal computing layer rather than a website you visit. Chat, memory, research, documents, email, and agents become parts of a workspace that you can inspect, modify, and host yourself. PewDiePie may be the reason many people hear about it, but the project's future will depend on whether its community can make that ambitious bundle trustworthy and boring enough to use every day.
References
Originally published at https://blog.jenuel.dev/blog/what-pewdiepie-is-building-in-ai-now-odysseus-july-2026
Thanks for reading! If you enjoyed this article and like this kind of content, you're always welcome to buy me a little coffee, but only if you'd like to. No pressure at all, and either way I'm truly grateful you stopped by. ☕️
Top comments (0)