Odysseus AI and Ajax Explained: Latest Updates and How to Use It

Quick answers

• What is Odysseus AI? A free, open-source, self-hosted AI workspace from PewDiePie's project that combines chat, AI agents, deep research, documents, email, notes, and calendar.

• Is Ajax AI out? No. As of October 15, 2026, Ajax (a fine-tune of Qwen 3.5 9B) has no official download.

• Is it free? Yes. Odysseus is licensed under AGPL-3.0-or-later. Hardware and paid APIs can still cost money.

• How do I install it? Clone the repo, copy .env.example to .env, run docker compose up -d --build, then open http://localhost:7011.

• Will it run on my PC? The app is lightweight. Running models locally is the demanding part, so small machines can connect an API model instead.

Odysseus AI is a free, open-source, self-hosted AI workspace: software you install on your own computer or server that brings chat, AI agents, deep research, documents, email, notes, and calendar into one browser interface. It is PewDiePie's (Felix Kjellberg's) AI project, which explains much of the attention it has received since its public GitHub discussions opened on June 1, 2026.

Ajax is something different. It is a planned AI model, a fine-tune of Qwen 3.5 9B built to run Odysseus tools reliably on modest hardware. As of October 15, 2026, Ajax has not been publicly released, and the project's own maintainers say its timing is handled outside the code repository.

This guide explains what each one is, what changed in October 2026, how to install Odysseus correctly today, and where its limits are.

What Is Odysseus AI? PewDiePie's Self-Hosted AI Workspace

The official Odysseus repository describes the project as "a self-hosted AI workspace for chat, agents, research, documents, email, notes, calendar, and local model workflows." The key word is workspace. Odysseus is not itself a language model. It is the application layer that sits on top of whichever models you connect to it.

Who made Odysseus AI?

The project's lead maintainer is the GitHub account pewdiepie-archdaemon, which signs posts as "Felix" (for example, in the June 5, 2026 fine-tune proposal). Several public signals tie it to PewDiePie. Community members address the account as PewDiePie in its pinned thank-you note. The Ajax model is announced on a pewdiepie.com subdomain, which maintainers reference directly in issue #6497. Contributors also describe Ajax training data as "PewDiePie's call for data."

The earliest pinned maintainer posts, including that thank-you note and a call for co-maintainers, are dated June 1, 2026. Other maintainers, notably Alexandre (alteixeira20), now lead much of the release engineering, so Odysseus is a community project, not a one-person tool.

Workspace vs model

A model (such as Qwen, Llama, or a hosted GPT or Claude model) generates text. A workspace decides what to do with that text: which tools to call, where to save a document, which emails to summarize, which sources to read. Odysseus is the second kind of software. You point it at a model running on your own machine, on another server, or behind a paid API, and Odysseus supplies everything around it.

How it differs from a hosted chatbot

With a hosted assistant such as ChatGPT, the provider runs the software, the model, and the storage. With Odysseus, you run the software and store your data in its local data/ folder. Whether your prompts leave your machine depends on which model and integrations you connect. That flexibility is the main appeal, and also the main source of setup work.

What Is Ajax AI?

Ajax is a planned, Odysseus-specialized language model fine-tuned from Alibaba's Qwen 3.5 9B. It is not yet publicly available.

The idea first appeared on June 5, 2026, when the lead maintainer proposed a dedicated Odysseus fine-tune on Qwen 3.5 9B. The stated goal was "a local Odysseus-specialized model that can run on more modest hardware while preserving reliable agent/tool usage." The proposal named three problems with current local models: they struggle with the full Odysseus agent prompt, they produce invalid tool calls, and users pick models that are not suited to agent work.

What fine-tuning means here

Fine-tuning takes an existing model and trains it further on examples of the behavior you want. The June proposal planned supervised fine-tuning (SFT) on about 20,000 filtered examples of successful Odysseus agent sessions, teaching the model to choose the right tool, produce valid arguments, recover from tool errors, and answer briefly. Its own example of the target behavior: a user says "remind me to call Ajax tomorrow" and the model calls the reminder tool correctly in under a second. The proposal listed llama.cpp (GGUF), vLLM, and a Q4 quantized build as deployment targets.

In the same thread, maintainer Alexandre argued for measuring prompt size and building evaluations before training, and the lead maintainer agreed to put prompt slimming first. Community posts in October 2026 confirm the fine-tune went ahead under the Ajax name, and an independent tester's discussion describes Ajax as fine-tuned from stock Qwen 3.5 9B.

What is confirmed and what is not

Confirmed from project sources: the base model (Qwen 3.5 9B), the goal (reliable local tool use inside Odysseus), and that Ajax was not released as of mid-October 2026. Ajax's announcement page is data.pewdiepie.com. We could not independently verify that page's capability or benchmark claims, the final weights license, or hardware requirements, so none are reported here. Treat any performance figures you see elsewhere as claims until the weights are public and independently tested.

Latest Odysseus AI and Ajax Updates in 2026

The biggest recent change is the "pre-Ajax migration": a large runtime and security overhaul merged into the public dev branch on October 6, 2026, which prepares Odysseus to run Ajax. Ajax itself has not shipped.

The pre-Ajax migration (October 3 to 6, 2026)

On October 3, maintainer Alexandre announced the migration. For several months, maintainers had developed tightly linked runtime changes in a separate environment so contributors would not face constantly shifting internals. That work covered process ownership, execution and containment boundaries, background jobs, browser execution, cancellation and recovery, and CI.

One change matters directly to users. The announcement says the runtime now distinguishes between a tool reporting success and Odysseus having evidence that the intended change actually happened. In practice, that should make agent actions such as "create a calendar event" more trustworthy.

The work went public as PR #6503 (tracked in issue #6502) and merged into dev on October 6, 2026. The maintainers stress that this is not a finished release: dev is now in a stabilization and backlog-cleanup phase, and nothing goes to main until broader validation is complete.

Ajax release status (October 4 to 13, 2026)

On October 4, a user opened issue #6497, "Bruh, wheres Ajax!?", reporting that data.pewdiepie.com returned a 404 error. Other users replied that the release was delayed. On October 13, a maintainer closed the issue, explaining that Ajax's release timing and the content of data.pewdiepie.com are not controlled by the repository, and that release news will come through the project's usual announcement channels.

Stabilization work since the merge

Commits on dev between October 12 and 15, 2026 are mostly fixes and refactors: authentication handling, email bulk-trash safety, ChromaDB volume mounting in Docker, Notes and Todos as agent context, and a new opt-in reasoning-effort setting. None of them ships Ajax. The repository's Releases page lists no formal releases; versions are published as Docker image tags instead.

All dates were checked against the linked GitHub pages on October 15, 2026.

Key Features of Odysseus AI

Odysseus bundles eight feature areas into one interface, according to the README. How well each works depends heavily on the model you connect and the integrations you configure.

AI Chat and Model Selection

You can chat with local models or API models. Providers are added inside Settings, not by editing code. Odysseus accepts OpenAI-compatible endpoints, which covers local servers such as Ollama, llama.cpp, and vLLM, and the roadmap lists provider setup checks for Anthropic, Gemini, Groq, xAI, OpenRouter, OpenAI, and DeepSeek as ongoing work. A Compare view runs blind side-by-side tests between models and can synthesize their answers.

AI Agents and Tool Integrations

Agent mode lets a model use tools: files, a shell, skills, memory, and MCP (Model Context Protocol) servers, a standard way to plug external tools into AI apps. A built-in browser tool based on Playwright is available after a one-time install. The setup guide notes that shell, Python, and file tools are admin-only by default, and recent commits added approval controls for running commands.

Deep Research and Web Search

Deep Research runs multi-step web research: it searches, reads sources, and writes a report. The pipeline is adapted from Alibaba's open-source Tongyi DeepResearch project, per the project's acknowledgments file. The Docker setup bundles SearXNG, a self-hosted metasearch engine, so searches go through your own instance by default.

Document Editing and Knowledge Work

The Documents editor is writing-first, with AI edits and suggestions, Markdown, HTML, CSV, and syntax highlighting. Optional extras add PDF viewing and text extraction from Office and EPUB files.

Email, Calendar, Notes, and Tasks

Email connects through IMAP/SMTP and offers triage, tags, summaries, reminders, and reply drafts. The setup guide warns that Outlook and Microsoft 365 accounts generally need OAuth, which the IMAP password login does not support yet. Notes, Tasks, and Calendar include reminders, scheduled agent tasks, and CalDAV sync. Push reminders use the bundled ntfy service.

Local Model Support

A local model runs on your hardware, so prompts stay on your machine. A remote API model runs on a provider's servers, so prompts are sent there. Odysseus supports both. Its Cookbook feature scans your hardware, recommends models that fit, and downloads and serves them, building on the open-source llmfit project. If you already run Ollama, you can simply add its endpoint in Settings instead.

How to Install and Set Up Odysseus AI

The recommended route is Docker Compose: four commands, then open http://localhost:7011 and log in with the generated admin password. The steps below follow the official README and setup guide as checked on October 15, 2026. Since Ajax is not released, these steps cover Odysseus only.

Step 1: Check the System Requirements

Separate the app from the model. The setup guide calls the app itself "lightweight"; local model serving is "the heavy part" and depends on the model, runtime, GPU, and VRAM.

• For Docker installs: Git, plus Docker with Docker Compose.

• For native installs: Python 3.11 or newer. Cookbook also needs tmux on Linux and macOS, and Git for Windows on Windows for background downloads and the agent shell tool.

• For running models locally: enough RAM or GPU memory for the model you choose. Larger models and higher-precision files need more memory; quantization (compressing a model to fewer bits, such as Q4) reduces that. For first tests on 8 GB laptop GPUs, the guide suggests starting with GGUF/Q4 models on llama.cpp.

Two platform notes: Docker on Apple Silicon Macs cannot use the Metal GPU, so local models run on the CPU there. vLLM and SGLang serving need Linux or WSL2 with NVIDIA or AMD GPUs. If your hardware is limited, you can skip local models entirely and connect an API provider.

Step 2: Choose a Supported Installation Method

Three methods are documented:

1. Docker Compose (recommended). Starts Odysseus plus ChromaDB (memory storage), SearXNG (search), and ntfy (notifications).

2. Native Python on Linux, macOS, or Windows. On Apple Silicon, the ./start-macos.sh script is the way to get GPU-accelerated local models. On Windows, launch-windows.ps1 does setup in one command.

3. Stack managers such as Portainer or Coolify, using the official image ghcr.io/odysseus-dev/odysseus.

Step 3: Download the Project

Clone the official repository. Pay attention to branches. dev is the default and gets changes first, but the setup guide says it "may be unstable," and it is currently in post-migration stabilization. main is the curated branch. Beginners should use main:

git clone -b main https://github.com/odysseus-dev/odysseus.git

cd odysseus

Step 4: Start Odysseus

cp .env.example .env

docker compose up -d --build

The first line copies the example settings file so defaults are explicit (the guide calls this optional but recommended). The second builds and starts all containers in the background. Wait until docker compose ps shows them as healthy.

For a production server, pin an exact image build rather than the moving :latest tag, by setting ODYSSEUS_IMAGE in .env to an immutable tag in the form X.Y.Z-<sha>. Current tags are listed on the container package page.

Alternative: Windows without Docker

The official launcher creates a Python environment, installs dependencies, runs setup, and starts the server. It is safe to re-run. From inside the cloned folder in PowerShell:

powershell -ExecutionPolicy Bypass -File .\launch-windows.ps1

You need Python 3.11 or newer, plus Git for Windows for Cookbook downloads and the agent shell tool. For a local model on Windows, the setup guide recommends Ollama as the easiest path. The app opens at http://localhost:7000.

Alternative: Apple Silicon Macs with GPU acceleration

Docker on macOS cannot use the Metal GPU, so run natively for faster local models:

./start-macos.sh

The script installs Homebrew dependencies, sets up Python, and starts Odysseus at http://127.0.0.1:7860. It uses llama.cpp or Ollama for Metal acceleration.

Alternative: native Linux

python3 -m venv venv

source venv/bin/activate

pip install -r requirements.txt

python setup.py

python -m uvicorn app:app --host 127.0.0.1 --port 7000

This creates an isolated Python environment, installs dependencies, runs first-time setup (which prints the admin password), and starts the server on port 7000. Install tmux if you plan to use Cookbook.

Step 5: Open the Interface

With Docker, open http://localhost:7011. Watch for a documentation mismatch: the README and the current docker-compose.yml map host port 7011 to the app's internal port 7000, while one paragraph of the setup guide still says http://localhost:7000 for Docker. For native installs started with uvicorn, the address is http://localhost:7000. The macOS start script uses port 7860 because AirPlay often occupies 7000.

On first boot, Odysseus creates an admin account (username admin by default) with a temporary password. Find it with:

docker compose logs odysseus

Log in, then change the password in Settings. By default the interface is only reachable from your own computer, which is the safe setting.

Step 6: Connect an AI Model

Add model providers in Settings. You have three main options:

• Existing Ollama on the same machine: start Ollama with OLLAMA_HOST=0.0.0.0:11434 ollama serve, then add the endpoint http://host.docker.internal:11434/v1 (Docker) or http://localhost:11434/v1 (native).

• Cookbook: let Odysseus recommend, download, and serve a model that fits your hardware.

• API provider: paste an API key for a hosted provider. These calls are billed by the provider and send your prompts to its servers.

Step 7: Run a Simple Test

Start a normal chat and ask something short, such as "Summarize what a self-hosted AI workspace is in two sentences." A reply confirms the model connection works. Then try agent mode with a harmless task, such as creating a test note.

If something fails, the guide's checks are docker compose ps and docker compose logs --tail=120 odysseus. If the port is taken, set APP_PORT in .env and recreate the container. On Windows, if .env changes are ignored, re-save the file as UTF-8 without BOM.

To follow Ajax, watch the Announcements discussions and the Ajax page. Do not download "Ajax" weights from unofficial sources.

How to Use Odysseus AI for Everyday Tasks

The examples below are illustrative prompts, not tested demonstrations. Results will vary with the model you connect, especially for agent tasks, which smaller local models handle less reliably.

Researching a topic

Goal: a sourced overview without opening twenty tabs.

How: open Deep Research and try:

Research the current state of home battery storage for small households in the UK. Compare three common system types, list typical trade-offs, and cite every source you used.

What affects the result: Deep Research depends on the search backend (the bundled SearXNG by default) and on the model's ability to read and synthesize long sources. The project's roadmap lists hardware-based model presets for Deep Research as unfinished work, so expect to experiment with model choice. Always open the cited sources before relying on a claim.

Summarizing a document

Goal: the key points of a long report.

How: upload the file to a chat or open it in Documents, then ask:

Summarize this document in five bullet points, then list any figures or deadlines it mentions.

What affects the result: chat uploads are capped at 10 MB by default (adjustable with ODYSSEUS_CHAT_UPLOAD_MAX_BYTES). Office and EPUB extraction need the optional markitdown package. Small models with short context windows may miss details in long files.

Organizing notes and tasks

Goal: turn messy notes into actions.

How: recent dev commits let you attach selected Notes and Todos as agent context. In agent mode, try:

Read my attached meeting notes, create a todo for each action item, and set a reminder for anything with a date.

What affects the result: this relies on correct tool calls, the exact weakness Ajax is meant to address. Reminders reach your phone only if ntfy is configured as described in the setup guide.

Experimenting with agents and comparing models

Start agents on low-risk tasks, such as creating notes or drafting (not sending) email replies, before allowing anything that changes external accounts. To judge which model suits you, run the same prompt through Compare, which tests models blind side by side.

Odysseus AI vs ChatGPT, Ollama, and Open WebUI

These tools solve different problems. ChatGPT is a hosted assistant: OpenAI runs the software, the models, and the storage, and you install nothing. It is the simplest option, but you cannot self-host it or swap in your own local model. The other three are self-hosted, and they often work together: many people run Ollama to serve a model and connect it to Odysseus or Open WebUI as the interface.

Rows are based on each project's official README as of October 15, 2026.

Odysseus stands out for built-in personal productivity (an IMAP inbox, CalDAV calendar, reminders, and notes) alongside agents and a hardware-aware model Cookbook. Open WebUI has a longer track record as a shared chat platform. Neither is better for everyone: choose Odysseus for a personal hub, Open WebUI for a team chat interface, and ChatGPT if you want no setup at all.

Is Odysseus AI Free, Private, and Safe to Use?

Is Odysseus Free and Open Source?

Yes. Odysseus is licensed under AGPL-3.0-or-later, an open-source license that lets you use, modify, and self-host it for free. If you modify it and offer it to others over a network, the AGPL requires you to share your modified source. Models you connect have their own licenses, separate from Odysseus. Ajax's license has not been published.

Free software does not mean free to run. Expect costs for hardware or a server, electricity, and any paid API provider you connect.

Does Odysseus Run Locally?

The application, its database, and the bundled search, memory, and notification services run on your machine, and user data is stored in the local data/ folder. Data leaves your machine when you use a remote API model (your prompts go to that provider), web search and Deep Research (queries reach search engines and websites), email and calendar sync (Odysseus talks to your mail and CalDAV servers), and remote MCP servers. With a local model and no external integrations, chat stays on your hardware.

What Should Users Know Before Connecting Accounts or Tools?

The setup guide says to treat Odysseus "like an admin console," because it can run shell commands, read files, download models, and act on email and calendars. Its main recommendations:

• Keep AUTH_ENABLED=true and LOCALHOST_BYPASS=false on any network-accessible install.

• Never expose the port directly to the public internet; use HTTPS behind a trusted reverse proxy or a private network such as Tailscale.

• Disable open signup after first boot, and keep only your own account as admin.

• Use separate API tokens per integration, and rotate any key that appeared in a screenshot or log.

Agents can also misread instructions. The roadmap lists an audit of prompt injection (malicious instructions hidden in notes, documents, or web pages) as high priority. Until you trust a setup, keep agents on draft-and-review tasks rather than sending email or changing calendars unattended.

Limitations of Odysseus AI and Ajax

Odysseus is young, fast-moving software, and its own roadmap opens by saying the project is "on a voyage, but not home yet."

Limitations the project documents

• Setup and maintenance. You install, update, secure, and back up the system yourself. The roadmap lists fresh-install testing on Linux, macOS, Windows, Docker, and WSL as a high priority, and asks contributors to check whether integrations "even work."

• Branch churn. dev is in post-migration stabilization, with more than a thousand open issues being re-evaluated. The documentation also lags in places, such as the 7000 versus 7011 Docker port mismatch.

• Small local models. The roadmap says agent mode is "too heavy" for smaller local models because tool schemas, skills, and memory can fill the context window before the user's request. This is the problem Ajax and prompt slimming aim to fix.

• Cookbook reliability. Model downloads and serving are flagged as the area most likely to break across different GPUs, drivers, and Python setups.

• Email. Outlook and Microsoft 365 mailboxes generally will not connect yet, and IMAP performance is listed for review.

• Fully offline use. The roadmap notes front-end assets still load from CDNs, with full vendoring planned.

Potential limitations to consider

• Hardware. Larger local models need substantial RAM or GPU memory, and Apple Silicon users lose GPU acceleration in Docker.

• API costs. Agent and Deep Research tasks can make many model calls, which adds up on paid APIs.

• Reliability of agent actions. The October 2026 runtime work adds verification of whether an action really took effect, but any model can still misunderstand a request.

Ajax: what remains uncertain

Ajax was unreleased as of October 15, 2026. Its weights, license, recommended settings, hardware requirements, and real-world performance are unconfirmed. Independent comparisons against stock Qwen 3.5 9B, such as the one prepared in discussion #6509, can only begin once the model is out.

Conclusion: Should You Try Odysseus AI?

Odysseus AI is worth trying if you are comfortable with Docker or Python and want one self-hosted hub for chat, agents, research, and personal productivity, with full control over which models it uses. Its open AGPL license, built-in email and calendar tools, and hardware-aware Cookbook set it apart from simple chat front ends.

Ajax is the missing piece the project is building toward: a small, tool-focused local model meant to make Odysseus agents dependable on ordinary hardware. The October 2026 pre-Ajax migration laid the groundwork, but the model itself had not shipped as of October 15, 2026.

If you want zero setup or need production-grade reliability today, a hosted assistant is the simpler choice. If you enjoy tinkering, start with the main branch, connect a model you already have, and keep agents on low-risk tasks at first.

Next step: visit the official Odysseus repository to install it, and watch the project announcements for Ajax's release.