About this video
12% off Minimax coding plan: https://samuelgregory.co.uk/minimax Try Odysseus: https://github.com/pewdiepie-archdaemon/odysseus PewDiePie just released Odysseus, a free open-source, self-hosted AI workspace that runs entirely on your own hardware. So I jumped straight on and got MiniMax M3 running inside it. No script, no prep, just a live walkthrough of the Docker setup, wiring up M3 through a custom endpoint, and putting it to work. It went well in places and hit a wall in others, and the reason why is the interesting bit. Spoiler: the model was never the problem. It's all about the harness. KEY TAKEAWAYS - Odysseus is genuinely local-first and privacy-first. No telemetry, no subscription, your data stays on your machine. - The Docker route is the easy way in, but it boxes the agent in. No headless browser and no reaching outside the container, so web tasks fall over. - MiniMax M3 is seriously capable. It handled file tasks instantly and barely broke a sweat. It's a frontier-level open-weight model and a fraction of the cost of the big closed ones. - The bottleneck in local AI right now is rarely the model. It's the harness around it that decides what it can actually do. - Two quick gotchas: watch out for the port 80 clash, and set your password in the .env file BEFORE you run docker compose up. - If you want full capability, skip Docker and run it natively on Apple Silicon instead. Does it replace OpenClaw, Hermes or Pi for you? Drop a comment and let me know. And tell me if you want the local model or native Apple Silicon version next.
Odysseus and MiniMax M3: A Dockerised Reality Check
Docker is effectively a prison for your AI agents and you are the one who built the bars. While we often praise containerisation for its ability to protect our systems, we rarely discuss how it lobotomises the very tools we are trying to build.
The Odysseus Experiment
Taking a leaf out of the book of the UK cheese farmer and PewDiePie, I decided to dive into Odysseus. The goal was simple: get MiniMax M3 sorted and see what this model can actually do when paired with a proper harness.
I opted for the Docker route. On paper, it makes sense. It protects your computer and manages dependencies. In practice, it turned into a battle against port allocations and hidden admin credentials.
The Password Problem
If you are following the GitHub instructions, word of advice: set your password in the environment variables before you even think about running Docker Compose. The terminal output might not give you what you need, and if you do not have a tool like Warp to help you navigate the logs, you might find yourself locked out of your own creation.
MiniMax M3: The Model is Great, the Harness is Limited
Once I got the API integrated (using the Anthropic endpoint workaround), MiniMax M3 performed admirably. It is fast, intelligent, and handles complex queries with ease. However, the Docker container acted as a wall.
When I asked the agent to list files for deletion or download a podcast, it hit a hard limit. It cannot reach outside the container. It cannot install a headless browser on the fly.
The Verdict
If you want a simple tool for email integrations or direct API tasks, the Docker setup for Odysseus is fine. But if you want a true autonomous agent that can interact with your local environment, you have to run it natively on your hardware.
MiniMax M3 is a powerhouse, but it deserves to be let off the leash.
Keep on vibing, and remember to check the links for a discount on the MiniMax coding plan.
Transcript▾
Shout out to the UK cheese farmer for giving me the idea for Odysseus. I forgot what PewDiePie said the name was, but I should have made this video about getting MiniMax M3 sorted on Odysseus. I have literally not planned anything, I am jumping on and just going to do this video because it should be fairly easy to set up.
If you go to PewDiePie's GitHub, I will leave links below. If you scroll down, we have some quick starts. If you are going to run MiniMax M3, go for the Docker route because we want to containerise the whole installation. This protects your computer, however, if you are going to run local models, we have to run it natively. In the meantime, let's go the Docker route.
I use Warp for my terminal. We are going to clone Odysseus and CD into it. We will copy the environment variables and then run docker compose up on localhost 7000. My Docker was already running something on 8080, so we had to kill that old process. Once started, I had to find the admin password. My suggestion: set the password in the ENV file before you even run Docker.
I integrated the MiniMax API. It picked up M3 immediately. I tried to ask it to list candidates for deletion in my downloads folder, but it could not. It is inside Docker, it cannot reach outside. I tried to get it to download a podcast, but it lacks a headless browser. This is the problem with running stuff inside Docker, you have to give it everything it needs before you run the software.
MiniMax M3 handled the logic easily, it is the harness itself that has limitations. If you want a video on running this natively on Apple Silicon, let me know below. Like and subscribe, and get 12% off the MiniMax coding plan with my links below. Keep on vibing.