OmniRoute pools free provider quotas into one combo so DeepSeek Harness codes without paid APIs.
OmniRoute is a self-hosted GitHub repo with over 50,000 stars that connects to over 50 providers, many requiring no credit card. It intelligently determines which providers have free quota, picks those models, and serves them into your IDE, such as the DeepSeek harness, through one API.
Through the connected providers you can get approximately 1.5 billion tokens of usage continuously each month. Some providers give an additional sign-up bonus, so if you take advantage of those, your first month can reach just over 2 billion tokens of usage.
The free API calls are not private, so anything you send should be something you are happy to share publicly. Also, some providers might not be happy that you are routing and optimizing for free usage, so there are warnings and you do this at your own risk.
Go to omniroute.online, copy the command npm install g omniroute and paste it in a terminal; installation takes about five minutes. Then run the omniroute command to spin up the dashboard and server. The presenter suggests first asking ChatGPT or Claude to audit the code for safety.
Copy the API key from the terminal, open DeepSeek settings, go to models and add a custom provider named omni-route. Paste the base URL, keep the OpenAI completions protocol, use a placeholder API key like 1234, and add a model set to auto for immediate free usage.
Sign up on OpenRouter, create a new API key with a useful name and copy it. In the privacy settings, turn on the options that are disabled by default, since free models collect your information. Then paste the key in OmniRoute, check it works, and import only free models.
NVIDIA has Nemotron models that are fast and very smart. You sign up for a free account, generate an API key in settings, paste it into OmniRoute and import free models. The GPT OSS 120 model tested successfully, and OpenRouter plus NVIDIA alone built the demo apps.
A combo plugs all your providers into one API call for DeepSeek. The presenter named his bart-combo using the free stack template, kept OpenRouter's auto endpoint, and chose only working NVIDIA models: GPT-OSS 120B, 20B, Nemotron 3 Super and Nemotron 3 Ultra, using the default round robin strategy.
If you plug in all providers and the coding agent hits a model that does not work, that breaks the entire coding session. So you test each endpoint, cherry pick out the failing ones and leave only working models, though you should still check every so often.
No. With 10 or 20 different models plugged in, quality varies: one call might get a big smart model and another a small, less smart one. The presenter would not use it for anything production, only for fun, though a Pong game was built in about 30 seconds.
“In this video, I'm going to show you how to use the DeepSeek Harness completely for free,”
“This is probably the best tower defense game that I've built.”
“OmniRoute is a tool that will be self-hosting on our computer, and we'll be plugging in a bunch of different providers directly into OmniRoute.”
“And the two things you have to watch out for are while these API calls are free, they're not private.”
“And then make sure that you change your password straight away. Do not leave it as change me.”
“NVIDIA has got a bunch of really cool Nemotron models, which are fantastic.”
“Use DeepSeek Harness for FREE — No VRAM, No Paid API.” https://www.youtube.com/watch?v=3g28CmoapOw. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/use-deepseek-harness-for-free-no-vram-no-paid-api.html.