Keep exploring · more guides
The tool is free. The AI brain inside it is free too.
Most people install it, hit the box asking for a credit card, and quietly give up.
You do not have to.
I am going to show you the exact switch that runs it at zero, on my own machine.
Every number on this page came off that machine today.
You download a free tool.
You install it. It opens. It looks brilliant.
Then it asks for an API key before it will do a single thing.
So you paste in a card, and now every question you ask has a tiny price on it.
You stop experimenting. You start rationing.
The tool was free. The thinking was not.
That is the meter, and this page turns it off.
This one has a 200,000 word working memory and answered me at 108 words a second today.
It is DeepSeek's own fast model. You will see it run further down this page.
DeepSeek Harness landed on 13 August as a developer preview.
Free, open source, MIT licensed — download it, change it, build a business on it, ask nobody.
A model can think, but it cannot touch anything — no files, no memory, no browser.
The harness is the hands, the memory and the workspace wrapped around it.
Two separate bills, and only one of them is zero by default.
You need Node.js first — it is a free download from nodejs.org, click through it like any installer.
Then open Terminal on a Mac, or PowerShell on Windows, and type this one line.
What you are watching: the real install running on my machine — it pulls everything down, then hands you a local address.
It runs at 127.0.0.1:3080 and opens in your normal browser — chat window, sidebar, folder picker.
The page is yours: it runs on your machine, nobody else is hosting it.
What you are watching: the Harness running locally — this is the whole interface, in a browser tab.
This is the whole trick, and it takes about twenty seconds.
You are adding a second brain next to the paid one, and it costs nothing to run.
What you are watching: the real Settings → Models page on my machine. DeepSeek carries a red dot because I never gave it a card. OpenCode Zen is green — ready, and free.
Click the model name at the bottom of the composer, open Model, and it is sitting under OpenCode Zen.
Then give it a real job and watch it answer.
What you are watching, unedited and at real speed: I open the model picker, choose DeepSeek V4 Flash Free, type a real client task, and it writes the note back. No key. No card.
These are the numbers the Harness and the provider reported back on that exact job.
I tested both today so you do not waste an evening on the wrong one.
Sign up at opencode.ai, copy your key, paste it into the provider. The free models still bill at zero — the key just identifies you.
The free endpoint answers with no key at all, but refuses a wrong one. A tiny local relay strips the header. This is what is running in the videos above.
One signup, no moving parts, nothing to keep running. Route B is there if you want zero accounts.
If you want Route B, the relay is one file and one command.
Then point the provider's base URL at http://127.0.0.1:8788/v1 and give it any key you like.
Route A is the normal, supported way and will not break.
Route B is only for people who want no account at all — and I have told you exactly what it does, so you can judge it.
Free does not mean infinite. Push it hard enough in one day and you get this back, word for word:
I hit it myself while testing for this page, so you would find out here rather than halfway through a job.
For a normal day of drafting, summarising and tidying files you will not notice it — and when you do, you switch the dropdown to a local Ollama model, or to Pro, and carry on.
The same Add provider button takes OpenRouter's free models, or Ollama running fully offline on your own machine.
Flash is fast and free, so it should do most of your day.
Pro is DeepSeek's flagship and it is not free — save it for the big builds.
Click the folder icon and point it at one tidy folder — that is the only place it can read or write.
Then write what you want in plain English, the way you would brief a person.
What you are watching: a real brief going in, the agent listing the folder, opening files one at a time, and writing a new file back out.
This page shows you the free brain. The Boardroom is where you point it at the jobs that actually grow a business.
Start a new session, switch the preset to Creator mode, and describe something you wish existed.
The dropdown you watched me use earlier was built exactly this way.
What you are watching: Creator mode taking one plain sentence and building a working panel into the interface, then asking permission to switch it on.
A free harness on its own is good. A free harness inside a system that already knows your business is a different thing entirely.
What you are watching: the Harness running as one tab inside my Agent OS, next to Claude, Hermes and the rest — all reading the same memory vault.
Wrong: "Free AI means weak AI that cannot do real work."
Right: V4 Flash read seventeen thousand tokens and answered in five seconds on a real client job. Free is the price, not the quality.
Wrong: "This is too technical for me — I am not a developer."
Right: You install one thing, paste one line, and click a dropdown. The hardest part is choosing which folder to point it at.
Wrong: "I will wait until all this settles down."
Right: It will not settle. The people who learn the free stack now are the ones who will not be paying for it later.
Members post their wins every day — agency owners, ecom founders, course creators and solo operators across 38 countries, in their own words.
Read the 158-page wins doc →No — that is the biggest myth about it.
The everyday ninety percent runs on free local models on your own machine, and free APIs like the one on this page.
For the heavy work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, so you are not paying twice.
And there are full token-efficiency tutorials inside the Boardroom, so you learn to cut usage to the bone.
Free, open source, running on your own machine at localhost:3080.
DeepSeek V4 Flash Free through OpenCode Zen — 200k context, 108 tokens a second.
It reads and writes there, and nowhere else on your computer.
Brief it like a new assistant and it does the job in the place the work lives.
Creator mode builds panels into the interface while you watch.
Drop the harness into an Agent OS and every agent shares the same memory.
The harness is free and the brain is free. What still costs you a year is working out which jobs to hand it, and building the system that runs them while you sleep. That part is already built, and it is waiting inside the Boardroom.
