DeepSeek Harness Setup: A Free Coding Agent You Actually Own

What DeepSeek Harness actually is

DeepSeek Harness is a coding agent you install on your own machine, and it is free because the part you were paying for was never the agent. The model is the brain. Everything wrapped around that brain is the harness: your context, your tools, your plugins, your MCP servers, your permissions, your memories. Claude Code is a harness where somebody else made every one of those decisions and then closed them. This ships the same job with every decision still open while it is running. It hit 190,000 stars faster than any repo on GitHub, and the install takes about five minutes.

The distinction matters more than it sounds. For two years the agent and the model came as one product. If you wanted the good model, you took whatever harness was bolted onto it. That stopped being true.

How to install DeepSeek Harness

Copy the install command from the front page of the site or from the GitHub repo, paste it into your terminal, and press enter. It takes roughly five minutes and opens a local web dashboard on its own when it finishes.

The first time that dashboard loads it asks for a DeepSeek API key. You do not need to paste one in. Click configure it later. This is the step most people get stuck on, because it looks like a paywall and it is not one.

How to point it at a model you already pay for

Go to Settings, then Models, then add a provider.

The provider list that ships with it includes Anthropic, OpenAI, Bedrock and Vertex. So DeepSeek wrote the harness and then put its competitors in the settings menu. You are not stuck with DeepSeek's paid API just because DeepSeek wrote the software.

I pointed mine at a GLM subscription I already pay $10 a month for, which is more than enough for this. Apply it and the indicator goes green.

For anything not in the list, use add a custom provider. You need four things:

• Provider ID
• Display name, which can be anything you want
• Base URL
• API key

That is also how you run it against a local model. If you have Ollama running, the base URL points at your own machine and there is no meter at all.

Is it actually free

The harness is free forever. The model is the only part anybody can ever charge you for, and you get to choose who that is.

Be careful with the middle option. Free hosted models have caps. Google's free Gemini tier works here, but they stopped publishing their limits, so you find the ceiling by hitting it. The only genuinely uncapped free path is running the weights on your own machine, which needs a decent graphics card. Ollama has a one line install for exactly that.

I do not have that hardware, so I am not going to show you something I cannot actually run.

What "everything is a plugin" really means

It is the headline on DeepSeek's own front page, and they mean it far more literally than you would expect.

Open Settings, then Plugins. The sidebar you are looking at is a plugin. The commands are plugins. You can disable any of them with a click, and enable them again with the play button. There is no privileged core that you are locked out of.

That has a practical consequence. Pull a component out of a normal coding agent and it breaks. Lose a dependency halfway through a task and it breaks, unless you hardcoded a plan B in advance. Here, every component declares what it needs, and the runtime keeps an undo available for every change it makes. Nothing is hardcoded, so nothing has to restart.

There is a preset called creator mode that loads the plugin development skill automatically, which is what you want if you are building rather than using.

Building the plugin every agent should already have

I gave it a real job rather than a demo: a plugin that fires a desktop notification when a run finishes, with the token count and what it cost me inside the notification.

That is the thing I actually wanted. I lose runs constantly. An agent works for nine minutes in another window, finishes, and I do not notice for half an hour because nothing told me.

It built it. The first version had black text on a black background, which you could only read by highlighting it, and a broken counter. I told it to fix both and it did. That is roughly the shape of the work.

The part that made me pay attention: traceability

While the job ran, I could watch everything it did while it did it.

Open the trajectory on any run and you get:

• The exact system prompt it used
• The context it injected, word for word
• Which skills loaded, which did not, and why
• Every tool call, with the payload it sent and the result it got back
• The thinking
• How long each step took

Not a summary of those things. The actual values. If you have ever tried to work out why an agent did something strange and had nothing but its final message to go on, you already know what this is worth.

Killing a run mid task and resuming it

Everything above about components coming and going safely sounds like marketing until you test it, so I killed a task while it was running.

Then I told it to resume where we left off. It picked up with the same context, the same user preference, the same assistant and tool calls it had already listed. Not a re-run from the top. A continuation, because all of that state is still on disk.

The part almost nobody is talking about

There are two packages sitting inside this repo: subagent claude code and subagent codex.

These are not model providers. This is not an Anthropic key in a settings box. They spawn Claude Code and Codex as actual child processes and hand them a task. Your agent gets two new tools, and it can decide by itself that a job is better handled by Claude Code, then go and run Claude Code.

Three things will catch you out:

1. It is one shot. Every call starts a fresh process and a conversation that cannot be resumed. One task, one answer, no follow up.
2. Both providers load dormant. The preset has to decide to give your agent the tool. It is not switched on the first time you open it.
3. It strips credential shaped variables out of the environment on purpose. Whatever key is already in your shell does not reach the child. You have to pass it in explicitly.

This is why the framing everyone has used is wrong. Every take on this has been a versus: this versus Claude Code, this versus the agent you already pay for. It is not a competitor to that agent. It is a layer that can drive it.

So should you switch

The question is not DeepSeek or Claude. Add a provider, pick Anthropic, put your key in, and you are running Opus inside the free harness with the trace and the plugins.

The harness and the model were always two separate purchases. One of them just went to zero.

One real limitation on that: it is an API key, not your Claude subscription, and the Claude API is not famous for being cheap. Running Opus through here is not the same bill as running Opus through the thing you already subscribe to.

What you should know before you install it

This is a developer preview. The only tagged release is a release candidate, and the readme says in capital letters that there will be breaking changes. Things will go wrong for you that did not go wrong for me, and things went wrong for me that will not go wrong for you.

That is the trade. When this breaks, it can come back, because the state is traceable and the components are not hardcoded. That is the same property that makes everything else in this article possible.

There is a formal paper behind it and a framework called Cordis. I am not going to read a paper at you.

What to do next

Install it, point it at a free or cheap model for a week, and run it on the work that does not need a frontier brain. That is the honest test: not whether it can do your hardest task, but how much of your month it quietly covers.

Then look at what is left. If the answer is a small slice, you now know exactly what you are paying a frontier model for, and you are paying for the brain instead of the software wrapped around it.