Ten days working with DeepSeek Harness

I've spent ten days with DeepSeek Harness inside my real work, not a toy project, comparing it with Claude Code on the same tasks. This is the diary of those ten days: what a harness is, what this preview version does well, where it gets stuck and whether it's worth trying. Everything I describe is an experimental test, and that's how it should be read.

When it came out, the noise was the usual: "replaces Claude Code", "your free programmer", "no more paying for coding". If you believed the headlines, by now every programmer would be looking for work as a gardener.

I tried it for a more boring reason: I already work with Claude and with DeepSeek, and it was time to see whether this piece fitted my setup or not. I chose tasks I do every week: updating a VPS configuration, fixing small bugs on one of my blogs, writing a script to sort backups and documenting a project I'd abandoned.

A clear warning before going on: DeepSeek Harness is a preview release and this is an experimental test. What I describe holds for the version I used; later ones may change a lot, for better or worse.

First, the word. A harness is the program that wraps the model and gives it hands: tools, file access, permissions and a loop of asking, acting and checking the result. On its own, the model only produces text; the harness turns that text into actions. It's usually summed up like this: model plus harness equals agent.

DeepSeek Harness is DeepSeek's open-source coding agent. It was presented in 2026 as a developer preview, under the MIT licence (which lets you use and modify it freely), alongside the DeepSeek-V4-Pro model. What sets it apart is that everything in it is a swappable plug-in: the model, the tools, the sessions, the interface. It can even use Claude Code or Codex as helpers, and it comes with a local web app as well as the command line (checked in September 2026).

And DeepSeek, the company, in one line: a Chinese AI company, based in Hangzhou, offering chat, an API and open models you can download.

I installed it on the test server on my internal network, an old recycled machine running Ubuntu that I use as a sandbox. Why there and not on my work computer? Because an agent touches files and runs commands, and a new experiment doesn't go on the machine where I keep client projects. You test first where breaking things doesn't hurt.

I connected it to DeepSeek-V4-Pro through the DeepSeek API, in the cloud. I also tried a smaller open model locally, with Ollama, but it fell short for agent work and I shelved it on the second day.

What failed: the first install clashed with the machine's Python version; then the local web app wouldn't start because its default port was already taken by another service. Nothing serious, but nothing like "five minutes" either. It took me about four hours spread over two afternoons to get it working the way I wanted.

I was cautious with permissions: I gave it access to a single working folder, with a system user without administrator privileges and without access to my keys for other servers. And everything it touched was under version control (a history that lets you go back). I don't even trust myself, so imagine a newly arrived agent.

First task: the backup script. I asked for a script that sorted backups by date, deleted those older than thirty days and warned me if one was missing. It wrote it, tested it on a sample folder and fixed a date bug on its own. Good result first time, with just one correction from me about where to store the log.

Second task: small bugs on a blog. Three layout bugs in a template. It fixed two properly and "fixed" the third by changing something I hadn't asked for. I reverted that change and explained; second time round it got it right.

Third task: documenting an abandoned project. This is where it surprised me in a good way. It read the whole project, which even I didn't remember, and left me a clear document of what each part did and what was half-finished. It's the task where it saved me the most time.

Numbers for those three days: about six hours of work between asking, reviewing and correcting, around €3 of API usage and seven corrections from me in total.

On day one I pictured myself sitting with a coffee watching the agent work on its own. On day five I still had the coffee, but I was going over every change with a magnifying glass. Reality has a habit of letting your coffee go cold.

Here came the crash, and I'll tell it straight.

Updating the VPS configuration was a half-disaster. It had to change the web server configuration and renew the way certificates were managed. It got into a loop: apply a change, the check fails, undo, try again with a variant… Twenty minutes going round in circles until I stopped it.

I tried to rescue it by breaking the task into smaller steps and giving it the server documentation. It managed half; I finished the other half by hand. I also tried having it use Claude Code as a helper for that part, which is one of its selling points, and it worked, but then the question was obvious: what did I need the middleman for?

On day 7 there was an update to the tool that improved the local web app, but it changed nothing about the loop problem.

Another smaller snag: in a long session it started forgetting instructions I'd given it at the beginning, something I'd already lived through with chats. Nothing that doesn't happen to any agent, but it happened.

On the last two days I repeated equivalent tasks with both, on the same machine and from the same starting point.

What each does better: I liked DeepSeek Harness for reading and documentation tasks, and its local web app is handy for seeing what it's doing. Claude Code was more reliable on multi-step tasks where each step depends on the last, and gets less tangled when something fails.

Now the honest part: I'm comparing against the tool I know best. I've been with Claude Code for a long time, I know how to ask it for things and I know its quirks. I've had DeepSeek Harness for ten days. That skews the comparison, and I'm saying so for you to bear in mind.

If you use the DeepSeek API, what you send goes to its servers, which are in China according to its own privacy policy (checked in September 2026). With an open model running locally things change, because nothing leaves your machine, though the good models need a powerful computer.

That's why I gave it no client data or access to their projects: every task was on my own stuff. Not out of any special distrust, but by the same rule I apply to any new service.

And a warning about the opinions on DeepSeek doing the rounds: they mix up the chat, the API, the open models and now the harness, which are different things. Being wary of where your data goes in the app is not the same as assessing an open-source tool you can connect to a model on your own machine.

Am I still using it? For documenting projects and for simple tasks on the test machine, yes. For client work and long, delicate tasks, I'm sticking with Claude Code. I'd recommend it to someone who already knows their way around agents and fancies tinkering with an open, highly configurable tool. To beginners, no: learn first with a more mature tool.

What's true in the promise? It's a real coding agent, open and with good ideas. What's smoke? The "replaces Claude Code" and "free programmer" bits: the tool is free, but the cloud model is pay-per-use, and nobody gives away the review time. The ten days cost me about twenty hours and about €8 of API.

Maintenance note: since this is a preview release, I'll revisit this article when the tool changes and mark what's been updated and what has stopped working.