DeepSeek Harness: Is the Fastest-Growing Repo Worth It?
DeepSeek Harness: Is the Fastest-Growing Repo Worth It?
DeepSeek open-sourced its own coding harness, and it just became the fastest-growing GitHub repo ever — 167,000 stars in under a week. Think Claude Code, but for DeepSeek: a full agentic coding environment where everything is a plugin, including the harness itself. I installed it, built a custom plugin from scratch, and stress-tested the pitch. Here's exactly how it works, where it beats Claude Code, and whether it's actually worth switching.
What Is the DeepSeek Harness?
The DeepSeek Harness is an open-source agentic coding tool — DeepSeek's answer to Claude Code and Codex. You give it a prompt, it plans, calls tools, edits files, and runs commands inside a workspace. On the surface it looks like every other harness: workspaces (your folders) on the left, chat windows underneath, a prompt box up front, slash commands, permission levels, and a model/effort picker.
The thing that makes it different is that everything is a plugin — and that includes the plumbing of the harness itself. In Claude Code or Codex, plugins usually mean skills, MCPs, or CLIs — outside tools you bolt onto a fixed core. The DeepSeek Harness lets you edit and change the core itself. That's the whole pitch: an infinitely customizable harness that can, in theory, write its own plugins and partially improve itself.
It racked up 167,000 stars in a handful of days, which is why it's blowing up across dev Twitter and YouTube right now. But star count isn't the same as "you should use this." Let's get into how it actually works.
How Do You Install the DeepSeek Harness?
Search "DeepSeek Harness" on GitHub, find the repo, and run the single install command in your terminal. It spins up a web UI — no complicated setup. The layout follows the conventions you already know from other harnesses, so there's almost no learning curve to get oriented.
One detail worth knowing upfront: it's open source, so you're not locked into DeepSeek's own model. In my setup I'm not using the DeepSeek API at all — I'm running it through OpenRouter, which gives me access to basically every model out there. To wire that up, go to Settings, then Models, add a provider, and paste in your API key. You can also make the whole thing run completely local with whatever model you want.
What Are the Four Modes in the DeepSeek Harness?
Before you prompt, you pick a mode. There are four, and they change how the agent behaves:
- Standard mode — exactly what it sounds like. This is the default, closest to working inside Claude Code. Tool calls happen one at a time.
- PTC mode — built for multi-step tasks. Instead of firing tool calls one by one, it writes a script that runs all the tools essentially in one go. For complex tasks, this cuts the context bloat you get from calling tools individually.
- Minimal mode — for very simple tasks where you don't need the overhead.
- Creator mode — where you build your own plugins for the harness.
You also get slash commands (compact, goal, and more), permission levels for read-only, workspace, write, and full access, plus a model and effort selector. If you've used Codex, one setting will feel familiar: enter behavior while busy — set it to "queue" to wait until the current tool call finishes before your next message lands, or "steer" to inject your message as soon as the current tool call wraps.
What Does "Everything Is a Plugin" Actually Mean?
Open the plugin list and you'll see essentially every single thing the harness can do, defined as a plugin. Then there's the plugin configuration, which exposes parts of the harness you can actually edit. Look at something like the agent loop — that's core plumbing, and you can start tuning it, right down to how parallel tool calls behave. Go deeper with Creator mode and you can rewrite far more.
This is the real line between the DeepSeek Harness and Claude Code. In Claude Code you can approximate a lot of these changes with skills and hooks, but you're not touching the actual plumbing under the hood. Here, you are. That's the trade the whole tool is built around.
The engine everything runs on is a program called Cordis — you'll see the name all over the harness. Cordis is the kernel, the chassis of the car; the plugins are the engine and everything else you bolt on.
Is It Safe to Install Random DeepSeek Harness Plugins?
Here's the part you need to pay attention to, because it's the biggest red flag. As of right now — and remember, this is a developer preview — every plugin you add gets full shell access and full access to your entire file system.
That matters. There's a huge, exploding library of community plugins on GitHub, and installing one is often as easy as running a single command. But if you install a sketchy plugin from a bad actor, it can read your API keys and do real damage, because it has permission to do exactly that by default.
When I needed web search — which requires a DeepSeek API key out of the box, so it doesn't work if you're running another model — I installed a free-search plugin that lets me use Bing or DuckDuckGo instead. Before running it, I had Claude Code review the plugin's code to make sure it wasn't doing anything shady. That's the move. This whole situation is reminiscent of the early days when tools first hit the scene with wide-open permissions: be wary, know what you're installing, and check the code. Hopefully default permissions get tightened as it matures.
How Do You Install and Create Plugins?
For installing, the best resource right now is the "awesome DeepSeek Harness plugins" page on GitHub — a big curated list. Say you want a UI enhancement, like an effort-level slider. You find it on the list, follow the install instructions (usually one terminal command), restart the harness — about two seconds — and the slider shows up.
A quick myth to correct: you'll see people claim there's a built-in hot-reload system so you never have to restart. That's only sort of true — it depends on the plugin. Some plugins do require a quick restart. It's fast, but don't expect zero restarts.
Creating your own plugin is the real selling point. Drop into Creator mode and describe what you want in plain language. As a simple test, I asked it to build a plugin that changes the UI color scheme to look like the Matrix, with falling ASCII characters in the background. It went off and built it. The result was ugly as sin — but that's the point: it shows how easy it is to spin up a custom plugin that changes the harness itself.
One gotcha: custom plugins aren't saved automatically. When it builds one, it's not permanent. You have to explicitly tell it to save the plugin — or save it with a toggle — or you'll lose it at the end of the session.
Why Is the Trajectory View a Big Deal?
The second real selling point is visibility, and this is where the DeepSeek Harness genuinely beats Claude Code. At the bottom of the screen you get info most harnesses hide: cache hit percentage, input and output tokens, and tokens per second, live.
It goes deeper than that. You can see the context injection and the actual system prompt itself. And there's a feature called the trajectory — deep insight into everything happening in a turn: the user input, the context, the assistant, the tools. Click into any of them and you get a summary, a preview, and the raw source. This is far more visibility into what the AI is actually doing than a harness like Claude Code gives you. When you're troubleshooting and need to know where things actually went wrong, this is where you find it.
You can also download the entire session log as a zip — full JSON, broken down by turns and calls. For debugging agentic runs, that level of transparency is rare.
Should You Switch From Claude Code to the DeepSeek Harness?
The short answer: probably not — but you should still try it.
Don't get me wrong, the DeepSeek Harness is genuinely impressive. I'm glad they didn't just clone Claude Code. The "everything is a plugin" approach and the deep customizability are legitimately cool, and the trajectory view is best-in-class. The dream — a self-improving harness that writes its own plugins and reshapes itself around your exact task — sounds incredible.
But in reality, that dream isn't here yet, and it might never fully arrive. There's a real question buried in the hype: is a fully custom harness actually moving the needle enough to justify the time and energy to build it? For most people, most of the time, the answer today is no. It's tough to argue it's a clear step above Claude Code right now.
Here's why you should still install it anyway: it's free, open source, and fully local. You're not stuck with one tool — you can use both. You should be tool-agnostic. Tinker with it, see how the harness internals work, and pull the ideas worth pulling. Just don't rip out your Claude Code workflow expecting a night-and-day upgrade.
Frequently Asked Questions
What is the DeepSeek Harness?
The DeepSeek Harness is an open-source agentic coding tool from DeepSeek — its equivalent of Claude Code or Codex. Its defining feature is that everything, including the harness's own core plumbing, is a plugin you can edit. It hit 167,000 GitHub stars in under a week, making it the fastest-growing repo ever.
Do you have to use DeepSeek's model with it?
No. Because it's open source, you can run it with any model. In my setup I use OpenRouter instead of the DeepSeek API, which opens up nearly every model available. You can also run the whole thing completely locally.
Are DeepSeek Harness plugins safe to install?
Be careful. In the current developer preview, every plugin gets full shell access and full access to your file system by default. A malicious plugin could read your API keys. Only install from trusted sources, and have a tool like Claude Code review the plugin's code before you run it.
What is Cordis in the DeepSeek Harness?
Cordis is the kernel — the core engine the entire harness runs on. If Cordis is the chassis of the car, the plugins are the engine and the add-ons you bolt onto it. You'll see the name referenced throughout the harness.
Should you switch from Claude Code to the DeepSeek Harness?
For most people right now, no — it's hard to call it a clear step above Claude Code today. But since it's free, open source, and fully local, it's worth installing and experimenting with. You can run both and stay tool-agnostic rather than committing to one.
If you want to go deeper into building with agentic coding tools, join the free Chase AI community for templates, prompts, and live breakdowns. And if you're serious about building with AI, check out the paid community, Chase AI+, for hands-on guidance on how to make money with AI.
