DeepSeek Harness Review: Fast, Cheap, But Is It Good?

Item: DeepSeek Harness
Rating: 3.5
Author: Julian Goldie

This DeepSeek Harness review comes down to one score: I give it a 7 out of 10.

It is the fastest and cheapest AI coding agent I have put up against Claude Code, and it is the easiest agent harness I have handed to a beginner.

But when I put its builds next to Claude's, Claude's still looked more like something a real business would ship.

I tested it twice on camera.

The first test was a straight race against Claude Code with my mate Kasra Dash, using the exact same prompt.

The second test was a workflow comparison against Hermes, which is the open-source agent I use every day.

Everything in this review comes from those two tests, so you are getting what I actually saw rather than the launch hype.

The Verdict: DeepSeek Harness Scores 7 Out Of 10

Here is my verdict before the detail.

DeepSeek Harness is worth trying if you want a fast, cheap, beginner-friendly coding agent that you can reshape however you like.

It is not a Claude Code killer yet, because Claude still has a big head start on the quality of what it builds.

It is a v0.1 developer preview, which means it is roughly a tenth of what it could become, and it will break from time to time.

My co-host Kasra Dash was harsher than me, and he said it did not blow his socks off.

I scored it higher because I value what it does to the market, since serious competition pushes every tool to get better and cheaper.

DeepSeek Harness Review Scorecard

I did not invent sub-scores out of thin air, so this scorecard shows what each test actually showed.

What I tested	Result	How DeepSeek Harness did
Speed against Claude Code	It finished in 11 minutes while Claude was still running at 30.	This is a clear win for DeepSeek Harness.
Cost per build	The test build cost about 5 cents.	This is a clear win, because DeepSeek is roughly 56 to 57 times cheaper per token.
Output quality	Claude's build looked more professional and the game felt smoother.	Claude Code still wins on polish.
Token efficiency	It used 483,000 tokens on the one-page site.	It is very verbose, which the low price hides.
Beginner friendliness	It broke less and felt easier to set up than Hermes.	This is a clear win over Hermes.
Coding against Hermes	It replied faster and produced better builds.	This is a clear win over Hermes.
Memory and learning	It has no equivalent of Hermes's self-improvement loop out of the box.	Hermes still wins here.
Stability	It is a v0.1 developer preview.	Expect breakage and fast changes.
Overall	My rating is 7 out of 10.	It is worth trying, but it is not a full replacement for Claude Code.

What DeepSeek Harness Actually Is

DeepSeek Harness is a local AI agent that DeepSeek released as a developer preview at the same time as its DeepSeek V4 Pro model.

The agent runs on your computer, but the model it uses does not.

It can read your files, run commands, change code and search the web.

Its official slogan is "Everything is a plugin," and that is not marketing fluff.

The model, the tools, the memory, the session locks, search, sub-agents, scheduling and even the main loop are all plugins you can swap without touching the core.

I think of it like a box of Lego compared to a finished toy.

The plugin framework underneath it is called Cordis, which has powered an open-source chatbot project for years, so the foundations feel solid even though the harness is brand new.

It runs locally as a website in your browser, so there is no extra desktop app to keep clicking into.

The community jumped on it fast, because it passed 100,000 GitHub stars within about two days and the community had published 421 plugins one day after launch.

Test 1: DeepSeek Harness vs Claude Code

This was the test I cared about most, because Claude Code is the tool most people would be switching from.

Kasra and I gave both tools the exact same prompt.

We asked for a 3D animated accountancy website plus a Tetris game, which is a mix of a business page and a working piece of interactive code.

We kept it like for like, with DeepSeek V4 Pro on the harness and Claude Opus 5 on high inside Claude Code.

Speed went to DeepSeek Harness

DeepSeek Harness finished the whole build, start to finish, in 11 minutes.

Claude Code was still going at the 30-minute mark.

That is not a small gap, and if you are iterating on lots of small builds, that speed adds up fast.

Cost went to DeepSeek Harness by a mile

Kasra topped up $10 of credit, and the full build cost about 5 cents.

DeepSeek works out at roughly 56 to 57 times cheaper per token than Claude.

There is a catch hiding in that number, though.

DeepSeek used 483,000 tokens on the one-page site, while Claude was at 48,000 tokens at the 20-minute mark.

So DeepSeek is a very verbose model, and it is only the tiny price per token that makes the bill look so small.

Output quality went to Claude Code

This is where the race flipped.

Claude's build looked like a professional website.

The animation followed the mouse, it picked up local context and mentioned the North West of England, and the Tetris game felt smoother and nicer to play.

DeepSeek's build was animated but cartoony.

Kasra put it well when he said a cartoon game on an accountancy website is a mismatch.

Neither build was ready to ship, and both would have needed more back and forth, more context and better prompting.

The ratings from that test

Kasra rated Claude Code 9 out of 10, and he said it changed everything for him, especially paired with Fable 5.

I rated DeepSeek Harness 7 out of 10.

Both of us are staying on Claude for day-to-day work, but we would both move certain jobs over to DeepSeek.

If you want to see how Opus 5 stacks up against the other flagship models on 50 builds, read my Claude Opus 5 vs GPT-5.6 verdict.

🔥 Want my DeepSeek Harness and Claude Code setup? Inside the AI Profit Boardroom, I show how I split work between Claude Code and cheaper agents like DeepSeek Harness, with step-by-step video tutorials, weekly coaching calls and 3,400+ members testing these tools as they launch. → Get access here

Test 2: DeepSeek Harness vs Hermes

The second test answered a different question.

I wanted to know whether DeepSeek Harness could replace Hermes, the open-source agent framework from Nous Research that I already run every day.

Beginner friendliness went to DeepSeek Harness

Hermes is powerful, but it is fiddly and it breaks more often.

DeepSeek Harness breaks less, it is easier to set the model, and it feels a lot like using Claude Code.

If a beginner asked me which one to start with, I would point them at DeepSeek Harness.

Customisation went to DeepSeek Harness

This was the moment that impressed me most in the whole review.

The harness has a creator mode where a new session can rebuild the interface itself.

On camera, I added a three-tasks panel, a mission control view and a daily task scheduler just by asking for them.

In Hermes, building my own custom workflows took hours of coding, so this felt like a huge shortcut.

Coding output went to DeepSeek Harness

For coding and building, DeepSeek Harness produced better results and replied much faster.

Hermes tends to time out on big coding tasks, because it is really built for scheduled and smaller jobs rather than heavy coding.

DeepSeek V4 Pro was built for the harness, and it shows.

Memory and messaging went to Hermes

Hermes still has two things that DeepSeek Harness does not match out of the box.

The first is learning, because Hermes has a three-layer memory and a self-improvement loop that saves what it learned as a skill and gets faster next time.

The second is reach, because Hermes talks to you through Telegram, WhatsApp, Discord, Microsoft Teams and even iMessage.

There is already a community plugin that pairs DeepSeek Harness with a phone, but Hermes does it natively.

DeepSeek Harness vs Claude Code vs Hermes Compared

Feature	DeepSeek Harness	Claude Code	Hermes
Speed on my test build	It finished in 11 minutes.	It was still running at 30 minutes.	It times out on big coding tasks.
Cost	It cost about 5 cents with DeepSeek V4 Pro.	It is roughly 56 to 57 times more per token.	It depends on the model you plug in.
Output polish	It was cartoony in my test.	It was the most professional build.	It is not really built for heavy coding.
Beginner friendliness	It is the easiest of the three for beginners.	It is easy but tied to Claude.	It is powerful but fiddly.
Customisation	Everything is a plugin, and creator mode rebuilds the UI.	It has a fixed product design.	It is customisable with more coding work.
Self-improving memory	It has no built-in learning loop yet.	It has no built-in learning loop either.	It has a three-layer memory and a self-improvement loop.
Messaging apps	A community plugin adds phone access.	It is not built for messaging apps.	It supports Telegram, WhatsApp, Discord, Teams and iMessage.
Maturity	It is a v0.1 developer preview.	It is a mature product.	It has been out since February 2026.

DeepSeek Harness Pros And Cons

What I liked:

It finished my test build in 11 minutes, while Claude Code was still running at the 30-minute mark.
The build cost about 5 cents, which makes experimenting almost free.
The plugin design means you can swap the model, memory, tools and even the main loop.
Creator mode let me add a mission control and a scheduler without writing code.
It is the easiest agent harness I would hand to a beginner right now.
Session logs are easy to download when you want to review what it did.

What I did not like:

Its output looked less professional than Claude's on the same prompt.
It is very verbose, burning 483,000 tokens on a one-page site.
It is a v0.1 developer preview, so it will change fast and break at times.
It does not have Hermes's self-improving memory out of the box.

Is DeepSeek Harness Free?

The harness itself is free to run.

But a harness is just the body, and it needs a brain plugged in to do anything.

You can pay for DeepSeek V4 Pro, which cost about 5 cents for my full test build.

Or you can plug in a free brain, such as DeepSeek V4 Flash through OpenCode.

That means the tool and the model can both cost you nothing, which is a big shift from paying for a premium coding subscription.

You can see where DeepSeek V4 Pro sits on my GoldieBench DeepSeek V4 Pro page, and compare it with the scored models in my ranking of the best AI models for coding.

Who Should Use DeepSeek Harness?

You should try it if you are a beginner who wants an agent that feels like Claude Code without the bill.

You should try it if you run lots of small builds and speed and cost matter more than polish.

You should try it if you like tinkering, because the plugin design rewards people who want to reshape their tools.

You should skip it for now if you need client-ready output on the first pass, because Claude Code still wins there.

You should skip it for now if you need your agent to learn over time or live inside your messaging apps, because Hermes does that better today.

How I Actually Use It

I do not think the right answer is choosing one agent and throwing the others away.

I run DeepSeek Harness and Hermes side by side and give each one the lane it is best at.

DeepSeek Harness gets the coding and building jobs.

Hermes gets the scheduled tasks, the memory-heavy work and anything I want to trigger from my phone.

Claude Code still handles the work where output quality matters most.

I manage all of them from one place, which I call my Agentic Operating System, so an orchestrator sets up the job and delegates it to DeepSeek Harness without me touching the harness directly.

That way, it does not matter which harness launches next, because the system just gets another worker.

🚀 Want to build the same multi-agent setup? Inside the AI Profit Boardroom, I walk through how I run DeepSeek Harness, Hermes and Claude Code together, with video tutorials, weekly coaching calls and 3,400+ members sharing the setups that work for them. → Join the Boardroom here

DeepSeek Harness Review FAQ

Is DeepSeek Harness any good?

Yes, with limits.

In my test it finished a website and game build in 11 minutes for about 5 cents while Claude Code was still running at 30 minutes.

Claude Code's output looked more professional, so I rate DeepSeek Harness 7 out of 10.

Is DeepSeek Harness better than Claude Code?

It is faster and far cheaper, but it is not better on output quality.

DeepSeek is roughly 56 to 57 times cheaper per token than Claude, yet Claude Code produced the more polished build.

Kasra Dash rated Claude Code 9 out of 10 against my 7 out of 10 for DeepSeek Harness.

Is DeepSeek Harness better than Hermes?

For coding and beginners, yes.

It replied faster, produced better build results and broke less often than Hermes in my test.

Hermes is still better for self-improving memory and for talking to your agent through apps like Telegram and WhatsApp.

Is DeepSeek Harness free?

The harness itself is free to run, but it needs a model plugged in as its brain.

You can pay for DeepSeek V4 Pro, which cost about 5 cents for my test build, or plug in a free brain such as DeepSeek V4 Flash through OpenCode.

Is DeepSeek Harness stable enough to use?

It is a v0.1 developer preview, so expect things to change quickly and occasionally break.

I would use it for real coding tasks today, but I would not build a whole business process on it without a fallback.

Final Verdict

DeepSeek Harness is fast, cheap, flexible and friendly to beginners.

It beat Claude Code on speed and cost, and it beat Hermes on coding and ease of use.

It lost to Claude Code on polish and to Hermes on memory and messaging.

For a v0.1 preview, that is a seriously strong start, and it is why I am happy to give it a 7 out of 10.

That is my honest DeepSeek Harness review.

Also From Julian

I publish more agent comparisons on the AI Profit Boardroom blog.
My full agent setup walkthroughs live on AgentOS.guide.
I track how the underlying models score on the GoldieBench AI leaderboard.

About Julian

I'm Julian Goldie, an AI entrepreneur, SEO expert, and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation, and SEO.

I have 400,000+ YouTube subscribers who watch my AI tool tests every week.
I built a 7-figure agency from the ground up.
I run daily AI training inside the Boardroom.
I wrote two Amazon best-sellers on SEO and agency growth.

→ Get my best AI training inside the AI Profit Boardroom