I Built DeepSeek Harness Before Lunch. Here's the Six-Hour Window Nobody Mentions.
Plain-English guide · 8 minutes · one line to install.
DeepSeek Harness is DeepSeek’s open-source “agent harness”: the software around an AI model that reads your files, runs commands and remembers the conversation, with every part built as a swappable plug-in.
Who this is for: you use Claude Code (or want to), you don’t code, and your AI bill is starting to show up in your budget. Who should skip it: if you need an agent for client work today, stay on Claude Code; this is a developer preview.
After reading this you’ll know whether dsh belongs in your week, which model to plug in, and when to run it so it’s nearly free.
Prices last verified 2026-09-22 on Ollama Cloud. They will change; the ratios are the point.
So what is it?
Claude Code is the best tool I’ve used for getting real work done from my files. It is also the most expensive thing in my stack.
DeepSeek Harness does the same job. It’s free. And if you run it outside a six-hour window each weekday, the model behind it costs about a third of the daytime price. Most people running AI agents don’t know the window exists.
I built it from source one morning and had it in my browser before lunch.
Is it the same as DeepSeek?
No. This is the confusion that costs people money.
DeepSeek, the model
- The AI brain. DeepSeek V4 and friends.
- Made by DeepSeek, a Chinese lab.
- You pay per word it reads and writes.
- Runs on someone's server, not yours.
DeepSeek Harness, the tool
- The software that uses a brain.
- Free. Open source. Runs on your machine.
- Works with DeepSeek's models or anyone else's.
- The privacy decision is which brain you plug in, not the tool.
From the project README: “It is built on an everything-is-a-plugin architecture and powered by Cordis.” Every part is swappable: tools, memory, the web page, the model. That’s why a non-coder should care, and I’ll get to why.
One more thing they say plainly: it’s a “developer preview.” Things will break between versions. Run it for the work below, not for anything a client is waiting on.
What does it cost me?
The tool costs nothing. The model is where the money goes.
I run the models on Ollama Cloud, on the $20 plan, which gives you monthly credit, fast hardware, and a no-log, no-train hosting policy. You pick a model by name and paste one key.
Which model should I plug in?
Two honest options.
| Model | Price (in / out per M) | Pick it if | |
|---|---|---|---|
| Default | gpt-oss:120b — OpenAI’s open-weight model | $0.15 / $0.60 | You want zero questions about where your data goes. US lab, same speed, same hosting policy. Enough for files, formatting, drafts, checklists. |
| Power option | deepseek-v4-pro | $0.66 / $1.98 off-peak | You want the strongest open model for reasoning and coding, and you’ve made peace with it coming from a Chinese lab. |
I run the power option, off-peak, knowing that. For a client with a compliance person, I’d run the default. Neither matches Claude on long, messy, multi-file jobs. Which is the next section.
How do I install it?
You need Node.js (a one-time download). Then one line in your terminal:
npx @deepseek-ai/dsh webIt opens a page on your own machine at http://127.0.0.1:3080. That page is the whole thing: a chat that can see your files. Connect a model, and you’re working.
When do I use it instead of Claude Code?
Most people frame this as dsh versus Claude Code. Wrong fight. I run both. The rule for which gets a job is the one I learned the hard way with Jev:
Route by the hardest step in the job, not the average.
A job with five steps where four are easy and one threads a change through three files is a Claude Code job. The cheap model will do the four and quietly stop at the fifth. A job where every step is easy, or where 80% right is fine because you’ll rewrite it, is a dsh job.
My week, split:
- dsh, cheap model, off-peak: transcript clean-up, note reformatting, first-pass summaries of recorded calls, checklists, renaming and filing, the boring half of any draft.
- Claude Code: anything touching my website’s code, anything with more than three files, anything I’ll ship to a client, anything where “it stopped early” would cost me a day.
- Jev in front of both: the yes/no and which-one questions, before either spends a token.
The second thing nobody mentions: the harness becomes where you keep your own skills. Because every part is a plug-in, the prompts and checklists scattered through your notes can live inside dsh as named tools. Same move as in Claude Code for Non-Coders: a repeated prompt becomes a file you call by name. Do it in a tool you own and the file is portable.
Where does it go wrong?
I nearly framed it as a replacement. My first question was “can I use DeepSeek instead of Claude Code?” The research said the strongest open model still trails Claude on long, follow-the-instructions work. The right question was “which jobs?” Check: decide per job, not per tool.
Cheap models stop early and don’t tell you. I saw this with a cheaper Claude tier first: four of five steps, then “ready for integration.” Expect the same shape here. Check: give the cheap harness jobs where stopping early is visible and harmless.
“Developer preview” means it. The README says breaking changes are coming. Check: if a setting from a tutorial doesn’t work, look at the tutorial’s date before assuming you did it wrong.
Cheap is only cheap when the routing is right. A cheap run that stops early and gets re-run on Claude costs more than one Claude run. Check: the routing rule above.
Do this in 10 minutes
- 01Install it.
npx @deepseek-ai/dsh web. A page opens. Nothing is changed yet. - 02Plug in gpt-oss:120b. Ollama Cloud, $20 plan, one key. Start with the no-questions model; switch later if you want more power.
- 03Give it one boring job, after lunch. Reformat a folder of notes, or summarize one transcript. If it finishes, you have a cheap worker. If it stops early, you've learned the rule for free.
Take the rule with you:
Which harness gets this job?
1. What is the HARDEST step? (not the average)
2. Code, >3 files, or ships to a client → Claude Code.
3. Everything easy, or 80% is fine → dsh, cheap model.
4. Yes/no or pick-one first → Jev, before either.
Off-peak: avoid 12:00–18:00 UTC Mon–Fri = 5–11am Pacific.
Default model: gpt-oss:120b. Power option: deepseek-v4-pro.Send this to the friend who says Claude Code is too expensive. It is, for the boring half.
Questions people ask
Is DeepSeek Harness free? The tool, yes: open source on GitHub. The model you connect is what you pay for.
DeepSeek Harness vs Claude Code: which is better? For the easy 80% of work, close enough that price wins. For the hard 20%, Claude Code, and I keep it for exactly that.
Do I have to use DeepSeek’s models? No. It’s model-agnostic. I recommend gpt-oss:120b as the default.
Does it send my files somewhere? The tool runs on your machine. The model sees whatever you send it, wherever it’s hosted. The model choice is the privacy decision.
Can I run it on a server? Yes, the web page is built for that. If you want agents running while you sleep, that’s a different tool: Hermes Agent.
Where do I read more? The docs and the repo. Read their SAFETY.md before giving it access to anything that matters.
Sources and further reading
- DeepSeek Harness — GitHub repo · documentation. The “everything-is-a-plugin” line is theirs.
- Ollama Cloud — ollama.com: the hosting and the prices quoted above, checked 2026-09-22.
- Cordis — the plug-in framework underneath dsh; its design paper is linked from the README.
- My routing rule — from the Jev technical post, where I learned it the expensive way.
The tools I run, in plain English
- What Is Jev? — the $0.0002 yes-or-no AI that sits in front of everything.
- What Is Hermes Agent? — the self-hosted team lead that runs agents while you sleep.
- Jev in front of Claude Code — the technical build log with the numbers.
- Claude Code for Non-Coders — start here if none of this makes sense yet.