DeepSeek Harness vs Claude Code: Is the Free Rival Ready?

DeepSeek Harness vs Claude Code is the comparison developers are searching for since DeepSeek put its own coding agent on GitHub. DeepSeek Harness (you run it with the command dsh) is free, open-source software under the MIT license, and DeepSeek presented it as an open alternative to Anthropic's Claude Code. So what does DeepSeek Harness do, what does it really cost, and should you switch?

The software costs nothing, the model behind it does, and nobody has yet published a measured head-to-head between the two tools. What exists are vendor benchmark scores for the underlying models, a pricing change that lands today, and a neat trick that lets Claude Code itself run on DeepSeek's engine.

This post covers what DeepSeek shipped, what it will cost you, and how to run a fair test on your own code before you trust anyone's leaderboard.

What is DeepSeek Harness?

DeepSeek Harness is an agent harness: the layer of software that sits around a language model and gives it hands. The model on its own can only read and write text. The harness hands it a codebase to open, a terminal to run commands in and external tools to call, then loops: the model proposes an action, the harness carries it out, and the result goes back to the model for the next step. That is the same job Claude Code does for Anthropic's models.

According to the deepseek-harness repository on GitHub, the project uses an "everything-is-a-plugin" architecture, so tools, commands and behaviors are add-ons you can swap. It is labeled a developer preview, and the README warns that updates may break compatibility. When we checked on November 1, 2026, the repo showed 267.7k stars, 32.5k forks and 24,922 commits, which is a lot of attention for a tool that is less than three months old.

The timeline is short. DeepSeek reportedly formed a team to build Harness in May 2026. On August 12, Bloomberg reported that DeepSeek was preparing to compete with Claude Code, and the company set up a "DeepSeek Harness Team" WeChat account, according to PYMNTS. The next day DeepSeek announced Harness v0.1 alongside its V4 Pro model, as Forklog reported.

What is Cordis in DeepSeek Harness?

Cordis is the engine underneath. The README says Harness is powered by Cordis, which is described in a paper titled "A Programming Paradigm for Spatiotemporal Composability" (arXiv 2608.25512). The repo does not say much more than that, so if you want the theory, the paper is the place to read it.

How to install DeepSeek Harness and the desktop app

You need Node.js. One command starts it: npx @deepseek-ai/dsh web opens a web interface at http://127.0.0.1:3080 in your browser. Add --no-open if you would rather open the page yourself. To build from source, clone the repo, then run pnpm install, pnpm run build and pnpm dsh web.

If you prefer an app window over a browser tab, DeepSeek Harness desktop builds shipped on October 2, 2026 for macOS on Apple Silicon and for Windows, per yuv.ai's write-up. The sources list no other desktop platforms; the npx route above is the documented way to run it elsewhere.

As for whether DeepSeek Harness is safe to run, the same caution applies as to every coding agent: it can run commands on your machine. Start it in a throwaway project, read what it asks to execute, and check any plugin before you install it. Our guide on the Claude Code extension backdoor walks through why add-ons for coding tools deserve that scrutiny.

Is DeepSeek Harness free?

Yes, the harness is free: MIT-licensed open-source code you can download, modify and run. What you pay for is the model it talks to, billed per token through DeepSeek's API.

The yuv.ai article lists peak-hour pricing for the deepseek-flash model at $0.30 per million uncached input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens, with off-peak hours at half price. That model is tied to DeepSeek-V4.1-Flash-1017, which the article describes as 763 billion total parameters with a 1 million token context window and open weights under MIT.

The peak and off-peak split itself is newer than many users realize. It took effect on August 16, 2026, according to Forklog, and one X user called it "basically a 2x price increase." That is one person's reading, not a DeepSeek figure, but it is a reminder to check what hours your team actually works.

The bigger change starts today. Search and Fetch calls (the tools an agent uses to look things up on the web) have been free through the DeepSeek API whether you called them from DSH, Claude Code or Codex. From November 1, 2026 at midnight Beijing time, they are billed at $0.75 per 1,000 successful requests. An agent that searches often during a long task can rack those up quickly, so watch your first invoice.

Can Claude Code run on DeepSeek's engine?

It can, and this is what most people looking for a DeepSeek Harness Claude Code plugin actually want. You do not need a plugin at all: DeepSeek exposes an Anthropic-compatible endpoint, so Claude Code keeps its familiar interface while DeepSeek's model does the thinking.

Per yuv.ai, you set ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic and point the models at deepseek-flash. There is a catch. Requests for "claude-opus" models are mapped to the pricier deepseek-v4-pro unless you also set ANTHROPIC_DEFAULT_OPUS_MODEL. The snippet below shows only the settings the sources name; DeepSeek's and Anthropic's docs list the rest, including authentication. If you forget the override, any step that asks for the heavyweight model quietly bills at the higher rate.

export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-flash"
claude

This setup is useful for testing, because it separates two questions that usually get tangled. Swapping the model inside Claude Code tells you how much the model matters. Running DSH on the same model tells you how much the harness matters.

DeepSeek Harness vs Claude Code benchmark: what the numbers say

No one has published a benchmark that measures both tools on the same tasks. Neither Forklog nor yuv.ai reports one. The scores circulating online describe models, and they come from the companies selling them, so they cannot settle whether DeepSeek is better than Claude for coding.

For V4 Pro, DeepSeek reports 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 74.1 on Toolathlon-Verified, according to Forklog. V4 Pro also launched with native Responses API support, Codex integration and three reasoning levels (low, high and max). For the Flash model, yuv.ai cites vendor-published Terminal-Bench 4.0 results rising from 31.2 to 42.3, against 51.8 for Anthropic's Opus 5.

Read those numbers with care. They are self-reported, they use different benchmark versions, and a model score says little about how a particular harness plans work, recovers from errors or asks before deleting files. On the figures we have, Flash trails Opus 5 on that one test. Whether the gap matters depends on what you pay and what you build, which only your own test can show.

DeepSeek Harness vs Claude Code vs OpenCode: how to run a fair test

The same gap applies if you add OpenCode or Codex to the comparison: we found no sourced test of DeepSeek Harness vs Claude Code vs Codex, or of any three-way lineup. You can still run a useful one in an afternoon, because Claude Code and Codex can both talk to DeepSeek's API.

A fair test keeps everything constant except the thing you are measuring:

• Same model. Point every harness at deepseek-flash first, so differences come from the harness. Then rerun Claude Code on Anthropic's own model to see what the model adds.
• Same tasks. Pick five to ten real tickets from your own repo: a bug fix, a small feature, a refactor, a failing test. Public benchmarks may already sit in a model's training data.
• Same starting point. Reset the repo to the same commit before each run, with identical instructions.
• Score what matters to you. Did the tests pass? How many times did you have to step in? How many tokens and Search or Fetch calls did it use, now that those are billed?
• Note the rough edges. DSH is a developer preview, so log crashes and breaking updates separately from bad answers.

If you want a broader sense of where these tools are heading, our explainer on computer-use AI agents covers the next step: agents that operate whole desktops.

What to watch next

The DeepSeek Harness vs Claude Code question will look different in a few months. Watch for an independent benchmark that runs both harnesses on the same model and tasks, since that is the missing piece today. Watch the first bills under the new Search and Fetch pricing, which will show whether DeepSeek's low token prices hold up for agents that browse a lot.

Watch, too, how quickly DSH leaves developer preview. Breaking changes are acceptable in a side project and costly in a team workflow, and a stable release is the signal that it is ready for the second kind of use.