DeepSeek Harness: setup, models and our Windows test
Independent guide / developer preview. DeepSeek Harness (dsh) is DeepSeek’s local coding-agent runtime. It connects a model to tools that can inspect a project, edit files and run commands; it is neither the DeepSeek web chat nor a model you download. The official project currently calls it a developer preview and warns that interfaces may change.
Start here: in a disposable project folder, run the official npx @deepseek-ai/dsh web command, then choose a model and workspace. Our Windows test was partial: on September 23, 2026, we installed the official npm package and opened its Web UI. In a separate coding task, the agent read our files and reasoned about a repair, but its Windows sandbox failed before it could write the file. The guide separates documented setup steps from that observed result; we did not obtain a working agent-made fix.
By China AI Chat Team. Product sources and the test reviewed September 23, 2026 (Africa/Cairo). Research and review were AI-assisted; see our methodology. ChinaAI is independent of DeepSeek.
First, is Harness the right DeepSeek product?
- Want to ask questions or write in a browser? Use the official DeepSeek Chat. Harness is not required.
- Want to call a hosted model from your own app? Start with our DeepSeek API guide. Harness uses a model provider, but the API is a separate service with its own billing.
- Want a downloadable model on your computer? Our Windows/Ollama guide covers a specific R1 distill. Installing Harness does not install those model weights.
- Want an agent to work through a codebase? Harness is the relevant route. Start with a disposable project, limited access and a task you can verify.
How to start DeepSeek Harness
The sequence below follows DeepSeek’s official README and Web UI guide. We confirmed the package and UI launch on one Windows computer; we did not verify that every later step works on every system.
1. Prepare a small workspace and start the Web UI
Install a Node.js version supported by the current package and check the version your terminal actually runs. The official source repository currently declares ^22.19.0 || >=24.0.0; our 0.1.5-rc.3 test used Node.js 24.19.0. A declared range is not a guarantee that every release works on every version. In a folder containing only the project you intend to test, run:
node --version
npx @deepseek-ai/dsh webThe official command normally opens a local Web UI at 127.0.0.1:3080. Use the address printed by your own process if it differs, keep that terminal running while you use Harness, and keep any access token in its URL private. For a reproducible experiment, record the exact npm version you ran; our package was @deepseek-ai/dsh 0.1.5-rc.3, not a promise about the version a future unpinned command will fetch.
2. Connect a model
In the Web UI, open Settings → Models. The DeepSeek card accepts a DeepSeek Platform API key; other providers and custom endpoints have their own setup. DeepSeek’s model-configuration guide says the managed key is stored in $DSH_HOME/.credentials.yaml and is shown only as a redacted descriptor after saving. Create and store your own key through the official platform. Do not put it in a project file, screenshot, website page or browser script. A successful save does not prove that a model request will work: check the chosen model and provider with a small task.
3. Choose a workspace and a modest first task
Select Choose workspace and point it at the disposable project. The official guide says a new Web UI needs a selected workspace before its composer is available. Check the access mode before sending a task. Our UI displayed Workspace Write; that setting did not make our Windows file operation succeed, and a workspace setting alone should not be treated as a guarantee that all reads, processes or network activity are isolated.
Begin with a read-only request such as Summarize this project and name the files you used. Check those filenames yourself. Then try one small change with an existing test command. For example: Read TASK.md. Change only src/parseCsvLine.js. Run npm test once after editing and report the actual result. Compare the final file diff with the request; a convincing explanation is not a substitute for changed files and passing tests.
Our bounded Windows test: what happened
We used a fictional, four-file JavaScript project whose CSV parser incorrectly split quoted commas. We set the six test cases before the run. npm test initially passed three and failed three. We launched the Web UI separately, then submitted the coding task through Harness’s command-line headless profile with the same disposable workspace. This distinction matters: the coding result below is not a claim that we completed the task through the Web UI.
| Check | Observed result on September 23 |
|---|---|
| Setup | Official package 0.1.5-rc.3 launched on Windows with Node.js 24.19.0; the local Web UI loaded and showed the selected workspace and DeepSeek Flash. |
| Task and model | One synthetic CSV repair task; deepseek-flash; at most ten API requests through our test limit. |
| Agent’s attempt | It read the task and tests and reasoned about a parser change. Glob failed to launch its search command. Write and Edit then failed before saving the proposed change. |
| Blocking error | SetFileSecurityW EACCES (Win32 5) while preparing a file write; a shell attempt also failed at SetNamedSecurityInfoW ... grantWrite. |
| Verified outcome | No source file changed. An independent check after the stopped run still showed three passing and three failing tests. Harness did not run the requested post-edit test. |
| Recorded usage | The dedicated DeepSeek key’s Usage view reported ten requests, 97,253 tokens and a rounded cost below $0.01. This was one limited run, not a price-per-task estimate. |
Download the original four-file test project plus a README (ZIP) to inspect the exact task, original parser and tests. The README was added after the run; the four tested files are unchanged. The ZIP contains no API key, private site data or fix. Run npm test inside the extracted folder to see the baseline; your result may differ with another Node version or a changed Harness release.
How to read this result: the failure happened in this computer’s Windows sandbox/ACL setup before the tool could write. It is not evidence that every Windows installation fails, or a benchmark of the model’s ability to repair CSV code. We did not give the agent broad/full access to force the run to continue. If you see this error, keep the test project isolated, review the local permissions and the exact Harness release, and report a reproducible case to the official project discussions. Do not treat a wider access mode as a harmless troubleshooting switch.
Does DeepSeek Harness cost money?
The Harness code is published under the MIT license, but the model provider is a separate cost decision. A DeepSeek Platform key uses the API’s current token billing; another provider may bill differently. The DeepSeek pricing guide and calculator explain current API rates without assuming a fixed number of messages per dollar. Our small run’s dashboard displayed <$0.01; it does not predict what a larger repository or a longer agent session will cost.
What to check before giving an agent your code
Harness can run commands and act on local files. DeepSeek’s Safe Use Policy recommends a limited-privilege VM or container, avoiding sensitive data, reviewing generated code and tests, and requiring human approval for consequential operations. A fresh sample repository is a sensible first workspace. Keep credentials and unrelated documents out of it; inspect requested permissions and the actual diff before using a change in production.
Common questions
Is DeepSeek Harness the same as DeepSeek Chat?
No. Chat is the official consumer assistant. Harness is a local developer tool that connects a model to project and command tools. Use the DeepSeek overview to choose the right route.
Does it require a DeepSeek API key?
It needs a usable model provider. The built-in DeepSeek route uses a DeepSeek Platform key, while the official provider guide also documents other and custom providers. Check the account, model and pricing for the route you select.
Is it ready for important work?
The official repository labels it a developer preview with expected compatibility-breaking changes. Our one Windows test reached the UI but failed at a file-write step. Use a reversible project, pin the version you evaluated, and keep human review in the workflow.
Sources and corrections: official repository and launch instructions, Web UI guide, provider configuration, and Safe Use Policy, reviewed September 23, 2026. The test observations above are ours and limited to the recorded setup. Report a correction.