Status correction: The DeepSeek Harness repository, read on September 22, 2026, calls the project a developer preview and says there will be compatibility-breaking changes. Earlier wording on this site that called it generally available, quoted a universal task cost, or treated an rc.8 feature list as current has been withdrawn. This page did not install the project.
What the repository establishes
DeepSeek Harness (dsh) is published by DeepSeek AI under the MIT license. The README describes an everything-is-a-plugin architecture and points to Cordis for that design. This review did not read the linked paper, so the architecture claim stays the README’s description, not an independent design review.
The default branch is master. The GitHub API showed a push on September 22, 2026. Pin the commit you evaluate. A later README is not evidence about the build you ran. The README links documentation at https://deepseek-harness.github.io/deepseek-harness/. That link is the project’s stated docs entry. This check did not audit every page there.
The documented npm launch is npx @deepseek-ai/dsh web. The README says that command starts a Web UI at http://127.0.0.1:3080 by default and opens a browser for a local launch. An SSH launch only prints the host URL. --no-open starts the server without opening a browser. The source path is clone, pnpm install, pnpm run build, then pnpm dsh web. Those are installation instructions from the README. They are not a completed installation on this site, and they do not set a price per task.
Read the safety notice before the first command
SAFETY.md says the project is experimental developer-preview software, has not had a security audit, and must not be treated as secure or production-ready. It can execute model-generated code and commands, load third-party plugins, and access the network, processes, credentials and files made available to it. Incorrect model output, defects, misconfiguration, malicious input or untrusted plugins may change or delete files or disclose credentials.
The same notice says sandboxing, approval prompts and permission controls can reduce risk but do not guarantee isolation. Even a correctly enforced restriction cannot protect a resource the process is allowed to access. Do not use Harness as the only control for an untrusted workload. The notice’s own conditions are: least privilege, a disposable virtual machine or container, backups of files the process can reach, no sensitive credentials unless you accept the risk, and review of plugins and proposed commands before they run. The software is provided without warranty under the MIT license.
Evaluate the runtime separately from the model
Choose one released commit and save its README and SAFETY.md with the log. In an isolated example directory, record how that build reads files, proposes edits, runs checks, handles a denied action and stops a task. Keep the diff and logs. A model’s API compatibility does not determine this runtime’s filesystem or network permissions. Adding a plugin changes the permissions under test, so repeat the same checks after the plugin is loaded.
A model list price, including the dated DeepSeek rates on the cost worksheet, does not become a Harness task price. Harness can add retries, tool calls and local compute that the list rate does not include.
Use a small acceptance task, then stop
The first-task pack provides fixed inputs and checks. Try it without private code, retain failed attempts, and record setup work. Compare that evidence with the current workflow comparisons. Do not describe the result as a MonkeyCode integration test. No such test was run for this update.
Frequently asked questions
Is DeepSeek Harness generally available?
No. The deepseek-ai/deepseek-harness README, read on September 22, 2026, calls it a developer preview and warns of compatibility-breaking changes. Earlier wording on this site that called it generally available has been withdrawn.
Does the safety notice make a local run safe?
No. SAFETY.md says the project has not had a security audit, is not production-ready, and that sandboxing does not guarantee isolation. It can run model-generated commands and load third-party plugins.
Does a documented npm command prove a MonkeyCode integration?
No. The README documents npx @deepseek-ai/dsh web and a source build. This site did not run either command, and it did not test a MonkeyCode plugin or shared permission boundary.