In one line
DeepSeek Harness composes model connections, tools, session logs, and the agent loop as plugins to build an agent's working environment.
Repository: deepseek-ai/deepseek-harness
What should you separate when adding an internal tool to an agent that reads code? DeepSeek Harness, featured in our resource library, uses a Cordis-based plugin structure. Start with model configuration and a workspace, then explore a repository, try a small fix, and examine a tool integration.
What can you do with it?
The CLI is named dsh. Model adapters, the tool registry, session logging, and the agent loop are plugins replaceable through configuration. The Web UI supports file reading and editing, command execution, delegation, and planning.
The project is a developer preview with compatibility-breaking changes expected. Its safety notice says it has not undergone a security audit and should not be treated as production-ready. We recommend beginning with one task in a dedicated environment without sensitive data.
Getting started
Install Node.js, then start the Web UI:
npx @deepseek-ai/dsh web
The default address is http://127.0.0.1:3080, and a local launch opens the browser. Save a model key under Settings → Models, then add and select the project directory through Choose workspace. A fresh UI requires a selected workspace before accepting session input. Keep keys in configuration rather than prompts.
Scenario 1: Get an introduction to a repository
The official Web UI guide begins with summarizing a repository and identifying its main packages. You could add explicit boundaries to that request:
Explain this repository's main packages and their roles.
Identify entry points and connections between packages, citing supporting files.
List anything you could not verify separately. Do not edit files.
Open the referenced files and compare them with the explanation before narrowing the next question. The guide says the UI requests approval where the active permission policy requires it. If a command is proposed, check whether it is needed for the exploration task.
Scenario 2: Investigate a failing test
The official Python SDK example includes a request to inspect a repository and fix failing tests. In the Web UI, you can choose a similar bounded task and begin with investigation:
I want to investigate failing tests in this repository.
First find the existing test procedure and explain the commands to run.
After confirming a failure, propose causes and a minimal change scope,
then wait for approval.
After a change, compare the original failure with the rerun result. The official SDK example also shows an explicit Harness home and workspace. Its minimal SDK profile has a different tool composition from the Web UI, so keep their configurations distinct.
Scenario 3: Design an internal tool connection
Before adding a plugin, describe the inputs and outputs the agent should be allowed to use. The official architecture separates model adapters from tools exposed to the model. Choose one read-only task and try:
I want to connect internal document search as a read-only tool.
First define inputs, outputs, permitted documents, and error handling.
Find the relevant extension point in the official architecture.
Wait before installing anything or changing configuration.
Review the proposed plugin's provenance and code, the API it would call, and the data it would expose. Check how far each permission reaches before broadening the task. Record package and plugin versions alongside configuration.
Enterprise adoption checks
Useful capability: Separate extension points let a team examine model and tool integration independently. Use the official Web UI guide and SDK example to choose the execution form the task needs.
Operating burden and security: Review plugins and compatibility changes. The safety notice recommends least privilege, a disposable VM, container, or dedicated environment, and file backups. Sandboxing and approval do not guarantee isolation; assess exposed files, networks, processes, and credentials separately.
License: The repository uses MIT and provides third-party dependency notices. Review both documents and assess model API terms and costs separately.
Original sources and further reading
- DeepSeek Harness repository
- README and startup command
- Official safety notice
- Web UI guide
- Architecture document
- LICENSE and third-party notices
- Read alongside OpenCode's plan and build agents and team practices in mattpocock/skills.
Wrapping up
The Hypemarc Builder Community will keep sharing the AI news and open source projects we explore each week. We'll continue gathering useful resources to share with fellow builders.
- Open chat: AI Builders
- Adoption advice and inquiries: Contact form
- Get the next post: Blog RSS · AI resource library