DeepSeek Harness—whose command is dsh—is an open-source agent harness developed by DeepSeek AI. It is not the DeepSeek model itself. It combines a model, tools, permissions, sessions, persistence, and runtime interfaces into a system that can complete tasks.
The model reasons. The harness lets it use tools, preserve state, and work within defined boundaries.
Where it sits in the stack
A model that receives only text can produce only more text. A harness adds the machinery required to carry out work:
- Describe file, shell, web, and other tools to the model.
- Execute the selected tools and write their results into the session.
- Apply permission policy to decide which operations run directly and which require approval.
- Continue the agent loop until the task completes, fails, or is stopped.
- Preserve an event log so the session can be resumed, replayed, or forked.
What “everything is a plugin” means
DeepSeek Harness is built on Cordis. Model adapters, the tool registry, session logging, and even the agent loop register services, events, and reversible effects on a shared ctx through plugins.
This means more than installing extensions. A capability can be replaced, composed, or unloaded: you can swap a model provider, tool implementation, shell, file-system policy, or UI without rebuilding the entire runtime.
Shortest start
npx @deepseek-ai/dsh webThe default Web UI is available at http://127.0.0.1:3080. You still need to configure model credentials and select a workspace after opening it.
Current maturity
The project is currently labeled Developer Preview. Official guidance warns that breaking changes may occur, so tutorials, plugins, and configuration should be tied to a version and verification date. DSH101 currently tracks CLI 0.1.0-rc.5 and master@47f943859b.
Where to go next
- Want to try it now? Follow the 10-minute Web UI quickstart.
- Need terminal automation? Read Run one-off tasks with Headless.
- Want the internal model? Open the architecture map.
- Ready to extend it? Build your first plugin.
Evidence and revision
Primary sources
This guide is intentionally concise; use the official source or documentation below as the authority for commands, behavior, and risk boundaries.