Last updated: 9/30/2026Last verified: 2026-08-14
DeepSeek Harness (dsh) is DeepSeek AI's open-source, local-first AI agent framework built on an everything-is-a-plugin architecture powered by Cordis. Launch the Web UI or CLI locally with npx @deepseek-ai/dsh, then compose models, MCP servers, tools, sandboxes, and workflows into custom agents with parallel sub-agents, browser control, and resumable trajectory logs.
What is DeepSeek Harness?
DeepSeek Harness, also called dsh, is DeepSeek AI's open-source, local-first AI agent framework for building custom agent workflows. It can be launched locally through a Web UI or CLI using npx @deepseek-ai/dsh. The framework is built on the Cordis plugin architecture, allowing models, MCP servers, tools, sandboxes, workflows, and UI components to be composed into flexible agent systems.
Key Features
- Everything-is-a-plugin runtime powered by the Cordis kernel, where models, tools, skills, sessions, sandboxes, loops, scheduling, and UI components are swappable.
- MCP integration that lets MCP servers plug in as tools while model adapters can be swapped across DeepSeek, OpenAI-compatible, and other supported providers.
- Parallel agent orchestration for building workflows with sub-agents that can run and coordinate tasks.
- Append-only Trajectory logs that support workflow resume, fork, and replay for agent runs and sub-agent workflows.
- Browser control capabilities for agents that need to interact with real web pages.
- Vision and image-understanding plugin support for workflows that need to process visual information.
- Local Web UI and CLI access through npx @deepseek-ai/dsh.
Best For
Pricing
DeepSeek Harness is listed as a free tool in the provided information. Because it is described as open-source and local-first, users should still review the official project page or repository for the current license, model-provider costs, and any external service fees that may apply.
Pros & Cons
Pros
- Local-first design gives developers control over how the framework is run and configured.
- Plugin-based architecture makes it possible to swap models, tools, sessions, sandboxes, and UI components.
- MCP support helps connect external tool servers into agent workflows.
- Parallel sub-agent orchestration is useful for more complex workflow design.
- Trajectory logs support resume, fork, and replay, which can improve debugging and iteration.
- Browser control and vision plugins expand the range of tasks agents can attempt.
Cons
- Requires comfort with local development tools such as npx, CLI workflows, and plugin configuration.
- Users may need to configure model providers, MCP servers, sandboxes, and tools before getting full value from the framework.
- The provided information does not include details about hosted deployment, enterprise support, or managed infrastructure.
- Actual performance and reliability will depend on the chosen models, plugins, tools, and local environment.
Alternatives
A popular framework for building LLM applications and agent workflows with integrations for models, tools, retrievers, and orchestration components.
A framework focused on connecting LLMs to data sources, retrieval workflows, and agentic applications.
An open-source framework for building multi-agent AI systems with conversation-based agent coordination.
A multi-agent orchestration framework for designing role-based agents and collaborative workflows.
A developer framework for building agentic applications with tools, handoffs, tracing, and model integrations.
FAQ
Details
Platform
Features
- Everything-is-a-plugin agent runtime on the Cordis kernel — models, tools, skills, sessions, sandboxes, loops, scheduling, and UI are all swappable components
- MCP integration: MCP servers plug in as tools while model adapters are swappable plugins (DeepSeek, OpenAI-compatible, and other providers)
- Parallel agents and orchestration with an append-only Trajectory log supporting resume, fork, and replay for sub-agent workflows
- Browser control and vision/image-understanding plugins extend the harness to see and drive real web pages
Languages
Known limitations
- Developer preview: the README explicitly warns there will be compatibility-breaking changes as the project iterates
- Not a turn-key coding agent — you must configure a model provider (DeepSeek or OpenAI-compatible API key, or a local model) and assemble agents from plugins first
- Running local models on-device needs capable hardware (GPU/VRAM); otherwise you rely on API keys
- Browser automation and vision come from third-party plugins of varying maturity, so stability is not yet guaranteed
