DeepSeek Harness is an open-source agent harness from DeepSeek AI. Its core philosophy is simple: everything is a plugin.
Every capability—including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI—can be swapped or recomposed.
Official installation paths
Use the package for a quick local launch, or clone the official repository when you want to inspect and build the source.
npx @deepseek-ai/dsh webgit clone https://github.com/deepseek-ai/deepseek-harnessThe Web UI is served at 127.0.0.1:3080 by default.
DeepSeek Harness Plugins
DeepSeek Harness plugins extend the runtime with terminal interfaces, agent collaboration, visual analysis, model integrations, and other focused capabilities.
Vision / Model integration
ModLens is a DSH Web bundle that routes pasted-image evidence for text-only DeepSeek and GLM model variants through a configurable vision engine.
View listingTerminal UI
DSH CC TUI is a Claude Code-style full-screen terminal interface bundle for DeepSeek Harness, with streaming thought display, status information, rollback handling, and context indicators.
View listingAgents / Collaboration
A DeepSeek Harness Web Bundle for natural-language multi-agent teams with a current-session captain, resumable members, dependency-aware tasks, direct messages, and a live activity panel.
View listingVision / Developer tools
DSH Vision Toolkit is a native DSH bundle for structured visual analysis, OCR, grounding, screenshots, artifact output, and Web settings.
View listingOpen source
The official repository is published under the MIT License.
Developer Preview
DeepSeek says the project is iterating rapidly.
Compatibility changes
The official warning explicitly includes breaking changes.
First run
Install Node.js first, use the documented npm command, and treat the first profile as an evaluation environment. The runtime can work with files, tools, credentials, and subprocess policy; review the configuration before bringing it near sensitive work.
The official installation instructions list Node.js before running dsh.
Run the official command, then open the local address it reports.
Configure models and workspace access for a disposable first test.
The model is not the harness
DeepSeek Harness is the runtime that composes context, tools, sessions, and execution policy around a model request. That distinction matters: a change to a tool, adapter, or profile can alter what an agent can see and do without changing the model itself.
Read the official architectureCordis context
DeepSeek documents Cordis as the framework beneath dsh. Plugins contribute services, typed events, and reversible effects, so an extension can be mounted beside the rest of the system rather than patched into a privileged core.
Profiles & bundles
Profiles compose ordered bundles and patches. The official architecture names web and headless as templates, while bundles package configuration rows with the code they mount.
Replaceable seams
Model adapters, tools, persistence, sandbox policy, credentials, and telemetry are described as layers in the composed runtime. That flexibility is useful, but it also makes configuration review part of using the system safely.
Turn flow
The outline below paraphrases the official event flow. It is useful for understanding where an integration changes behavior, not a substitute for the upstream implementation details.
The driver claims the next available input and a queued message.
Registered prompt sections and tool schemas are assembled before the model request.
The model response can produce tool calls that pass through the guarded execution pipeline.
Durable session events preserve the turn; another input or owed tool response can start the next step.
Primary sources
This page is an independent guide. The links below are the primary sources for the runtime, its architecture, documentation, and public package distribution.
FAQ
DeepSeek Harness, also called dsh, is an open-source agent harness developed by DeepSeek AI. Its official architecture describes a plugin-based system in which model adapters, tools, sessions, and the agent loop are composable parts.
DeepSeek labels DeepSeek Harness as a Developer Preview and explicitly warns that compatibility-breaking changes will happen. Assess it against your own change-control, security, and recovery requirements before using it for important work.
After installing Node.js, the official repository documents the command npx @deepseek-ai/dsh web. It says the Web UI is served locally at http://127.0.0.1:3080 by default.
A profile is a named composition of bundles and configuration patches. A bundle is a distribution format for Cordis configuration rows and the code they mount. The official architecture documentation identifies web and headless as shipped profile templates.
The official repository suggests the dsh-plugin GitHub topic for discoverability. AgentPlugin.net separately indexes public community projects with source and compatibility evidence; a listing here is not an official endorsement or security approval.
DeepSeek Harness is the runtime around a model, not the model itself. The official quick start asks you to configure a model provider before starting a session, and the architecture keeps model adapters composable.
AgentPlugin.net is an independent community directory, not operated, endorsed, or supported by DeepSeek. It records source and review evidence for public projects; it does not replace upstream documentation, code review, or a safe evaluation environment.