Overview
Released on August 13, 2026, DeepSeek Harness (DSH) is an open-source Agent runtime framework distributed under the MIT license. The repository accumulated 15,000 stars within only five days. Its core design philosophy is everything is a plugin: models, tools, skills, dialogue management, sandboxes, storage, scheduling, UI and cross-agent collaboration capabilities can all be extended or replaced. The core Cordis module is responsible for plugin loading, dependency management and lifecycle control.
Unlike all-in-one integrated tools such as Claude Code and Codex, DeepSeek Harness decouples every capability into modular pluggable components. Developers can swap the default LLM model, replace the frontend UI, connect custom storage services, or expand tool sets via community plugins without modifying core source code. Another critical feature is full execution traceability. Every action observed by the model is logged into persistent dialogue history, including system prompts, reasoning chains, tool inputs and outputs, agent scheduling events, and context injection. The built-in Trajectory view supports trace review, branching, restoration and playback, which greatly simplifies Agent debugging.
The rapid growth to 15k stars (9.5k stars achieved within the first 48 hours) reflects clear market demand. Developers require a flexible, auditable, modular Agent runtime environment free from vendor lock-in. This guide covers two deployment approaches, official and custom model provider configuration, four built-in runtime modes, and curated high-quality community plugins selected from the 15,000-star ecosystem.
Core Architecture: Agent = Model + Harness
At the heart of DeepSeek Harness sits the Cordis scheduler, whose sole responsibility is plugin orchestration: loading, unloading and resolving dependencies. All functional modules are implemented as independent plugins: models, tools, skills, dialogue sessions, sandboxes, storage, scheduling logic and the web UI. All components can be freely combined through the configuration layer. This architecture allows developers to build tailored Agent workflows rather than adapting to fixed tool pipelines.
Two Deployment Methods
1. Quick Start via NPX (Recommended for First-Time Testing)
After installing Node.js, run the single command below to launch the web service directly. There is no requirement to clone the repository locally, making it ideal for rapid validation.
npx @deepseek-ai/dsh web
The service runs locally at http://127.0.0.1:3080. You can open the address in any browser to complete initial configuration.
2. Source Code Installation (For Custom Development)
If you plan to build custom DSH plugins or modify framework logic, deploy from source:
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm run web
Important Note: DeepSeek Harness remains in developer preview. Official documentation warns compatibility-breaking changes may be introduced between versions. Production users should lock specific release versions.
Model Provider Configuration
Option 1: Connect Official DeepSeek API
After launching the Web UI, navigate to Settings → Models. The DeepSeek provider card is pre-configured.
- Enter your API Key obtained from
platform.deepseek.com - Save settings; the model becomes available immediately without server restart.
Credentials are stored inside $DSH_HOME/.credentials.yaml. The web interface only references keys and does not persist plaintext secrets in frontend configurations. You can override the default workspace path with an environment variable:
export DSH_HOME=/path/to/your/dsh-home
npx @deepseek-ai/dsh web
Alternatively, pass the API key via environment variables during startup:
export DEEPSEEK_API_KEY=sk-your-deepseek-key
npx @deepseek-ai/dsh web
Option 2: Add Custom OpenAI-Compatible Providers
DeepSeek Harness supports any OpenAI-compatible endpoint as a custom model provider. This design suits teams that aggregate multiple LLM services. To configure via the Web UI:
- Open Settings → Models and click Add a custom provider
- Fill in provider identifier, display name, base URL, protocol and API key
- Click Fetch available models to pull the model list, select required models and save
Teams running multiple LLM backends can manage unified access control and traffic routing with an API gateway such as Treerouter to streamline provider integration.
For production deployment, static configuration via settings.yaml is preferred. Once configured, select the provider in the model selector, and the chosen model will become the default for new dialogue sessions. If you enable multimodal support for certain models, add corresponding input flags in the configuration file to avoid image request rejection.
Four Built-in Runtime Modes
DeepSeek Harness provides four preset modes for different task scenarios:
-
Standard Mode (Default) Full-featured mode for daily coding Agents. It supports file editing, shell execution, web search, skill invocation, planning, sub-agent scheduling and workflow management. Suitable for most software engineering tasks.
-
Program-Tool-Call Mode (PTC Mode) Built on Standard Mode. The model outputs TypeScript programs to execute multi-step tool invocations instead of sequential single-step calls. Optimized for complex multi-stage operations and workflow logic design.
-
Minimal Mode Only retains two core tools:
readandreplace. Extra tool capabilities are removed. Targeted for capability benchmark testing or resource-limited environments. -
Creation Mode Inherits all Standard Mode capabilities, with additional runtime inspection, in-memory plugin experimentation and preset creation interfaces. Developers building custom Agent workflows are recommended to start here.
Inside the Web UI, switch presets via the mode selector on the left sidebar. Source deployments allow users to create and save custom presets for automatic loading at startup.
Curated High-Star Community Plugins
Hundreds of plugins have been published under the GitHub dsh-plugin topic. Below are high-quality projects sorted by community popularity (data captured August 18, 2026):
open-design (88.3k stars)
Supports prototype generation, landing page design, slide creation, image and video synthesis. It can export outputs as HTML, PDF, PPTX and MP4 files. Compatible with mainstream design models. Installation command:
dsh plugin --profile web add nexu-io/open-design
OpenViking (28.9k stars)
A long-term context database for Agent memory, unified knowledge base and RAG management. It solves persistent information storage across long sessions and multi-agent scenarios. Historical work records, code standards and project documentation can be stored, indexed and retrieved continuously.
Voyager (19.5k stars)
A browser extension and prompt manager toolkit. It offers a visual prompt library management interface and works seamlessly with DeepSeek Harness Web UI for teams maintaining standardized prompt templates.
Archify (13.8k stars)
Generates architecture diagrams, workflow charts, sequence diagrams and data lifecycle visualizations. Outputs self-contained HTML files with structural validation and export functions, ideal for technical documentation and system design reviews.
EverOS (12.0k stars)
Markdown-first persistent storage for cross-workspace tool and task data. It emphasizes local data ownership and offline usability, distinguishing it from OpenViking’s centralized knowledge base positioning.
First Task: Run Your Initial Agent Workflow
After finishing deployment and model setup, select a project directory as the workspace, then submit the sample instruction:
Summarize this repository and identify its main packages.
The Agent will read the directory structure, analyze module dependencies and generate a summary report. If shell execution requires approval, the Web UI will pop up confirmation prompts before running commands.
Additional beginner task templates for capability verification:
List all TODO comments in the codebase and create a summary.
Run the test suite and summarize the results.
Find all functions longer than 100 lines and refactor the longest one.
The Trajectory view is one of the most valuable debugging tools within DSH. Developers can observe complete Agent reasoning chains, all tool call inputs and outputs, and execution latency for each step. The trace architecture records system prompts, reasoning chains, tool invocations, sub-agent activities and context injection in chronological order.
Frequently Asked Questions
Q: What fundamental differences exist between DeepSeek Harness and Claude Code / Codex?
A: Claude Code and Codex are closed, end-to-end integrated products. DeepSeek Harness is an open, modular runtime framework. It separates the model, harness and agent loop, allowing flexible swapping of each component without vendor constraints.
Q: Can DSH be used in production during developer preview?
A: Official documentation confirms breaking changes may occur between preview versions. For personal projects and experimental workflows, it is usable. For production services, lock specific stable versions and review changelogs thoroughly before upgrades.
Q: What is Cordis?
A: Cordis acts as the internal plugin scheduler. It only manages plugin loading and dependency resolution and does not implement any agent logic. All functional capabilities are delivered via loosely coupled plugins.
Q: Can I modify the ID of custom model providers?
A: Direct editing is not supported, as saved dialogue history and configuration reference the provider ID. The correct workflow is to add a new provider entry and remove the old one. Display names, base URLs and API keys can be modified freely.
Conclusion
DeepSeek Harness redefines the boundary of Agent runtime design with its "everything is a plugin" architecture. Developers are no longer limited by fixed tool chains. They can freely switch LLM backends, expand tool sets, connect custom storage and build multi-agent collaborative pipelines. The learning curve is gentle: users can complete end-to-end deployment within two minutes and start building workflows powered by official APIs or self-hosted model endpoints.
The rapid growth of community plugins shows strong ecosystem momentum. Open-design, OpenViking, Voyager and Archify immediately extend core capabilities for design generation, persistent memory, prompt management and architecture documentation. As the framework matures, more reusable modules will lower the threshold for building production-grade custom Agents.
All statistics in this article are sourced from official DeepSeek pages, GitHub repository metrics and community plugin data as of August 18, 2026. Since the project remains in preview, installation parameters and compatibility rules may evolve. Always refer to official README documents before deployment.
Learn more：https://treerouter.com
