Overview
DeepSeek Harness is an agent runtime framework officially open-sourced by DeepSeek on August 13, 2026, under the MIT license. Its positioning formula is Model + Harness = Agent, benchmarking against Claude Code. The framework is based on a microkernel architecture with the Cordis plugin system, where all components can be freely replaced and combined. It offers four operating modes: Standard, Code, Minimal, and Creative, aiming to provide developers with a foundational runtime environment for building flexible and extensible agent applications. Shortly after its release, it quickly gained significant community attention, surpassing 50,000 GitHub Stars within 12 hours and reaching 149,000 Stars in five days, and has received support from the National Supercomputing Internet.
Usage climbed sharply after the desktop build shipped on 30 September. Bloomberg reported on 8 November that daily active users reached 1.54 million in under three months, a figure counting only official API users, with the seven-day average rising from 271,000 when the desktop version launched. For an open-source project positioned as a runtime framework, that volume suggests it has moved from developer experimentation into day-to-day workflows.
Key Features
- Microkernel Plugin Architecture: Based on the Cordis plugin system, all components such as models, tools, skills, sessions, sandboxes, storage, scheduling, and UI are plugins that can be freely replaced and combined, achieving high modularity and customizability.
- Unified Execution Trace: Uses an append-only event log to record every message, tool call, and intermediate state, supporting history replay and error isolation, improving debugging and auditing efficiency.
- Four Operating Modes: Standard mode provides a complete agent environment with shell and web search; Code mode supports multi-step tool calls via SDK; Minimal mode offers persistent shell sessions and text editing for benchmark evaluation; Creative mode supports plugin configuration diagnostics.
- Multimodal Input Support: Version v0.1.0-rc.8 adds native image requests and mixed text-image commands, expanding the agent's perception and interaction capabilities.
- Sub-agents and Tool Calls: Supports sub-agent mechanisms and tool calls, facilitating the construction of complex task workflows, while also adding Windows terminal support to enhance cross-platform usability.
- Quick Start and Integration: Can be launched with the npx @deepseek-ai/dsh web command, and already supports official QQ Bot plugin integration, lowering the integration barrier.
Use Cases
- Developers can quickly build general-purpose agent applications with shell and web search capabilities using Standard mode.
- Use Code mode to integrate multi-step tool calls via SDK, suitable for automated workflows or backend services.
- Utilize Minimal mode for benchmark evaluation, verifying the agent's stable performance in persistent shell sessions.
- Diagnose plugin configurations through Creative mode to optimize compatibility and performance of custom plugin combinations.
- Combine multimodal input and sub-agent features to build complex interaction scenarios supporting mixed text-image instructions.
Pros
- Open-sourced under the MIT license, allowing free use, modification, and commercial use, with a friendly ecosystem.
- The microkernel architecture is flexible, with all components replaceable and combinable, adapting to diverse needs.
- Unified execution trace supports history replay and error isolation, improving development and debugging efficiency.
- Four operating modes cover various scenarios from complete environments to minimal evaluation, with broad applicability.
- High community activity, gaining 149,000 Stars and over 5,100 plugins within five days of release, showing rapid ecosystem growth.
- Received support from the National Supercomputing Internet and integrated with official QQ Bot plugins, offering convenient integration.
Pricing
Free and open source under the MIT license, self-hostable with no seat or call limits. The runtime itself costs nothing; spending comes from whichever model service you connect. Standard and Code modes require your own model API key and are billed at that provider's rates. The official npx @deepseek-ai/dsh web command starts it locally at no extra cost.
Summary
DeepSeek Harness, as an open-source agent runtime framework by DeepSeek, centers on a microkernel plugin architecture and four operating modes, providing a flexible and extensible solution for building agents. Its unified execution trace and multimodal support enhance practicality and debugging capabilities, with enthusiastic community response and rapid ecosystem growth. The developer preview version may contain breaking changes, so it is recommended for evaluation and internal testing. For specific pricing and the latest information, please refer to the official website.
Version History
- Bloomberg: Harness reaches 1.54 million daily active users (2026-11-08): Bloomberg reported that DeepSeek Harness reached 1.54 million daily active users in under three months, counting only official API users, with the seven-day average rising from 271,000 when the desktop version launched on 30 September.