DeepSeek Harness Developer Preview Officially Launched: Rebuilding Agent Infrastructure with "Everything is a Plugin"
Beijing, August 13, 2026 — DeepSeek officially released DeepSeek Harness (DSH) v0.1 Developer Preview to global developers tonight, and simultaneously open‑sourced it under the MIT License. This marks a pivotal move by DeepSeek in the agent space, extending its strategic footprint from pure model capabilities to agent execution infrastructure.
DeepSeek Harness is not a new foundation model or an API client. Instead, it is a full‑featured agent runtime framework that bridges models with file systems, terminals, web browsers, code tools, and other agents, while orchestrating context, tool calls, and task execution. The official formula is simple: Model + Harness = Agent – the model handles reasoning, and Harness handles actual execution.
"Everything is a Plugin": Extreme Modularity
The core design philosophy of DeepSeek Harness is "Everything is a Plugin". All agent capabilities – models, tools, skills, sessions, sandboxes, storage, agent loops, task scheduling, and user interfaces – are assembled from plugins that can be freely swapped, combined, and reconfigured.
The framework is built on the Cordis plugin meta‑framework. Cordis handles only plugin loading, unloading, and dependency management; all business components of Harness are independent Cordis plugins that interact via Cordis’ service and event mechanisms, enabling flexible composition at the configuration level. Developers can customize or extend any module without modifying DeepSeek Harness’ source code – just by adding or replacing plugins.
This design delivers three direct benefits: First, models and the runtime environment are decoupled – developers can keep the same session and toolset while swapping model adapters. Second, components can be combined at configuration time, drastically lowering the bar for customisation. Third, agent capabilities form an independent ecosystem – developers do not need to maintain a forked version of Harness; they just publish their own plugins.
Four Modes for Diverse Scenarios
DeepSeek Harness comes with four pre‑configured runtime modes, each loading a specific set of plugins tailored to different use cases:
Standard Mode: Loads the full toolchain, providing comprehensive file editing, Shell, file/web retrieval, Skills, planning, sub‑agents, and workflow capabilities – suitable for general development tasks.
PTC Mode (Programmatic Tool Calling): The model generates TypeScript code to orchestrate multi‑step tool‑chaining, combining several operations in one go.
Minimal Mode: Includes only Shell and file editing tools, designed for model benchmarking in minimal environments.
Create Mode: Enables runtime state inspection and in‑memory debugging of Cordis plugins, allowing developers to generate entirely new runtime modes.
Open Source Ecosystem and Community Response
The DeepSeek Harness project is now open source on GitHub and can be launched directly via npm. Developers with Node.js installed can run npx @deepseek-ai/dsh web to start the Web UI locally. The official npm plugin ecosystem is also open for contributions.
Community reaction has been overwhelming – within half an hour, the GitHub repository surpassed 10,000 stars, and as of press time, it has exceeded 30,000 stars. Notably, DeepSeek has launched a separate WeChat official account for Harness, “DeepSeek Harness Team,” featuring a black whale logo, distinctly different from the blue whale used for DeepSeek’s model products.
In terms of positioning, DeepSeek Harness competes with OpenAI Codex and Anthropic Claude Cowork (Claude Code), aiming to be an AI productivity tool focused on programming and office scenarios. However, unlike those products, Harness is not a fixed coding agent – it is a model‑agnostic, open‑source agent infrastructure framework that developers can use to assemble their own agent systems.
Preview Status and Future Roadmap
DeepSeek emphasises that the current v0.1 release is still a Developer Preview: “There are still many details to be improved and polished, and core plugins and basic interfaces will evolve rapidly in upcoming iterations.” The team also warns that breaking changes may occur during these fast‑paced updates.
From internal project kick‑off in May 2026 to today’s public debut, DeepSeek Harness marks an extension of DeepSeek’s open‑source boundary – from models to agent engineering systems. With the simultaneous release of V4‑Pro, DeepSeek is building a complete technology stack that spans foundation models and agent execution layers – perhaps exactly what that surfacing “black whale” signals for DeepSeek’s next chapter.
