DeepSeek releases plugin-based agent runtime with swappable core components

DeepSeek has made its DeepSeek Harness available as open source under an MIT license, bringing a modular agent runtime designed around extensibility. The Node.js-based tool attracted over 33,000 GitHub stars within hours of its release.

The core philosophy driving the architecture is radical modularity: model adapters, tool registries, session logs, and even the agent loop itself are all plugins that developers can swap out. This means there is no immovable core requiring patches—extending the system simply amounts to adding a new plugin alongside existing ones.

Plugin-first design#

The harness is built atop Cordis, a framework developed by researchers at Peking University and DeepSeek that prioritizes what they call "spatiotemporal composability." The underlying concept is straightforward: modern AI systems need to be able to add and remove components dynamically, with each piece understanding its dependencies and able to interact with others without breaking the whole.

An academic paper by the framework's creators provides the theoretical foundation, with the DeepSeek Harness implementing those principles in practical code.

Four operational modes#

The harness ships with four preset configurations. Standard mode provides a full-featured coding agent complete with filesystem access, shell commands, web search, and the ability to spawn subagents. Minimal mode strips this down to just two essential tools: bash and a text editor. Code mode changes the interaction model—instead of making individual function calls to tools, it generates a TypeScript SDK that lets the model write a single program, reducing a multi-turn sequence into one call. Creator mode is aimed at developers building custom presets; it includes Standard's capabilities plus runtime inspection and plugin experimentation tools.

Model-agnostic compatibility#

Nothing in the harness locks it to DeepSeek's own models. It supports providers including Anthropic, OpenAI, AWS Bedrock, Microsoft Azure, and Google's Gemini, plus custom OpenAI-compatible endpoints. The harness also includes subagent providers that can delegate directly to Anthropic's Claude Code and OpenAI's Codex, with both disabled by default. Additionally, the tool can read configuration files from competing products (Anthropic's , Anthropic's , and others), treating these as compatibility bridges rather than preferred patterns.