DeepSeek has opened a developer preview of DeepSeek Harness, its open-source framework for building AI agents, released under the MIT license with full source code on GitHub (deepseek-ai/deepseek-harness).

The central design principle is that everything is a plugin. Built on the Cordis kernel, which handles plugin mounting, unmounting, and dependencies, the harness implements every agent capability — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI — as a plugin. Cordis services and events coordinate the plugins, and developers can select, swap, or extend any capability purely through configuration, without modifying the source code.

The second pillar is traceability: everything the model sees is written to an append-only session log, including system prompts, reasoning, tool calls and results, subagent scheduling, and context injections. A Trajectory view lets developers inspect records by source, and resume, fork, search, and replay all operate on the same event stream.

Four runtime modes are available. Standard mode is a full coding agent with file editing, shell, file and web search, skills, planning, goals, subagents, and workflows. Code mode exposes tools through a Code Mode SDK so the model can orchestrate multi-step operations in a single TypeScript program. Minimal mode reduces the agent to persistent bash and a str_replace_editor for benchmarking in a minimal environment. Creator mode adds runtime inspection, in-memory plugin experiments, and preset-authoring guidance for building custom agent presets.

Getting started requires Node.js: the Web UI launches via `npx @deepseek-ai/dsh web`, or the full source can be cloned from the repository. DeepSeek notes the preview is still evolving — core plugins and APIs will continue to change — and points developers to community plugins under the dsh-plugin GitHub topic.