DeepSeek's AI agent runtime "DeepSeek Harness", recently released as open source, is spreading rapidly in developer communities after receiving positive reviews.

DeepSeek Harness is based on Node.js and is on GitHub as a developer preview under the MIT licence.

A TheNewsTack report said DeepSeek Harness received more than 33,000 stars on GitHub within hours of its release, and the number is rising quickly. A community plug-in ecosystem has already formed actively.

Harness is a "work environment" that connects models, tools and execution processes so AI can perform tasks such as coding or search. It can be described as software that bundles and manages the functions needed for an AI agent to do actual work.

DeepSeek Harness is characterised by the idea that "everything is a plug-in". Model adapters, tool registries, session logs and even the agent loop itself are all plug-ins and can be replaced at any time. It explained that functions can be expanded by adding plug-ins without needing to touch separate core code. The overall structure follows Cordis, a meta-framework based on a paper jointly published by Peking University and DeepSeek researchers.

DeepSeek Harness provides four execution modes: standard, minimal, code and creator. Standard mode is a full coding agent with a filesystem tool, shell access, web search, sub-agents and a planning mode.

Minimal mode is a lightweight version with only two tools: bash and a string-replacement editor. Code mode has the model write and execute a TypeScript SDK instead of making individual function calls, handling in one go what would take five back-and-forth steps. Creator mode adds runtime checks and plug-in experimentation to standard mode functions, allowing users to create their own agent presets.

Session management is also a key feature. All information delivered to the model is recorded in an append-only session log, enabling conversation resumption and branching, replay, recording and telemetry.

Agent sandboxing is also applied strictly. Ri

DeepSeek Harness is not dependent only on DeepSeek's own models. It supports Anthropic, OpenAI, AWS Bedrock, Microsoft Azure and Google's Gemini Enterprise agent platform. It also pro[vides a sub-agent function that directly hands off tasks to Anthropic's "Claude Code" and OpenAI's "Codex".

DeepSeek said it is not currently accepting external developer code contributions, and encouraged GitHub Discussions and plug-in development.