DeepSeek Harness Holiday Update: Adds an Experimental Claude Code Compatibility Layer

Another Holiday Release

After a splashy Spring Festival release earlier this year, DeepSeek “worked overtime” again for the National Day holiday.

On October 3, DeepSeek Harness shipped a new version. Two things: stronger plugin architecture and cross-platform support, plus an experimental Claude Code Mods compatibility layer.

Harness is the full runtime system that lets large models actually act as agents — managing and executing tasks. DeepSeek’s Agent Harness team was only formed this March.

Claude Code Mods is Anthropic’s extension capability for Claude Code extension developers. DeepSeek’s move: some Anthropic extensions can, with adaptation, run inside the DeepSeek Harness framework.

The Lead Responded Personally the Next Day

The update spread fast through tech communities. On October 4, DeepSeek Harness lead Tianyi Cui posted a response.

He said the compatibility layer’s main purpose is to verify one thing: whether the extension capabilities Claude Code Mods offers plugin authors are a subset of what DSH’s “everything is a plugin” architecture provides.

“Everything is a plugin” is DeepSeek Harness’s core philosophy. Cui said it was set before the project started — before the first line of code. Building open and extensible was DeepSeek’s original intent.

He also poured cold water on it. Full seamless compatibility for all Claude Code Mods in DSH isn’t possible yet. If the community thinks it’s worth investing in, technically it could happen down the road.

Not Cloning a Claude Code

Cui was blunt. Cloning a Claude Code or Codex — a product indistinguishable in tech philosophy and user experience — would not be advancing AGI.

In his view, only doing innovative things advances AGI. This compatibility experiment is a concrete attempt to test the DSH plugin system’s extensibility.

The DSH team’s goal: world-class innovation, work that moves the industry — redefining the newborn category of Agent Harness, even earning recognition and follow-through from AI giants.

He also mentioned spending a good chunk of each day recruiting and interviewing for DeepSeek. The DSH team is still young, with plenty of rough edges, and a long road ahead.

Team and Product Background

Cui, born in the ’90s, studied computer science at Zhejiang University, won six ACM Asia regional gold medals, and spent nine years at top quant firm Jane Street before joining DeepSeek this March.

DeepSeek Harness released its v0.1 developer preview on August 13 this year, open-sourced under MIT. Within 24 hours of launch, it pulled in over 70,000 GitHub stars.

This holiday update lands less than two months after v0.1. Plugin architecture, cross-platform, compatibility layer — all three reinforce the same direction: extensibility. DeepSeek clearly wants DSH to become infrastructure for the agent era.