DeepSeek Harness v0.1 open-sourced under MIT, positioned against Claude Cowork and Codex
On the evening of August 13, DeepSeek open-sourced Harness v0.1 developer preview, hitting 30,000 GitHub stars within two days, a record. Built on the Cordis meta-framework, the everything-is-a-plugin design supports Standard, PTC, Minimal, and Creative modes.
On the evening of August 13, alongside the V4-Pro release, DeepSeek open-sourced the DeepSeek Harness v0.1 developer preview as an agent execution framework, positioning it against Anthropic's Claude Cowork and OpenAI's Codex for programming and office AI productivity scenarios. Released under the MIT license, the GitHub repo crossed 22,000 stars within roughly 1.5 hours, surpassing the previous records set by xAI Grok-1 and DeepSeek-R1, and topped 30,000 stars within two days.
Harness is built on the Cordis meta-framework and enforces an everything-is-a-plugin design philosophy. Models, tools, skills, sessions, sandboxes, storage, loops, schedulers, and UI are all composed from plugins, freely swappable and recombinable. This lets developers extend or replace any component without modifying source code. The dsh command line tool supports Standard, PTC (programmatic tool calling), Minimal, and Creative modes for benchmarking, plugin development, automated execution, and open-ended exploration respectively.
DeepSeek published the Cordis design paper that decomposes model adapters, tool systems, session logs, and Agent Loops into interchangeable components with runtime loading, unloading, and composition. This mirrors OpenAI's Responses API and Anthropic's MCP in spirit but takes a more aggressive fully open-source route. After the Harness release, the Chinese large-model competition has moved beyond base models and APIs into the next frontier of Agent Runtimes and plugin ecosystems.