Timeline

DeepSeek open-sources DeepSeek Harness, its coding-agent framework

MIT-licensed, built on the Cordis plugin framework and in developer preview; DeepSeek had used a minimal mode of it to benchmark its models, and the repository held about 260,000 stars by 24 October.

  • Open weights & ecosystem
  • Models & capabilities
  • Minor

DeepSeek released DeepSeek Harness, an agent harness for coding and general knowledge work, as open-source code under the MIT licence, on 13 August 2026. Forklog quoted DeepSeek’s post on X announcing “DeepSeek Harness v0.1” in “Developer Preview” and “opening it up to developers building agent harnesses worldwide”. The GitHub repository was created that day, and the earliest release listed on GitHub, a release candidate, is dated 17 August.

A harness is the layer around a model that gives it tools, a terminal, sessions and a loop for running long tasks. The repository’s README says DeepSeek Harness is built on an “everything-is-a-plugin” architecture using the Cordis framework, so models, tools and interfaces are all plugins. It runs from npx @deepseek-ai/dsh web as a local web interface, and DeepSeek’s product page also describes a desktop app. The README warns that “there will be compatibility-breaking changes” and points users to a safety notice before running it. Forklog framed the project as an open alternative to Anthropic’s Claude Code, a comparison DeepSeek did not make in the text quoted.

The release followed earlier mentions. When DeepSeek published its V4-Flash update on 31 July, its API change log said code-agent benchmarks for V4-Flash used a “DeepSeek Harness minimal mode (to be released soon)”. The harness thus entered public view as the framework DeepSeek used to score its own models, which matters when comparing its numbers with those from other vendors’ harnesses.

It has become part of DeepSeek’s model releases. The V4.1-Flash-1017 announcement in October asked users to upgrade to version 0.2.1-rc.1 or later for the best results with the model, and a further release candidate followed on 23 October. As of 24 October the GitHub repository showed about 260,000 stars and 31,000 forks. Stars measure attention rather than use, and the figure says nothing on its own about how many developers run the software.