The MIT licensed dsh (DeepSeek Harness) runtime runs AI agents in a browser through a single npm command, with DeepSeek positioning it as an open alternative to Anthropic's Claude Cowork and warning of breaking changes ahead.
DeepSeek has put an open-source harness for running AI agents in a browser into developer preview. The runtime, called dsh, ships under an MIT license and launches with the single command npx @deepseek-ai/dsh web, per the project's GitHub repository and DeepSeek's X announcement.
The architecture sits on the Cordis meta-framework with an explicit design principle: "Everything is a plugin." That seam turns the runtime into a substrate third-party developers can extend, rather than a finished agent product.
DeepSeek positions the preview against Anthropic's Claude Cowork, a closed agent harness. The framing appears in DeepSeek's own announcement and is echoed in TechNode, VentureBeat, and CryptoBriefing.
The repo is tagged as a developer preview, with DeepSeek warning that compatibility-breaking changes are expected. As of August 13, the project had roughly 27,500 stars and 2,000 forks on GitHub, a social-proof snapshot rather than adoption. The agent team is led by Cui Tianyi, a former Jane Street quant who joined DeepSeek in March.
What remains unsettled: stability through the preview, real-world parity with Claude Cowork, and whether the plugin seam attracts the third-party extension that would make the substrate stick.