DeepSeek Harness
Also known as: dsh, DSH, @deepseek-ai/dsh
DeepSeek Harness (dsh) is an MIT-licensed, TypeScript-based agent runtime released by DeepSeek AI on August 13, 2026. One command, npx @deepseek-ai/dsh web, starts a local web interface on your own machine. From there the harness gives any language model a workspace, a tool registry, a sandboxed shell, and a persistent session log to operate from. Without something like it, a model can only produce text; it cannot open files, run commands, or remember what it did two turns ago.
The architectural idea that made dsh go viral is radical simplicity: everything is a Cordis plugin. The model adapter, tool registry, agent loop, session storage, sandbox, and UI are all plugins. None of them is privileged. Any can be swapped out or replaced from configuration without touching core source code. That means you can point dsh at Claude, GPT, or any OpenAI-compatible endpoint, not just DeepSeek models, by adding an adapter in settings.
dsh is in developer preview with breaking changes expected before a stable 1.0 release. The project shipped alongside DeepSeek's V4-Pro-0813 model and a new peak/off-peak pricing switch. Its rapid adoption reflects growing demand for an open, frontier-lab-backed harness that builders can actually inspect and modify, rather than treat as a black box. Treat it like a power tool: extraordinary access for experimentation, but not something to expose on a network or run against credentials you care about until it reaches stability.