DeepSeek Harness went from open-sourcing to 160k stars in six days, and most of the discussion still files it under another agent project. I spent two days in the source and the rest of the week making it do real work. My first conclusion is about categories: dsh is not a model and not a framework — it is the first open-source agent runtime shipped as a finished product. A model lab took the harness it runs internally and released the whole rig.
What it is. A complete harness: session runtime, a tool surface (shell, files, web search and fetch, code execution, sandboxing), plan mode, step-by-step approvals, context compaction, a skill system — and four front ends: a CLI, an official web workbench, a headless mode, an editor protocol. Model access is open: DeepSeek’s endpoint is the default, but any OpenAI- or Anthropic-compatible endpoint is a few lines of config away. The most accurate analogy I have: dsh is to DeepSeek what Claude Code is to Anthropic — except open source, model-neutral, and shipping with a GUI.
What it does. The default persona is a coding agent, but the tool surface is general. This week I used it to reorganize a downloads folder, run a research report, and clean spreadsheets — all it took was writing a few Chinese skills, short playbooks that teach it how a class of task should be done. It works inside a local sandbox, proposes a plan before touching anything, and every step is approvable and replayable. That is the line between it and a chatbot: the output is not an answer, it is finished work.
Where it sits. Three comparisons cover the map. Against frameworks like LangChain and LangGraph: a framework hands you parts to build an agent; dsh hands you the assembled machine — you write skills and plugins, not orchestration code. Against its closest siblings, Claude Code and Codex: comparable capability, but both are closed and tied to their vendor’s models, and dsh opens both doors. Against aggregator clients like Cherry Studio: those are shells; dsh is the engine that gets mounted inside them — Cherry wired it in as one more runtime within three days of the release, which tells you exactly which layer each one lives on.
Frameworks are too raw, chat clients too cooked. dsh eats the layer in between: works out of the box, comes apart for rebuilding.
Where it loses. Honesty required. It is young: still in release candidates, with breaking interface changes landing every few days — building against it is a commitment to keep up. The default experience assumes a developer: install Node, open a terminal, edit a profile; ordinary users stall on step one. And the ecosystem barely exists — skills, themes, documentation, especially outside English. That gap is where my evenings have been going.
If you keep one sentence: for the first time there is an open-source, model-neutral agent runtime with a full interface, shipped ready to work. The remaining question is not whether it runs. It is what you want it to do.