The full toolset, stock.
File editing, shell, file and web search, skills, planning, goals, subagents, and workflows.
DeepSeek ships dsh · August 13
DeepSeek Harness — dsh — is an MIT-licensed agent runtime built on the Cordis framework, where everything is a plugin: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, even the UI. One npx command starts it. 112,000 GitHub stars landed in the first two days.
Developer preview v0.1. The README's warning is in capitals: THERE WILL BE COMPATIBILITY-BREAKING CHANGES. "The model is the soul of an agent" is DeepSeek's own framing. Star count checked August 15.
capability groups, each a swappable Cordis plugin. Select, swap, or extend any of them in configuration — without changing the harness source.
Four presets, one backplane
dsh ships four runtime modes built from the same plugins. Scrub through them — Minimal is the one to remember, because it's the rig the V4 benchmark numbers ran on.
File editing, shell, file and web search, skills, planning, goals, subagents, and workflows.
Standard's capabilities, with tools exposed through the Code Mode SDK so the model combines multi-step operations in a single TypeScript program.
Persistent bash plus str_replace_editor — the minimal environment DeepSeek used for V4's agentic benchmark results.
Standard's capabilities plus runtime inspection, in-memory plugin experiments, and preset-authoring guidance.
When DeepSeek V4 Flash shipped its API on July 31, the docs disclosed that its coding-agent benchmarks were produced on "DeepSeek Harness minimal mode (to be released soon)." On August 13 the harness arrived — open source, alongside the official DeepSeek-V4-Pro-0813 release — which makes the vendor's own benchmark rig reproducible by anyone who cares to check.
on deepseek-ai/deepseek-harness, roughly two days after the repository went public.
developer preview, runnable with one npx command; TypeScript, with a web UI served locally at port 3080.
The Cordis kernel manages plugin mounting, unmounting, and dependencies; everything else — models, tools, skills, sessions, sandboxes, storage, loops, scheduling, the UI — is a plugin wired together by services and events. Developers swap or extend any capability in configuration, without forking the harness source.
DeepSeek's framing: the harness is what lets that soul understand an environment, use tools, and keep working in real-world settings. The layer stops being an implementation detail.
Models increasingly swap behind standardized interfaces. The harness that holds your tools, sandboxes, session history, and orchestration is much stickier — and that's the layer DeepSeek just seeded, for free.
System prompts, reasoning, tool calls and results, subagent scheduling, every context injection — recorded to a session log you can inspect by source in the Trajectory view. Resume, fork, search, and replay all operate on the same event stream.
Four different operations, one log. If a run went sideways, the full record of what the model actually saw is the debugging surface.
Plugin authors tag repos with dsh-plugin for discoverability. The internal test in early August recruited developers of other open-source agent harnesses specifically.
"The model is the soul of an agent." The body is now open source.