DeepSeek Harness: What It Is, How It Works, and How to Get Started
DeepSeek Harness is DeepSeek's open-source agent harness, built around one idea: everything is a plugin. Here's what it does, its four run modes, and how to try it.
On this page
DeepSeek's own framing for this project is a simple equation: Agent = Model + Harness. The model supplies the reasoning; the harness is everything that turns that reasoning into an agent that can actually work — tools, memory, a sandbox, a loop that keeps calling the model until the task is done. DeepSeek Harness (often shortened to DSH) is DeepSeek's open-source take on that harness layer, and its defining decision is that literally none of it is fixed: "everything is a plugin."
What is DeepSeek Harness?
DSH is an agent harness released as a developer preview under the MIT license. Its own site states there's "no privileged core to patch" — the model adapter, the tool registry, the session log, and the agent loop itself all ship as plugins, and each one is swappable from configuration without touching source code.
That architecture is built on Cordis, an independent open-source plugin runtime (predating DSH by years, as the plugin kernel behind the Koishi chatbot framework) that DeepSeek vendors in and co-published a formal paper on. Cordis is what makes "everything is a plugin" a literal engineering property rather than a slogan: plugins register services, typed events, and reversible effects into a shared context, and the whole system stays fully composable — swap the model, swap the sandbox, mount a different agent loop entirely — without a restart. If you want the deep technical dive into how that actually works, that's what our Cordis article covers; this page stays at the DSH product level.
DeepSeek Harness at a glance
- Tagline: "Everything is a plugin"
- Core equation: Agent = Model + Harness
- Built on: Cordis, an open-source plugin runtime
- Status: Developer preview
- License: MIT
- Released: August 13, 2026
- Quick start:
npx @deepseek-ai/dsh web - Links: GitHub · Docs · Official site
Why "everything is a plugin" actually matters
Most agent frameworks give you a fixed core with extension points bolted on. DSH inverts that: the composable pieces DeepSeek names explicitly are "models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI" — and developers can "select, swap, or extend any capability in configuration." Want a different model provider? Swap the model plugin. Want a different sandbox? Swap the sandbox plugin. Want a completely different agent loop? Mount your own.
The payoff isn't just tidiness. An agent harness that's built this way can be reconfigured — or reconfigure itself — without a restart, and because every registration is a Cordis-tracked effect, a bad change can be cleanly unwound instead of leaving the system in a half-mutated state. That property matters more than it might look at first glance if you're thinking about where agent harnesses are headed: a harness an agent can safely edit while it's still running is a prerequisite for any serious attempt at self-improving agents, not just a convenience for human developers.
Four ways to run it
DSH ships with four run modes, each a different configuration of the same underlying plugin system:
- Standard mode — the full toolset: file editing, shell, search, skills, planning, and more. This is the baseline, closest to a conventional coding agent.
- Code mode — instead of one tool call per step, the model writes TypeScript that orchestrates multiple tool calls in a single program, cutting down on the back-and-forth a multi-step task would otherwise need.
- Minimal mode — a deliberately stripped-down, two-tool coding agent (a persistent bash session plus a
str_replace-style editor). Useful for benchmarking a model's raw capability without a heavy harness doing extra work around it. - Creator mode — for building new presets: it can inspect the current runtime, experiment with Cordis plugins, and assemble a new combination of plugins, tools, and prompts into a preset of its own.
Because all four are just different plugin configurations on the same runtime, moving between "full-featured coding assistant" and "bare-bones benchmark harness" is a config change, not a different codebase.
Nothing the model sees goes unrecorded
DSH's session model is an append-only event log — DeepSeek's own description is that "everything the model sees is recorded in an append-only session log." Every message, tool call, and tool result becomes part of that log, and the harness's Trajectory view reconstructs a full run from it, supporting inspection, resuming, forking, searching, and replaying a session. For anything that runs long enough to compress its context, restart after a failure, or hand work to a sub-agent, having the complete, replayable history — not just whatever's currently in the model's context window — is what makes a run debuggable after the fact instead of a black box.
How to get started
The fastest path is the quick-start command from the official site:
npx @deepseek-ai/dsh web
From there, the GitHub repository has the source, developer docs cover the plugin and Cordis APIs in depth, and DeepSeek's own harness page links to the community plugin directory.
Who should care about DeepSeek Harness
- developers who want a coding agent they can reconfigure down to the model adapter and agent loop
- teams evaluating models who want a Minimal mode-style harness that stays out of the way
- plugin authors and Cordis-curious engineers interested in a production system built on "reversible plugin" architecture
- anyone tracking how agent harnesses are evolving toward supporting safe, self-directed modification
It's less relevant if you just want a polished, install-and-go assistant — DSH is explicitly a developer preview, and its value is in how much of it you can take apart and reconfigure.
FAQ
What is DeepSeek Harness?
DeepSeek Harness (DSH) is DeepSeek's open-source agent harness — the runtime layer around a model that provides tools, memory, sandboxing, and an agent loop, built entirely out of swappable plugins on top of the Cordis plugin runtime.
Is DeepSeek Harness open source?
Yes, it's released under the MIT license and is currently in developer preview.
What does "everything is a plugin" mean in practice?
There's no privileged core — the model adapter, tool registry, session log, and agent loop are all plugins that can be selected, swapped, or extended purely through configuration, without patching source code.
What are DSH's four run modes?
Standard (full toolset), Code mode (multi-step tool orchestration via generated TypeScript), Minimal (a two-tool bare-bones agent), and Creator (for assembling new presets out of Cordis plugins).
How do I try DeepSeek Harness?
Run npx @deepseek-ai/dsh web, or see the GitHub repo and docs for other setup paths.
What is Cordis, and how does it relate to DeepSeek Harness?
Cordis is the open-source plugin runtime DSH is built on — it predates DSH and independently powers the Koishi chatbot framework. See our Cordis deep dive for how its revertible-effects and reactive-coeffects design actually works.