deepseek-harness

The DeepSeek Harness GitHub repository is the most consequential open-source release in agent tooling this year: a complete agent runtime, called dsh, built on one architectural bet — everything is a plugin. The model adapter, the tool registry, the session log, even the agent loop itself: none of it is privileged core, all of it replaceable from configuration. Launched in August 2026 by DeepSeek, it has pulled in more than 263,000 stars and 31,000 forks in roughly ten weeks. The DeepSeek Harness GitHub project is not a model; it is the scaffolding that turns a model into a working coding agent — designed so that you can own every piece of that scaffolding. The official home of the project is the deepseek-ai/deepseek-harness repository on GitHub, where you can browse the source code, releases and issue tracker.

What it is

Harness ships as a runnable coding agent with a local Web UI and a headless mode for automation, written in TypeScript and released under the MIT licence. Underneath sits Cordis, a composition framework from the cordiverse project, whose model of composability is what makes the plugin architecture coherent rather than a pile of hooks. Sessions are stored as an append-only event log, which buys resume, fork, replay, transcripts and telemetry almost for free. DeepSeek is explicit that this is a developer preview: the README warns, in capitals, that compatibility-breaking changes will happen. Treat the current release as a serious look at where agent runtimes are heading — not a finished platform.

Key features

• Everything is a plugin — model adapters, tools, sessions and the agent loop are mounted components, not privileged core.
• A real coding agent out of the box — inspect repositories, edit files, run shell commands, search files and the web.
• Plans, skills and subagents — structured work management with delegation built into the runtime.
• Approval policies — enforceable gates on what the agent may do without a human yes.
• Event-log sessions — resume, fork and replay any session; transcripts and telemetry come from the same log.
• Web UI plus headless mode — interactive development locally, or one-shot automated jobs with no server.

Who it is for

Teams building their own agent products, platform engineers evaluating runtimes, and organisations whose audit requirements make a black-box agent a non-starter: when every component is a plugin you can inspect, replace and log, the runtime stops being something you rent and becomes something you own. Agencies building bespoke agents for clients will find the plugin model maps cleanly onto how client requirements actually vary. It is also a reference architecture worth studying even if you never run it. It is not yet for teams who need stability guarantees — the preview label is honest, and production commitments should wait for the interfaces to settle.

Getting started

The fastest route in is one command from the DeepSeek Harness GitHub README — npx @deepseek-ai/dsh web starts the Web UI on your machine; choose a workspace, register a model API key, and describe a task. For automation, headless mode runs one-shot jobs without the server. Our advice for evaluation: spend your first session on the architecture document rather than the agent itself, then mount one small plugin of your own — replacing a default component teaches you more about the runtime’s real seams than a week of prompting it. Pin your versions, and expect to re-pin as the preview evolves.

Our take

Every agent framework claims extensibility; Harness is the first we have seen where extensibility is the load-bearing wall rather than a feature. If the bet pays off, the ecosystem effect is significant: plugins become the unit of sharing, and runtimes compete on composition quality instead of feature checklists. The risks are equally clear — preview software, breaking changes, and a young plugin ecosystem whose conventions have not yet been proven under load. We went deep on the architecture in our MTD blog analysis — read our full deep-dive — including where we would and would not deploy it today. For anyone building agents professionally, the DeepSeek Harness GitHub repository is the project to watch this year.