DeepSeek ships open-source Harness with plugin-based design

DeepSeek has released a developer preview of DeepSeek Harness (DSH), an open-source, plugin-based framework the company is previewing for what it calls "agent harness developers worldwide." A harness, in DeepSeek's framing, is the runtime layer that lets an agent understand its environment, use tools, and keep working in real-world settings, as distinct from the model itself, which the company describes only as "the soul of an agent." The project's source code is public on GitHub at deepseek-ai/deepseek-harness, and it can be tried immediately with a single command, npx @deepseek-ai/dsh web, or installed from source by cloning the repository and following its setup instructions.

Harness's central design principle, in the product's own words, is that "everything is a plugin": models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all provided by plugins, built on top of a plugin system called Cordis, whose services and events let the plugins work together. Developers can select, swap, or extend any capability through configuration without changing the Harness source code. The paired principle is that "every run is traceable": everything a model sees during a run, including system prompts, reasoning, tool calls and their results, subagent scheduling, and every piece of injected context, is recorded to an append-only session log. A Trajectory view lets developers inspect these records by source, and resuming, forking, searching, and replaying a session all operate on that same underlying event stream.

The preview ships with four runtime modes. Standard mode includes the full toolset. Code mode uses model-generated code to orchestrate multiple rounds of tool calls. Minimal mode keeps only a shell tool and a file editor, for benchmarking models in a minimal environment. Creator mode lets a developer inspect the runtime currently in use, test Cordis plugins in memory, and combine them into new modes. DeepSeek is explicit that this remains a developer preview: the company says Harness's core plugins and APIs will continue to evolve, that the project is still being tested by developers building agent harnesses, and that it hopes to explore the limits of intelligence together with developers worldwide, using open-source infrastructure built to be reusable and composable.

Key facts

DeepSeek put DeepSeek Harness (DSH), an open-source agent-harness framework, into developer preview, with source code public on GitHub at deepseek-ai/deepseek-harness.
Every capability, including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI, is a swappable plugin, built on a plugin system called Cordis.
Every run is written to an append-only session log covering system prompts, reasoning, tool calls, subagent scheduling, and context injections, viewable in a Trajectory view and resumable, forkable, searchable, and replayable from that same event stream.
The preview offers four runtime modes: Standard (full toolset), Code (model-generated code orchestrates tool calls), Minimal (just a shell tool and a file editor, for benchmarking), and Creator (inspect and recombine Cordis plugins into new modes).
It installs with one command, npx @deepseek-ai/dsh web, or from source via git clone; DeepSeek calls it a developer preview whose core plugins and APIs will keep evolving.

Why it matters

DeepSeek draws a sharp line between the model, which it calls only "the soul of an agent," and the harness: the software layer that lets a model perceive its environment, call tools, and keep operating in real-world settings. By open-sourcing that layer as plugin-based infrastructure rather than shipping another model, DeepSeek is competing on the part of the stack that shapes how a given model behaves once it is wired to tools and left running. Because every capability, from the model itself to storage and scheduling, is a plugin, and every run writes to a resumable, replayable session log, the harness bundles reconfigurability and basic observability into the same open package.

Who it affects

The preview targets developers who build agent harnesses and the broader infrastructure layer around agents: the people choosing or building the runtime that gives a model tools, memory, and persistence, rather than end users of a finished agent product. Because models are only one of the plugin types Harness supports, alongside tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI, teams already committed to a particular toolchain, sandboxing setup, or storage backend are a natural fit, since those pieces can in principle be swapped in without a rewrite. Anyone who has struggled to debug or audit what an autonomous agent actually did during a run is a direct audience too, given the append-only, replayable session log built into the harness itself.

How to use it

Two install paths are given. The fast path is a single command, npx @deepseek-ai/dsh web, which launches the web UI once Node.js is installed. The from-source path is git clone https://github.com/deepseek-ai/deepseek-harness, followed by the setup instructions in the repository. Once running, a developer can choose among Standard mode (the full toolset), Code mode (tool calls orchestrated by model-generated code), Minimal mode (just a shell tool and a file editor, aimed at benchmarking), or Creator mode (inspecting the runtime in use and recombining Cordis plugins into new modes). The source text calls the published code "open-source infrastructure" but names no specific license, and it gives no pricing information for using Harness.

How solid is it

Every claim here comes from DeepSeek's own landing page for the project, not from an independent report or benchmark. No performance numbers, adoption or download figures, or comparison to any other agent-harness framework appear anywhere in the text. No individual engineer, executive, or spokesperson is named as building or announcing Harness, only the product itself and a self-referential "we." The page does not say who built Cordis, the plugin system Harness runs on, or whether Cordis is DeepSeek's own project or an outside one, so a load-bearing piece of the architecture is effectively unattributed. What is here is an architectural description and install instructions, not measured results.

Risks and caveats

This remains a developer preview, and DeepSeek says its core plugins and APIs will keep evolving, so anything built against today's version should be expected to change as the project matures. The page gives no release date, version number, or rollout timeline, so there is no way to tell from the source alone how mature the current preview actually is. It also does not say whether Harness works with models from providers other than DeepSeek or is built specifically around DeepSeek's own models; "models" is listed only as one of several plugin types, with no further detail, leaving the actual breadth of the plugin ecosystem an open question at this stage.

“The model is the soul of an agent.”

— DeepSeek Harness site copy