DeepSeek Harness

Quick Verdict

DeepSeek Harness (dsh) is an open-source agent harness published by DeepSeek on 2026-08-13. MIT licensed, written in TypeScript, powered by the Cordis kernel. Its architecture has exactly one idea: everything is a plugin — models, tools, skills, sessions, sandboxes, storage, loops, scheduling and even the UI are all supplied by plugins, while the kernel only handles loading, unloading and dependency resolution.

The official framing is Agent = Model + Harness. The model sets the ceiling on intelligence; the harness decides whether that intelligence can actually finish work in a real environment. DeepSeek open-sourced the second half.

Whether to install it now is a simple call: it is a developer preview, and the README says in all caps that THERE WILL BE COMPATIBILITY-BREAKING CHANGES. Use it to try plugins, build experimental workflows and watch the ecosystem — not to run your daily production pipeline. If you want a stable, ready-to-use coding agent, Claude Code and Codex are still the lower-friction picks.

Best For

Developers already using agent coding tools. If you have hours in Claude Code or Codex, you know where agents fail, and only then can you judge whether a harness’s plugin design is any good. The community has already shipped migration plugins from other tools, so the cost of trying is low.

Teams building their own agent products. What you need is a runtime, not a finished app: session management, plugin loading, sandboxing and scheduling take months to build from scratch, and the MIT license means you can just take it and modify it.

People who enjoy plugin ecosystems. Repositories tagged dsh-plugin on GitHub are already in the thousands, ranging from serious multi-agent collaboration and memory evolution to pure fun — mini-games and virtual pets included.

Not for: anyone who wants something stable and maintenance-free that just delivers work. Breaking changes during preview will interrupt your workflow; that need is better served by a mature client such as Cherry Studio or a hosted first-party product.

Key Features

Plugin kernel. Built on Cordis’s dependency-injection model, capabilities load and unload as plugins. Swapping models, toolchains or the interface never requires touching core code or waiting on the vendor’s roadmap.

Web workspace. With Node.js installed, npx @deepseek-ai/dsh web starts the Web UI at http://127.0.0.1:3080 by default. From source: git clone, then pnpm install && pnpm run build && pnpm dsh web.

Sessions as logs. Sessions are recorded as an append-only event log; the Trajectory view lets you restore, fork, search and replay an execution. For debugging agent behaviour this beats a chat window that only shows the final output.

Four run modes. Standard, PTC, minimal and creative — different levels of autonomy and tool-calling strategy, switchable per task type.

Sandbox and workspace. Sandboxing and storage are plugin-level capabilities too, and the agent’s reads, writes and command execution are bound to the workspace you select. You must pick a workspace first or the composer stays disabled — the single most common first-run stumble.

Not locked to DeepSeek models. It defaults to a DeepSeek API key (entered under Settings → Models, no restart needed) but also supports other providers and any OpenAI-compatible endpoint, with a dedicated docs page for third-party model setup.

Vision input (since 2026-08-21). The 0.1.1-rc.1 DeepSeek adapter added the multimodal model DeepSeek-V4-Flash-Vision-Exp, so you can hand images to the agent directly in a session. Note that DeepSeek labels it experimental, ships it as API-only with no open weights, and caps each image at 384 tokens — test it against your own images before relying on it for fine-grained chart reading.

Use Cases

Assemble a coding agent that is genuinely yours. Official kernel plus community plugins: interface, tool set and memory strategy assembled to your habits instead of a vendor’s preset bundle.

Reproducible debugging of agent behaviour. Append-only session logs plus Trajectory forking and replay turn “why did it call the wrong tool there” into a record you can revisit, not something you reconstruct from memory.

Higher-isolation local automation. Sandbox plus workspace binding keeps the agent’s file access and command execution scoped — useful for batch repetitive edits inside one repository.

Multi-agent collaboration experiments. Agent-team plugins already exist, splitting a task across roles. This kind of experiment is hard to run inside a closed product.

Study material. If you want to understand what an agent harness is actually made of, reading dsh’s architecture docs beats any secondhand explainer.

Pricing

The software is free, MIT licensed, commercially usable and modifiable. The official product page lists no paid tiers.

The real cost sits on the model side: dsh is only the shell, and inference burns whatever API credit you configure. Using DeepSeek’s official API bills at its published rates; other providers or self-hosted OpenAI-compatible endpoints bill at theirs.

On versions: @deepseek-ai/dsh has been published on npm since 2026-08-10, and as of 2026-08-21 the latest GitHub release is dsh-v0.1.1-rc.1 — still a release candidate, not yet 1.0. What you get today is a preview, and interfaces will change.

Pros

• Uncompromising architecture. “Everything is a plugin” is not marketing — models and UI are plugins too, so replacing any layer leaves the core untouched.
• MIT license, no restrictions on commercial use or forking.
• Ecosystem growing unusually fast. Repositories tagged dsh-plugin reached the low thousands within days of the public launch — rare density for a framework this young.
• Sessions can be replayed and forked, making agent debugging far better than in a plain chat client.
• No model lock-in; any OpenAI-compatible endpoint works.
• Readable, self-hostable stack for teams that need to audit what their agent runtime does.

Cons

• Developer preview with breaking changes. DeepSeek says so in all caps; do not treat it as stable.
• No formal release. The version is still at rc, and the upgrade path is unclear.
• Uneven plugin quality. The flip side of explosive growth is no review — toy plugins sit next to production-grade ones, and you have to tell them apart.
• Setup barrier. Node.js, workspaces and plugin concepts make it unfriendly to non-developers.
• Docs trail the code. Some capabilities are only discoverable by reading source or community threads.
• Model cost is separate. The software is free; inference is not.

Alternatives

The one-line difference: Claude Code and Codex sell a product; dsh sells a runtime. With the former you configure inside the vendor’s boundaries; with the latter you assemble it yourself, and what it becomes is entirely your call.

FAQ

Is DeepSeek Harness the same thing as the DeepSeek chat product?

No. DeepSeek is the model and chat product; dsh is the framework that lets a model execute tasks in a real environment. In the official formula, the former is Model, the latter is Harness.

Do I have to use DeepSeek models?

No. It prompts for a DeepSeek API key by default, but the docs explicitly support other providers and custom OpenAI-compatible endpoints.

Can I use it for a real project today?

Not recommended. DeepSeek labels it a developer preview and states breaking changes are coming, and the latest npm build is still an rc. Fine for experiments and prototypes; risky for delivery work on a tight schedule.

What is the fastest way to get running?

Install Node.js, run npx @deepseek-ai/dsh web, open http://127.0.0.1:3080, enter an API key under Settings → Models, then select a workspace — without one the composer stays disabled.

Are community plugins safe?

Judge for yourself. Plugins can read and write workspace files and execute commands, so check the repository’s source, stars and last update before installing. Quantity is not quality.

Does it compete with MCP?

Not exactly. MCP addresses how a model connects to external tools; dsh addresses how the whole agent runtime is organised, and tool connectivity is just one of its plugin types.

Bottom Line

DeepSeek Harness is one of the most interesting open agent projects of late 2026 — not because it is polished today, but because it lays the “other half” of an agent open for everyone. An everything-is-a-plugin architecture plus an MIT license grew thousands of community plugins within days of launch, and that density is itself evidence the architecture extends well.

Be clear-eyed too: it is a developer preview, there is no formal release, interfaces will change and plugin quality varies. Install it to learn and experiment, not to replace the production tool you rely on. If you already run something like Claude Code, treating dsh as a side project on a second machine is the right way to use it right now.