DeepSeek Harness (dsh) is the MIT-licensed developer-preview agent harness DeepSeek published on 13 August 2026. It is not a new foundation model. The official formula is Agent = Model + Harness, so the object under review is the runtime layer that sits beside the model.
📑Table of Contents
Where it sits in the stack
It belongs in the same category as Claude Code and Codex: a local runtime that can read and edit files, run a shell, maintain a plan, and delegate work. What sets Cordis apart is that those capabilities arrive as swappable plugins. The decision is narrower than the launch headlines suggest. Should you launch it in an empty working directory today, and should you refuse to pin its config or plugins into a production repository while the preview still advertises breaking changes?
What you can take from this article
What DeepSeek Harness actually shipped
Official definition and preview status
The release is a swappable agent runtime, not a model card. The deepseek-ai/deepseek-harness README defines dsh as an OSS agent harness on Cordis, under the slogan Everything is a Plugin.
The contract at the time of research was:
- The license is MIT
- The status is developer preview
- The README warns in capitals that compatibility will break
- The repository had zero tags
Distribution paths and version skew
There are three official distribution paths:
- the product page
- the npm package
@deepseek-ai/dsh(0.1.0-rc.6at research time) - the GitHub source
The npm rc does not match the repository release commit (0.1.0-rc.5, 13 August 2026). That skew is itself a preview signal.
Running in production requires built artifacts; a raw TypeScript checkout is not the supported way to run it.
Same-day stack split
Keep the same-day stack split. 51CTO treats V4 Pro as the model layer, the Responses API as the protocol layer, and DSH as the runtime.
Mixing V4-Pro agent headlines or Codex setup into a trial of the harness leads you to the wrong target.
Popularity is a discovery signal
Treat popularity only as a signal for discovery.
- VentureBeat recorded about 27,500 GitHub stars and 2,000 forks on launch day
- The GitHub UI at research time showed about 124k stars and 12.2k forks
Star counts measure attention. They do not certify a stable API or fitness for production use.
Read this preview as a harness you can break into parts through configuration, not as the next model to benchmark first.
What “Everything is a Plugin” lets you swap
Swappable surface
The official product page lists these capabilities as plugins you can swap:
- models
- tools
- skills
- sessions
- sandboxes
- storage
- loops
- scheduling
- UI
The intended approach is not a fork-and-patch of the core. Cordis mounts, unmounts, and resolves dependencies so you can swap or extend capabilities in configuration. What helps is being able to break it into parts, not a kitchen-sink IDE.
Session tracking
Session tracking is part of that design. The runtime keeps an append-only log of:
- system prompts
- reasoning
- tool calls and results
- subagent scheduling
- context injection
The trajectory can resume, fork, search, and replay. That approach is closer to an event log than to a chat you try to reconstruct from memory.
Independent coverage and layer split
Independent coverage widens the list slightly while keeping the layers distinct.
- The Register includes filesystems and orchestration and compares Cordis composability to a VS Code extension host: dynamic add/remove plus dependency management.
- 36Kr treats MCP as a connection standard and the harness as the execution-control layer, then flags plugin supply-chain risk before any production install.
Being able to connect a tool is not the same as controlling how it runs.
How this sits beside other harnesses
MindStudio treats the ability to load Claude Code or Codex as sub-agent plugins as the real delta, and notes YAML routing toward other providers or local models.
That is a composability claim, not a replacement claim. Supervision of a machine-local session and swarm control live on other layers; see Paseo and Puppetmaster.
- DSH is inside the runtime
- Paseo watches a host session
- Puppetmaster is a control plane for multiple agents
Operational rule
The operational rule is small. In preview, do not install a dsh-plugin whose origin you cannot name. Decomposing capabilities in config is not the same as adding unaudited power. Scoring or repairing agent loops is a different job, closer to Better Harness.
How to pick among the four runtime modes
Official names versus press aliases
You do not need all four modes on the first night. Use the official names from the product and keep press aliases in a footnote. Pandaily and 36Kr sometimes call Code “PTC” and Creator “Creation” or “Creative”.
Mode comparison
| Mode | Official role | Main capabilities | Best first use |
|---|---|---|---|
| Standard | full coding agent | file edit / shell / search / skills / planning / goals / subagents / workflows | ordinary repository work |
| Code | Code Mode SDK | Standard-class tools, plus TypeScript programs that compose multiple tools | multi-step tool chains |
| Minimal | two tools | persistent bash + str_replace_editor | model-only benches |
| Creator | custom preset | Standard plus runtime inspection, in-memory plugin experiments, and preset guidance | building your own mode |
Source: official product page (as of August 2026).
CLI profile
The npm CLI is a second axis of decision.
dsh webis an alias for--profile webdsh --profile headless "job"runs one persisted session and exits
Choose a UI conversation or a one-shot job before you debate modes.
First-session rule
A simple rule is all you need.
- Start in Standard
- Use Minimal when you want the model with almost no harness extras
- Use Creator when you need to inspect the runtime before adding a plugin
- Reach for Code only when you already want TypeScript orchestration
Making Code the default from the start puts the orchestration cost up front, before you know whether the Web UI even boots.
Fastest path to the local Web UI
Prerequisites
You need Node.js. The directory you launch from becomes the default workspace root, but a fresh Web UI still has no workspace selected. The session composer stays locked until you pick a directory in the UI.
Shortest path
This is the shortest path:
- Run
npx @deepseek-ai/dsh web. - Open the default URL
http://127.0.0.1:3080from the README. To change the port, pass the flag to the web app, for exampledsh --profile web --port 8080. - Save a DeepSeek API key under Settings → Models. No server restart is required.
- Use Choose workspace to add and select the launch directory. Without that step, you cannot start a conversation.
- Keep the first task small. The official docs example is “Summarize this repository and identify its main packages.”
Approval surface
Operations that need approval under the current permission policy surface in the Web UI. The primary procedure is the official Web UI quickstart.
The agent can read and edit files, run commands, delegate, and maintain a plan. Do not point the workspace at a production repository before you have seen what the approval prompts actually cover.
Source checkout path
The source path is separate:
git clonepnpm installpnpm run buildpnpm dsh web
You need built artifacts. Because npm’s rc and the repository rc differ, record which path you used.
First-night observation
Pandaily recorded a first-night flow that asks for an API key before the main screen, and noted slowness on long tasks.
A successful boot is not evidence that long-horizon jobs are ready. One summary task is enough for the first check. Do not treat a large refactor as the acceptance test.
Trial now, or keep it off production
What actually decides a trial
Stars and “open rival to Claude Code” headlines do not decide a trial. Purpose, mode, approval scope, config location, plugin origin, and mix-ups across layers do.
- Is the goal a model eval, or a look at harness / plugin design?
- Is Standard enough for the first session (or do you truly need Minimal / Code / Creator)?
- After the API key and workspace, did you inspect the approval prompts?
- Will you keep config and plugins out of the production repo while breaking changes are promised?
- Will you refuse untrusted
dsh-pluginpackages? - Are you still mixing V4-Pro or the Responses API with this runtime?
Next action
What comes next is small.
- In an empty working directory, run
npx @deepseek-ai/dsh web - Add the API key
- Choose a workspace
- Run one summary task
- If that holds, try a small file or shell job in Standard
Do not pin shared plugins, team defaults, or CI to this preview.
How independent coverage lines up
Independent coverage lines up like this:
- 51CTO is explicit: do not make a preview a production dependency
- 36Kr flags plugin supply-chain risk
- VentureBeat frames the project as candidate open infrastructure beside Claude Code, not a finished product replacement
Together those points make a local trial reasonable and make pinning this for production premature.
FAQ
Is DeepSeek Harness a new model?
Q. Is DeepSeek Harness a new model?
A. No. The official split is Agent = Model + Harness. V4-Pro is the model layer. DSH is the runtime. Same-day launches still need separate evals.
What is the shortest launch command?
Q. What is the shortest launch command?
A. After Node.js is installed, run npx @deepseek-ai/dsh web. The default URL is http://127.0.0.1:3080. You cannot start a chat until a workspace is selected.
How do the four modes differ?
Q. How do the four modes differ?
A. Standard is ordinary coding work. Code composes tools in TypeScript. Minimal is bash plus an editor for model benches. Creator is for custom presets. Press names PTC and Creation map to official Code and Creator.
Can it replace Claude Code or Codex?
Q. Can it replace Claude Code or Codex?
A. It is an OSS harness in the same category. MindStudio reads Claude Code and Codex as loadable plugins. VentureBeat does not treat it as a finished-product replacement. That is not a reason to uninstall what you already run.
Is it safe for production?
Q. Is it safe for production?
A. The official preview promises breaking changes. A local trial is fine. Do not freeze config, plugins, or CI around it yet.
Related articles:
Summary
The DeepSeek Harness delta is not model quality. It is the ability to decompose models through UI as Cordis plugins.
The ship is an MIT developer preview, and the npm 0.1.0-rc.6 versus repository 0.1.0-rc.5 skew is a reason not to rush a dependency.
The next move:
- Boot the Web UI in an empty directory
- Run one Standard task
- Compare the approval surface and preview constraints with your own work
Production pins, shared plugins, and CI come after the compatibility contract settles.
Author
krona23
Over 20 years in the IT industry, serving as Division Head and CTO at multiple companies running large-scale web services in Japan. Experienced across Windows, iOS, Android, and web development. Currently focused on AI-native transformation. At DevGENT, sharing practical guides on AI code editors, automation tools, and LLMs in three languages.
🔥 Most Popular
- Claude Desktop Won't Install? Windows & Mac Fixes That Worked (2026)
- Claude Pricing: Free, Pro, Max & Team Plans Compared (August 2026)
- Claude Code CLI vs Web vs Desktop: A Daily User's Guide (2026)
- Obsidian Visualization Tools: Graph, Canvas, Excalidraw (2026)
- Can an 8GB GPU Run a 35B MoE? Check Host RAM, Cache, and Speed












Leave a Reply