DeepSeek Harness Creator Mode: Plugins, Setup and Limits

Last updated: November 3, 2026

DeepSeek Harness Creator Mode lets you extend DeepSeek’s open-source agent app by asking for new tools, skills or interface panels in plain language. The agent writes the plugin, installs it into your profile and can make it usable in the same session. Harness is free under MIT; model usage is billed by your provider.

Contents

What’s new at a glance
What Creator Mode does
How to install and set up
How to build a plugin in Creator Mode
Connecting an MCP server by chat
Scheduled tasks and automation
Recent release changes
Creator Mode vs. Standard and other agents
Limitations and safety
Frequently asked questions

DeepSeek Harness, also called dsh, is now in public preview worldwide, according to the official DeepSeek Harness page. As reported by Quartz, DeepSeek introduced Creator Mode on X on November 2, 2026, with examples such as a design workspace, a sandbox game and a 3D LLM visualizer built as plugins.

Interest is high. The GitHub repository showed more than 270,000 stars when we checked, and Google suggestions for “deepseek harness” include creator mode, desktop, github, cli and plugins.

What’s new at a glance

Feature	What it does	Where or when
Creator Mode	Builds tools, skills and UI panels as plugins through chat	“Let Agent create a plugin” entry, added in v0.2.1-rc.1
Runtime inspection	`cordis_inspect_list` and `cordis_inspect_query`, read-only	Creator Mode only
Plugin Manager	Installs config bundles into the current profile	Standard and Creator modes
Experimental plugins	Agent teams, auto approval review, scheduled tasks, voice input	Official experimental plugins
Change requests and worktrees	Review agent edits in isolated worktrees	v0.2.2 alpha builds
Delegation to Codex or Claude CLI	Experimental `subagent_oar` tool	v0.2.2-alpha.2

What Creator Mode does

In Harness, almost everything is a plugin built on the Cordis framework, including tools, model providers and parts of the interface. The product page shows Creator Mode building a Pomodoro timer: the agent loads a plugin-development skill, inspects the running app, writes package.json and client.js, then installs the bundle with the Plugin Manager.

The dynamic Cordis practice guide explains the difference between modes. Standard and Creator both include the Plugin Manager. Creator adds read-only runtime inspection, so the agent can see what services and slots exist before it writes code. The inspection tools cannot call service methods or run code.

Configuration belongs to the current profile and survives restarts. That makes Creator Mode useful for building a personal toolset over time rather than one-off scripts.

How to install and set up

Install Node.js. The 0.2.2 alpha builds require Node.js 24.9 or later.
Run npx @deepseek-ai/dsh web. The web UI opens at http://127.0.0.1:3080; add --no-open to skip opening a browser. The npm “latest” tag was 0.2.1-rc.5 when we checked, and @alpha pointed to 0.2.2-alpha.3.
Open Settings, then Models, and paste an API key from platform.deepseek.com.
Choose a workspace folder. Use a disposable folder or a copy of your project at first.
Start a chat. The agent asks before running operations that need approval.

The providers guide says keys are stored in $DSH_HOME/.credentials.yaml and are write-only in the interface. You can also add Anthropic, OpenAI, Moonshot AI or Z.ai, or a custom model using the openai-completions, openai-responses or anthropic-messages protocol. For costs, see our DeepSeek API pricing answer and the DeepSeek V4.1 Flash guide.

Developers can embed the same agent with the Python SDK (pip install deepseek-harness-sdk, Python 3.10 or newer) or run it headless with dsh --profile headless.

How to build a plugin in Creator Mode

Open plugin management and choose “Let Agent create a plugin.” This opens Creator Mode and keeps your draft until you send it.
Describe the result, not the code. Example: “Add a sidebar panel with a 25-minute focus timer and a button that logs finished sessions to a file in this workspace.”
Let the agent inspect the runtime first. Ask it to list the providers and slots it plans to use.
Review the generated files and approve the install step.
Test the plugin in the same session, then ask for fixes in follow-up messages.

Good prompts name the trigger, the data the plugin may touch and how you will check it. Ask for one capability per plugin so you can remove a broken piece without losing the rest.

Connecting an MCP server by chat

The dynamic Cordis guide gives a ready prompt: “Configure the MCP server at <endpoint> in this profile as demo. Make its tools available now, then call its ping tool and tell me the result.” The agent writes a bundle that adds @deepseek-ai/dsh-mcp-client and installs it with plugin_manager install_bundle.

Check the result. application: applied plus a working mcp__demo__ping call means the tools are live. restart-required means the change is saved but not active until you restart.

Scheduled tasks and automation

The scheduled tasks guide lists schedule_create, schedule_list, schedule_view and schedule_edit in the standard, cordis and ptc presets, but not in the minimal preset. Schedules use RFC 5545 recurrence rules with a minimum interval of one minute, and subagents cannot create them. A practical pattern is a weekday morning task that summarizes new issues in a repository.

Recent release changes

v0.2.1-rc.1 (October 17): added the “Let Agent create a plugin” entry, on-demand Claude Code and Codex bundles, a /compact command and automation tasks in the web UI.
v0.2.2-alpha.2 (October 28): experimental delegation to a local Codex or Claude CLI, developer statistics, change requests and worktrees, six new interface languages and a 50 MiB image limit. Custom models without declared limits now default to a 1M-token context and 128K output.
v0.2.2-alpha.3 (November 3): unified review for change requests and worktrees, an apply_patch tool that is off by default, plugin management in the TypeScript and Python SDKs, and read-before-write checks off by default. Users of the standalone worktree plugin must remove @deepseek-ai/dsh-experimental-tool-worktree from dsh.profile.bundles.

Creator Mode vs. Standard and other agents

Use Standard mode for normal coding and research work. Switch to Creator Mode when the missing piece is a tool or panel rather than an answer. Creator Mode needs more review because it adds code to your profile.

Compared with Claude Code or the Codex CLI, Harness is an extensible app first and a coding agent second. It is closer to OpenClaw in scope, but its plugins can change the interface itself. If you build agents in code, our Cursor Agent SDK guide covers a different approach.

Limitations and safety

The README calls Harness a developer preview and warns of compatibility-breaking changes.
The safety notice says it is experimental, not security-audited and not production-ready, and that agents can run model-generated code.
Since 0.2.2-alpha.2, Web Fetch no longer rejects non-public addresses, so treat local network services as reachable.
DeepSeek has not published separate plan limits for Harness; limits come from your model provider.

Run it with least privilege, in a VM or container where possible, keep backups and read any plugin before you install it.

Frequently asked questions

What is DeepSeek Harness Creator Mode?

Creator Mode is a mode in DeepSeek Harness, DeepSeek’s open-source agent app, where you describe a new tool, skill or interface panel in chat and the agent writes and installs it as a plugin. It adds read-only runtime inspection tools on top of the Standard mode.

Is DeepSeek Harness free?

The Harness software is free and open source under the MIT License. You still pay for model usage through your DeepSeek API account or any third-party provider you connect, at that provider’s published rates.

How do I install DeepSeek Harness?

Install Node.js, then run npx @deepseek-ai/dsh web and open http://127.0.0.1:3080. Add a DeepSeek API key under Settings, then Models, choose a workspace folder and start a chat. A Python SDK is also available through pip.

Can DeepSeek Harness use Claude, GPT or other models?

Yes. The providers guide lists built-in third-party providers including Anthropic, OpenAI, Moonshot AI and Z.ai, plus custom models that use the OpenAI Chat Completions, OpenAI Responses or Anthropic Messages protocols.

Is DeepSeek Harness safe to run on my main computer?

DeepSeek says the project is experimental, not security-audited and not production-ready. Because agents can run model-generated code and install plugins, it recommends least privilege, a disposable VM or container, backups and reviewing plugins before use.

Sources

Image: DeepSeek logo by DeepSeek, via Wikimedia Commons, MIT License (trademarked), placed on a white background and resized. FreeAnswers is not affiliated with DeepSeek.

DeepSeek Harness Creator Mode: Power-User Guide to Plugins, Setup and MCP