Most AI tools ask you to describe what you want and then hand you an answer. DeepSeek Creator Mode goes one step further: describe the tool you want in plain English, and it builds the tool for you. Announced on 2 November 2026, it is a new mode inside DeepSeek Harness, the company’s open-source AI agent app, and it turns a one-line request into a working plugin — installed, tested and ready to use inside the app itself. In this guide, we explain what DeepSeek Creator Mode is, how it works, what it costs in US dollars, what you can realistically build with it, and the safety warnings you should read before letting it run code on your machine.
Table of Contents
- What Is DeepSeek Creator Mode?
- What Happened on 2 November 2026
- How DeepSeek Creator Mode Works
- The 7 Official DeepSeek Creator Mode Demo Prompts
- What Is DeepSeek Harness?
- How to Install DeepSeek Harness and Try Creator Mode
- DeepSeek Creator Mode Pricing: What It Costs in US Dollars
- What Can You Actually Build with DeepSeek Creator Mode?
- DeepSeek Creator Mode vs Claude Code, Codex and OpenCode
- Safety First: Read This Before You Let It Run Code
- Who Should Try DeepSeek Creator Mode — and Who Should Wait
- The Bigger Picture: DeepSeek’s Push into AI Agents
- Frequently Asked Questions
- The Bottom Line
What Is DeepSeek Creator Mode?
DeepSeek Creator Mode is a chat mode inside DeepSeek Harness, DeepSeek’s open-source AI agent application. You type what you want in ordinary language — for example, “Write a Pomodoro timer plugin for me” — and the agent does the rest. It studies the app’s plugin framework, writes the plugin’s code, installs the plugin into the running application, and then checks that it actually loads and works. No extension code from you, no manual setup, no build step to run by hand.
The idea only makes sense because of how Harness is built. DeepSeek designed the app around one principle: everything is a plugin. Tools, skills and even parts of the interface itself are plugins, so adding a feature means adding a plugin. DeepSeek Creator Mode simply lets the agent write those plugins for you, through conversation, instead of writing them yourself.
It is important to be precise about what launched on 2 November. Creator Mode had already been present in Harness builds since late September 2026, and it kept improving through October. What happened on 2 November was the public showcase: DeepSeek posted a short demo video on X, published a collection of seven official prompts in English and Chinese, and put the feature in front of a much wider audience. That is why the coverage calls the DeepSeek Creator Mode showcase a launch — and why search interest spiked that week.
DeepSeek Harness itself is free and open source under the MIT licence. The model that powers it is billed separately through DeepSeek’s API, and we break down those costs in the pricing section below.
What Happened on 2 November 2026
On Monday 2 November 2026, the DeepSeek Harness team posted a 99-second video with a short line on top: “Turn ideas into plugins that you can use.” The video showed Creator Mode building plugins from prompts typed in everyday English, including a design workspace, a sandbox game and a 3D visualiser that shows how a language model works. Bloomberg’s coverage the same day added two more examples: a plugin that renders 3D models and a video editing workbench.
Alongside the video, DeepSeek published seven example prompts, in both English and Chinese, so anyone could try to reproduce the demos or adapt them into their own tools. The list is a good map of what the feature is aimed at: practical workbenches and visual tools, not chatbots. A paper-reading tool, a poster design tool and a dashboard showing conversation statistics were among them.
DeepSeek was careful to add a warning that many readers missed. Some of the plugins in the video were refined over several rounds of conversation, so a single prompt may not reproduce exactly what you see in the demo. That is honest, and it matters: DeepSeek Creator Mode is a real capability, but the polished video is the best case, not the guaranteed first try.
The announcement landed in a busy week for DeepSeek. Two days earlier, on 31 October, the company had open-sourced dsh-superagent, a separate research prototype of a coding agent that works on a GitHub repository. And on 17 October it had released DeepSeek-V4.1-Flash-1017, the model that Harness uses by default and that DeepSeek says was tuned specifically for the app. DeepSeek Creator Mode is the consumer-facing headline of that larger agent push.
How DeepSeek Creator Mode Works
Using DeepSeek Creator Mode is deliberately simple on the surface. The chat box in DeepSeek Harness has a mode menu, and Creator is one of four modes you can pick before you type. Once you select it and send your request, the agent follows a clear sequence:
- It loads the plugin-development skill. Creator Mode gives the agent a skill that explains how Harness plugins are structured, so it does not have to guess the framework from scratch.
- It inspects the app. The agent looks at where the new plugin can live — which part of the interface can host it, which tools it can call, and what already exists so it does not duplicate a feature.
- It writes the plugin. The agent writes the plugin’s files, including its code and its interface, in the app’s plugin format.
- It installs the plugin. The finished plugin is installed into the running app through the Plugin Manager, without you restarting anything.
- It verifies the plugin works. The agent checks that the plugin loads and behaves as requested, and reports back in the chat.
The official demo makes the flow concrete. DeepSeek asked for a floating Pomodoro timer — a small always-on-top window for tracking work sessions — and the agent built it, installed it and confirmed it was live in about five and a half minutes. That is the whole pitch in one number: an idea typed in one line, a working tool less than six minutes later.
Behind the scenes, Creator Mode leans on the same machinery as the rest of Harness. The app runs on DeepSeek’s API by default, using the deepseek-flash model, which serves DeepSeek-V4.1-Flash-1017. That model supports a 1-million-token context window and up to 384,000 output tokens, which is part of why the agent can hold a whole plugin’s design in view while it works. You can also connect other providers — Anthropic, OpenAI, Kimi and Zhipu — with your own API keys, or any endpoint that speaks the OpenAI or Anthropic API format.
The 7 Official DeepSeek Creator Mode Demo Prompts
DeepSeek published seven prompts to show the range of what Creator Mode can attempt. They are worth reading as a group, because they reveal the feature’s real target: tools that do a job, not toys.
- A Pomodoro timer plugin — the headline demo, a floating timer built in about 5.5 minutes from a single line.
- A design workspace — a canvas-style tool for laying out visual work inside Harness.
- A sandbox game — a playable game built as a plugin, showing that Creator Mode can produce interactive interfaces.
- A 3D visualiser for language models — a tool that shows how a large language model works, in three dimensions.
- A 3D model renderer — a plugin for viewing 3D models, reported by Bloomberg among the launch demos.
- A video editing workbench — an editing surface built from a plain-English description.
- A paper-reading tool and a conversation-statistics dashboard — practical utilities for research and for tracking your own Harness usage.
Two cautions apply to all seven. First, DeepSeek says some were refined across several conversation rounds, so your first attempt at the same prompt may look simpler than the video. Second, these are starting points, not a catalogue: the point of Creator Mode is that you describe your tool, not that you copy DeepSeek’s seven.
What Is DeepSeek Harness?
DeepSeek Creator Mode only makes sense inside its host app, so it helps to understand what DeepSeek Harness actually is. DeepSeek describes it plainly in its own README: “DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.” It is the runtime that sits around a model — the software that turns a language model into an agent that can work with files, run code, browse the web and carry out multi-step tasks.
Harness launched in public preview worldwide in late September 2026, and it is built on Cordis, an architecture where everything is a plugin. That design is what makes Creator Mode possible at all: because tools, skills and interface pieces are all plugins, an agent that can write a plugin can extend the app in any direction.
The app has four built-in modes, each one a bundle of plugins, and you pick one from a menu above the chat box:
| Mode | What it is for |
|---|---|
| Standard | The default: everyday work with code, files and information. |
| PTC | Jobs that call tools in batches, then filter, count or summarise the results. |
| Minimal | Terminal and image reading only, with context compaction — meant for testing and comparing a model’s basic performance. |
| Creator | Customising Harness itself by conversation: the agent writes plugins that add features or interface elements, or combines tools and prompts into your own mode. |
Beyond the modes, Harness covers the usual agent ground: organising documents, analysing spreadsheets, writing and fixing code, researching with citations, running scheduled background tasks, and inspecting execution traces when a tool call goes wrong. Its official plugin list includes Agent Teams (a team roster and shared task board), voice input with local transcription, a shell plugin, an agent loop, a subagent plugin and web search through DeepSeek’s search provider. Scheduled tasks, auto-approval review and a terminal are listed as experimental.
The project’s scale is striking for software only a few months old. The repository, github.com/deepseek-ai/deepseek-harness, was created on 13 August 2026 and had roughly 269,000 stars and 32,700 forks when checked on 2 November. That is the kind of traction that usually takes established projects years to reach, and it is a large part of why Creator Mode is being searched for right now.
How to Install DeepSeek Harness and Try Creator Mode
There are three ways to run DeepSeek Harness, and the right one depends on how comfortable you are with a terminal.
1. The desktop app (easiest). DeepSeek ships packaged desktop apps for Apple silicon Macs and 64-bit Windows, available from the official Harness page. The Windows installer is about 290 MB and the macOS build about 370 MB, and each bundles its own Node runtime, so nothing else needs installing. There is no Linux package yet — Linux users should use the web UI below. Download the app, open it, and you are in.
2. The local web UI. If you have Node.js installed, one command starts the web interface on your own machine:
npx @deepseek-ai/dsh web
That serves the UI at http://127.0.0.1:3080 in your browser. It is the same app, running locally, and it is the route most developers use on Linux or when they want the web version.
3. From source. Clone the full repository and follow the setup instructions inside it:
git clone https://github.com/deepseek-ai/deepseek-harness
Whichever route you take, the first-run setup has the same three steps. First, add a DeepSeek API key under Settings, then Models — paste a key from platform.deepseek.com into the DeepSeek card and save it. Second, choose a workspace folder, which is the directory the agent is allowed to work in. Third, pick a mode from the menu above the chat box and select Creator.
Then describe the plugin you want. Start small, the way DeepSeek’s own demo does: “Write a Pomodoro timer plugin for me.” Watch what the agent does, check the result, and then refine it in conversation — “make the timer float above other windows” or “add a 25-minute work session and a 5-minute break.” Creator Mode is designed for that back-and-forth, and the demos that look most polished were built exactly that way, over several rounds.
One practical detail for Windows users: on Windows, Harness uses a PowerShell plugin instead of Bash. DeepSeek’s shell documentation lists a pwsh-local executor that runs commands as fresh PowerShell processes, and a separate pwsh-sandbox executor that confines commands through the sandbox capability. If you plan to let Creator Mode run commands on a Windows machine, check which executor is actually active before you trust it with a repository you care about.
DeepSeek Creator Mode Pricing: What It Costs in US Dollars
The short answer is that DeepSeek Creator Mode itself costs nothing — and the model behind it costs very little. The longer answer has two parts, because the app and the model are billed separately.
The app is free. DeepSeek Harness is open source under the MIT licence, and no page we found lists a plan, a tier or a subscription for it. Downloading the desktop app, running the web UI and using Creator Mode all cost $0. Community plugins shared under the dsh-plugin tag on GitHub are free too.
The model is billed per token. Harness runs on a DeepSeek API key, and the default model is DeepSeek-V4.1-Flash-1017, served under the API name deepseek-flash. DeepSeek’s official pricing page lists the following peak rates, verified against multiple pricing trackers on 7 November 2026:
| DeepSeek-V4.1-Flash-1017 | Per 1 million tokens |
|---|---|
| Input (cache miss), peak | $0.30 |
| Input (cache hit), peak | $0.006 |
| Output, peak | $1.20 |
| Input, off-peak | $0.15 |
| Output, off-peak | $0.60 |
Off-peak means every hour outside weekday peak windows — peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays — so a plugin-building session on a weekend evening bills at half price. Put another way, a Creator Mode session that uses 1 million input tokens and 200,000 output tokens costs about $0.54 at peak rates, or about $0.27 off-peak. For most personal projects, that is pocket change.
Two cost details are worth knowing. First, DeepSeek also offers web search and fetch at $0.75 per 1,000 successful requests each, though calls made through DeepSeek Harness were free until 1 November 2026 — check the current pricing page before assuming that still applies. Second, if you connect a different provider’s key — Anthropic, OpenAI, Kimi or Zhipu — you pay that provider’s rates instead, which are usually higher than DeepSeek’s. The cheapest way to run DeepSeek Creator Mode is with DeepSeek’s own model, which is also the model DeepSeek tuned specifically for Harness.
For comparison, the model that powers most of the rival agent tools costs far more per token. The site’s earlier coverage of Cognition’s SWE-2 Flash shows the same pattern across the industry: a cheaper, faster workhorse model is now the default choice for agent work, and DeepSeek’s $0.30/$1.20 pricing is at the very cheap end of that scale.
What Can You Actually Build with DeepSeek Creator Mode?
The honest answer is: small, useful, single-purpose tools — and the demos are a fair guide to the ceiling right now. A Pomodoro timer, a paper-reading tool, a dashboard, a design canvas, a 3D viewer. These are real, working plugins, but they are also the kind of tool a single developer could write in an afternoon. DeepSeek Creator Mode’s value is not that it builds things you could never build; it is that it builds them in minutes, from a sentence, without you touching the framework.
That makes it genuinely useful in a few concrete situations:
- Personal productivity tools. A timer, a habit tracker, a small dashboard for your own work — the kind of tool that is too small to justify a weekend of coding but worth having.
- Workflow glue. A plugin that pulls your Harness conversations into a summary, reformats files the way your team likes them, or adds a shortcut for a task you repeat every day.
- Visual and research helpers. A 3D viewer for a model you are studying, a paper reader for research, a statistics panel for your own sessions.
- Learning the Harness framework. Reading the plugin code Creator Mode writes is one of the fastest ways to learn how Harness plugins work, because the agent explains what it built as it goes.
The limits are just as real. DeepSeek has published no independent success rates, generation costs, average build times or reliability benchmarks for Creator Mode. The demos that look most complete were refined over several rounds, so a single prompt will often produce a first draft that needs follow-up. Plugins are code, and code the agent writes can have bugs — including bugs that only appear when the plugin runs on your files. And because Harness is a developer preview, its core plugins and APIs are still changing, so a plugin built today may need rework as the platform matures.
It is also worth saying plainly what Creator Mode is not. It is not a no-code app builder for the general public, and it is not a way to ship a commercial product without a developer. It is aimed at professional developers and enterprise agent work, and the people who will get the most from it are the ones who can read the code it writes and spot when it has gone wrong.
DeepSeek Creator Mode vs Claude Code, Codex and OpenCode
Creator Mode does not exist in a vacuum. The agent-tool space is crowded, and the comparison that matters most is not “which agent is smartest” but “what does each one let you build, and where does it run?”
| DeepSeek Creator Mode | Claude Code | OpenAI Codex | OpenCode | |
|---|---|---|---|---|
| What it builds | Plugins that extend the agent app itself | Code in your repository | Code in your repository | Code in your repository |
| How you ask | Plain-English chat inside Harness | Terminal or IDE chat | Terminal, IDE or web app | Terminal or IDE chat |
| Licence | MIT, fully open source | Proprietary app, Anthropic models | Proprietary app, OpenAI models | MIT, open source |
| Default model cost | $0.30 in / $1.20 out per 1M tokens (peak) | Claude subscription or API rates | ChatGPT plan or API rates | Bring your own key |
| Runs on | macOS (Apple silicon), Windows, web UI | macOS, Linux, Windows, web | macOS, Linux, Windows, web | macOS, Linux, Windows |
| Maturity | Developer preview, no security audit | Generally available | Generally available | Open source, active |
The difference that matters is the object being built. Claude Code, Codex and OpenCode are coding agents: you point them at a repository and they write code in it. DeepSeek Creator Mode is a meta tool: you point it at the agent app itself, and it writes a plugin that changes what the app can do. That is a narrower idea, but it is also a new one — none of the big rivals lets you extend the agent’s own interface by chatting with it.
There is also a price story here. DeepSeek’s model is dramatically cheaper per token than the frontier models the rivals default to, and because Harness is free and open source, the whole stack can be run without a subscription. For a developer in the US weighing a $20-a-month subscription against a pay-per-token setup, that is a real choice — though it comes with the preview-software trade-offs covered in the safety section below.
One more comparison is worth making inside DeepSeek’s own family. The company also open-sourced dsh-superagent on 31 October, a research prototype of a coding agent that works on a GitHub repository. Where Creator Mode builds plugins for Harness, dsh-superagent is closer to a conventional coding agent. Our earlier guide to DeepSeek’s super agent covers that project in detail. The two together show DeepSeek attacking the agent problem from both ends at once: the agent that writes your code, and the agent that writes your tools.
The wider context is a race. OpenAI is moving Codex from a coding assistant towards code infrastructure — our guide to OpenAI’s planned acquisition of Pierre Computer Company explains how Code Storage fits that plan — while DeepSeek is betting that an open, plugin-based harness anyone can extend will win developers over. Both visions are plausible, and Creator Mode is DeepSeek’s most distinctive card in that hand.
Safety First: Read This Before You Let It Run Code
This is the most important section in this guide, so read it before you install anything. DeepSeek Creator Mode is powerful precisely because it writes and runs code on your machine — and that is exactly what makes it risky.
Harness has not had a security audit. DeepSeek says so itself, in the project’s safety notice on GitHub: the software is experimental developer-preview software that has not been through a security audit. A plugin the agent writes is code running on your machine. It can run model-generated code and commands, and if a plugin or a prompt goes wrong, it may damage files or leak credentials. That is DeepSeek’s own warning, not ours, and it should be taken literally.
Run it in isolation. Independent reviewers who tested Harness in early November gave the same advice: run it strictly inside an isolated container or on a secondary machine, not on the computer that holds your work, your passwords and your main projects. Use a dedicated workspace folder with nothing sensitive in it, and keep your API keys out of any folder the agent can read.
Check the sandbox. Harness has a sandbox executor that confines commands — on Windows, the pwsh-sandbox executor — but you should verify which executor is actually running before you let Creator Mode loose on anything important. The default matters more than the feature list.
Turn off telemetry if you want to. The desktop app sends usage analytics by default. There is a toggle in Settings that stops uploading session logs when you use DeepSeek’s official API. If you are building plugins for a client or an employer, switch it off before your first session, not after.
Read the code it writes. Creator Mode’s output is a plugin, and plugins are code. Before you install one into your daily workflow, read what it does — or at least skim it for file access, network calls and anything that touches credentials. The agent will explain its own work if you ask it to, and asking is free.
None of this means you should avoid DeepSeek Creator Mode — and the safety rules above are what make DeepSeek Creator Mode usable in practice. It means you should treat it the way you would treat any powerful developer tool in preview: useful, exciting, and kept at arm’s length from anything you cannot afford to lose.
Who Should Try DeepSeek Creator Mode — and Who Should Wait
Try it now if you are:
- A developer who already uses DeepSeek’s API. You have the key, you know the model, and the cost is pennies. Creator Mode is a natural next step.
- Someone who builds small tools for yourself. If you have ever wanted a tiny utility that no one else would bother to build, this is the fastest way to get one.
- A researcher or tinkerer. The 3D visualiser and paper-reading demos show the direction: tools for understanding, not just for shipping code.
- Curious about open agent infrastructure. Harness is MIT-licensed and its repository is public, so you can read every line of the framework Creator Mode writes against.
Wait if you are:
- Not a developer. Creator Mode is aimed at professional developers and enterprise agent work. If you cannot read the code it writes, you cannot check its work — and in preview software, checking the work is the whole job.
- Working on a machine with sensitive data. Until you have an isolated setup, the no-security-audit warning is a reason to hold off, not a footnote to ignore.
- Looking for a finished product. Harness is a developer preview with compatibility-breaking changes ahead. Plugins built today may need rework as the platform matures, and the polished demos took several rounds of conversation to reach.
- On Linux and expecting a desktop app. There is no Linux package yet, so you will be running the web UI — which works fine, but is not the one-click experience Mac and Windows users get.
The Bigger Picture: DeepSeek’s Push into AI Agents
DeepSeek Creator Mode is best understood as one move in a larger strategy that has unfolded across October and November 2026. The company that became famous for cheap, powerful open-weight models is now building the software layer around them — the harness, the agent, the plugins, the tools that turn a model into something that does work.
The sequence tells the story. In August, the Harness repository was created. In late September, the desktop app arrived in worldwide public preview for Mac and Windows. On 17 October, DeepSeek-V4.1-Flash-1017 shipped, tuned for Harness. On 31 October, dsh-superagent appeared on GitHub. On 2 November, Creator Mode was showcased with its seven demo prompts. Each step made the last one more useful: a better model, a wider app, a prototype coding agent, and then a way for users to extend the app themselves.
There is also a business logic to it. DeepSeek’s API prices — $0.30 in and $1.20 out per million tokens at peak, half that off-peak — are a fraction of what the big US labs charge. Cheap models only matter if developers actually build on them, and an open, extensible harness is how a model company turns a price advantage into an ecosystem. Every plugin a user builds in Creator Mode is another reason to stay inside DeepSeek’s world, and every plugin shared under the dsh-plugin tag on GitHub is free marketing for the platform.
Whether that bet pays off depends on the two things DeepSeek has not yet proved: reliability and safety. No independent benchmarks exist for Creator Mode, no security audit has been done, and the platform is still changing fast enough that today’s plugins may break tomorrow. The rivals — Claude Code, Codex, OpenCode — are more polished and more predictable. But none of them lets you rebuild the tool itself by talking to it, and none of them is free and open source from top to bottom. That combination is why DeepSeek Creator Mode is worth watching, and why DeepSeek Creator Mode is being searched for right now.
Frequently Asked Questions
Is DeepSeek Creator Mode free?
Yes, the feature itself is free. DeepSeek Harness is open source under the MIT licence with no subscription. You only pay for the model tokens it uses through DeepSeek’s API — $0.30 input and $1.20 output per million tokens at peak rates, half that off-peak, for the default DeepSeek-V4.1-Flash-1017 model.
What is DeepSeek Creator Mode?
DeepSeek Creator Mode is a chat mode inside DeepSeek Harness that builds working plugins from plain-English requests. You describe the tool you want, and the agent writes the plugin code, installs it into the running app through the Plugin Manager, and checks that it works. It was publicly showcased on 2 November 2026 with a demo video and seven official example prompts.
How do I use DeepSeek Creator Mode?
Install DeepSeek Harness — the desktop app for Apple silicon Macs or 64-bit Windows, or the web UI with npx @deepseek-ai/dsh web — add a DeepSeek API key under Settings, then Models, choose a workspace folder, and select Creator from the mode menu above the chat box. Then type what you want to build, for example “Write a Pomodoro timer plugin for me.”
Is DeepSeek Harness safe to use?
DeepSeek says Harness is experimental developer-preview software that has not had a security audit, and that model-generated code can damage files or leak credentials if something goes wrong. Run it in an isolated container or on a secondary machine, use a dedicated workspace folder, check which sandbox executor is active, and read the plugin code before you rely on it. The desktop app also sends usage analytics by default, which you can switch off in Settings.
What can I build with DeepSeek Creator Mode?
Small, single-purpose tools: a Pomodoro timer, a paper-reading tool, a design workspace, a 3D visualiser, a video editing workbench, a statistics dashboard. DeepSeek’s seven official demo prompts show the range. Larger or more complex tools are possible, but expect several rounds of conversation and some debugging — the polished demos were refined over multiple rounds.
Does DeepSeek Creator Mode work on Linux?
There is no Linux desktop package yet. Linux users can run the web UI with npx @deepseek-ai/dsh web after installing Node.js, which opens the interface at http://127.0.0.1:3080.
What model does DeepSeek Creator Mode use?
By default it uses DeepSeek-V4.1-Flash-1017, served under the API name deepseek-flash, with a 1-million-token context window and up to 384,000 output tokens. You can also connect Anthropic, OpenAI, Kimi or Zhipu API keys, or any OpenAI- or Anthropic-compatible endpoint.
The Bottom Line
DeepSeek Creator Mode is a small feature with a large idea behind it: if an AI agent can write code, why shouldn’t it write the tools it works with? Described in one sentence, it turns plain-English requests into working plugins inside DeepSeek Harness — installed and checked in minutes, not hours. The Pomodoro timer demo, built in about five and a half minutes, is the proof of concept, and the seven official prompts show a range that runs from productivity timers to 3D visualisers and video workbenches.
The case for trying it is strong. The app is free and open source, the model behind it costs $0.30 in and $1.20 out per million tokens at peak — half that off-peak — and the whole stack runs on your own machine. For a developer who already uses DeepSeek’s API, there is almost no reason not to spend an evening with it.
The case for caution is just as strong. Harness is a developer preview with no security audit, its APIs are still changing, and the demos took several rounds to polish. Run it in isolation, read the code it writes, and keep it away from anything you cannot afford to lose. If you do that, DeepSeek Creator Mode is one of the most interesting things DeepSeek has shipped this year — not because it replaces developers, but because it gives every developer a faster way to build the tools only they would ever want.
