Announced 2 Nov 2026 · Sources checked
What creator mode does
DeepSeek introduced creator mode for DeepSeek Harness on Monday, November 2, 2026, announced from the project's account and reported by Quartz and Bloomberg. The feature lets a user describe a plugin in conversation and have Harness build, install and expose it inside the app, rather than requiring the user to write code against a plugin API.
DeepSeek's own Harness page states the case plainly: "Install plugins, or create them through chat in 'Creator mode', to extend tools, skills, and the interface." The page frames the app as "now in public preview worldwide and open source, with composable plugins that extend what agents can do," and lists five work modes: everyday work, coding, research, background tasks and plugins.
Quartz reported that DeepSeek's examples included a design workspace, a sandbox game and a 3D LLM visualiser, and that the company's Harness documentation says users can create plugins through chat in creator mode to extend tools, skills and the interface. Bloomberg's report, as summarised by Quartz, described plugins for generating 3D models and a workbench for editing video, all created from prompts typed in everyday English.
That last detail is the interesting one. A plugin that appears in the sidebar, changes the appearance of the whole application, or adds an always-on-top overlay is a different class of output from a text answer or a generated image. It is software that keeps running inside the app after the conversation that produced it ends.
The seven demos, and what they reveal
DeepSeek published the prompts behind its demo video in a creator mode prompt collection so users could try to reproduce them. The seven are: a left-sidebar plugin that renders 3D models of the Eiffel Tower, the Great Wall, the Forbidden City and one building of the model's choosing; an interactive 3D demonstration of the computation inside a Transformer; a plugin that restyles the whole app with five options including a pixelated look; a poster design workbench with a shader background and editable title, text and shader colours; a video editing workbench with multi-file upload, a preview pane, a timeline and export; a paper reader in the right sidebar that preserves the original layout and offers highlight, note and ask-in-conversation actions on selected text; and a statistics plugin built from the user's actual conversations.
Read as a set, the demos map the surface area the company wants to demonstrate: 3D rendering, app theming, layout, media processing, document handling and self-inspection of conversation data. Two of the seven touch video and 3D, which are the areas where a language model's output is hardest to evaluate and easiest to demo.
DeepSeek also published the caveat, which matters more than the marketing: "Some plugins were refined through several rounds of conversation, so a single prompt may not reproduce exactly what you see in the video." That is an honest limitation and a useful calibration. The demos show what a determined user can reach with iteration, not what one prompt reliably produces.
Where Harness runs and who it is for
Harness is not a consumer chatbot. Quartz describes DeepSeek's agent as geared mainly towards developers rather than everyday consumers, and notes that DeepSeek says Harness is in preview with evolving core plugins and APIs. Users can launch the web interface by installing Node.js and running a command with npx, or clone the full source from the project's repository. DeepSeek described the infrastructure as open source, reusable and composable.
The timing fits a pattern. Quartz reports that DeepSeek formed the Harness team earlier this year with the aim of building agents that compete with AI coding services, that the company rolled out an upgraded coding assistant on the Saturday before the announcement, and that an October model update optimised DeepSeek V4.1-Flash for Harness, with a recommendation to run Harness 0.2.1-rc.1 or later.
Quartz also notes the financial backdrop: DeepSeek's latest funding round brought in roughly 100 billion yuan, about $15 billion, among the largest raises in China, as the company prepares for an initial public offering. A developer-facing agent platform with an open plugin ecosystem is a plausible thing to show investors, particularly when the alternative story about Chinese labs is that they compete on price.
Why plugin generation is the strategic move
Agent products have converged on the same bottleneck: the model can reason, but the tools it can reach define what it can actually finish. The usual answer is an integration marketplace, where a vendor ships connectors and developers write more. Creator mode inverts that by making the agent the author of its own tools, which lowers the cost of a niche workflow to roughly the cost of describing it.
That has a real effect on where the product's value sits. If a user can generate a paper reader or a video workbench in a chat, the marginal feature is cheap, and the differentiator moves to the model's ability to write correct, well-behaved software on the first few attempts, and to the runtime's ability to contain what that software does. DeepSeek's demos are chosen to show breadth; the harder test is whether generated plugins keep working after the underlying app updates its APIs, which DeepSeek concedes are still evolving.
There is also a competitive read. DeepSeek's advantage has been open weights and price. A plugin ecosystem is an attempt to build switching costs that survive a cheaper competitor, because a user's accumulated generated tools are worth something only inside the app that runs them. Developer tools that an agent wrote for you are, by construction, the stickiest kind of local customisation.
The open questions DeepSeek has not answered
The obvious one is permissions. A generated plugin with a sidebar entry that reads uploaded papers or edits video files is code running inside a desktop app with the user's files in reach, and the published material does not describe a permission prompt, a sandbox, a signing step or a review process for creator mode output. The demos also include a plugin that restyles the entire interface, which is exactly the kind of surface where a malicious or careless generated plugin could imitate trusted UI.
The second is reproducibility. DeepSeek's own note that some plugins needed several rounds of conversation implies the success rate of a single prompt is unknown, and no failure rates are published. For teams evaluating this for real work, the practical question is how often a generated plugin installs and runs correctly on the first attempt, and what happens to it when Harness changes.
The third is distribution. Harness is in public preview and open source, so anyone can run it, but a generated plugin ecosystem needs a way to share and version plugins, and DeepSeek's material describes composable plugins without describing a registry. Until that exists, creator mode is a demonstration of capability more than a platform.
None of that makes the feature trivial. A working agent that writes its own 3D renderer or video timeline from a paragraph of English is a meaningful step past chat, and DeepSeek shipped it in a preview product rather than a research paper. It just is not yet the thing the demo video implies: a reliable app factory.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover is an AI-generated editorial illustration, not a photograph or evidence of an actual event.
Our editorial standards
