DeepSeek adds Creator Mode, lets users build agent plugins by chat
What's the deal? DeepSeek has added a Creator Mode to its open-source Harness agent app, letting users build working plugins from plain-language requests. The Hangzhou-based company announced the feature on November 2, demonstrating plugins such as a design workspace, a sandbox game, a 3D model renderer, and a video editing workbench — all created from prompts typed in everyday English.
How it works: A harness connects large language models to the tools and systems that let them act, turning a model into an agent. DeepSeek Harness is built on an "everything is a plugin" design, so tools, skills, and parts of the interface can be swapped. Creator Mode lets the agent write those plugins itself.
Why now? The launch pushes DeepSeek deeper into the agent market, days after it showed a Super-agent coding prototype. It also follows rivals: OpenAI has expanded its Agents SDK with a harness and sandbox support, while Anthropic is adding plugins and integrations to Claude's agent products.
The details: Harness is in public preview worldwide and published under the MIT license. It runs as a desktop app for Apple silicon Macs and 64-bit Windows, or as a local web interface. On GitHub, the project has passed 269,000 stars since its creation in August.
What could go wrong? DeepSeek is clear that Harness is experimental. Its GitHub safety notice says the software has not had a security audit, can run model-generated code and commands, and may damage files or leak credentials if a plugin or prompt goes wrong.
The company recommends running it with the least privileges possible, ideally in a disposable virtual machine, and reviewing plugins before allowing them to run. DeepSeek also warns that results vary, noting "some plugins were refined through several rounds of conversation, so a single prompt may not reproduce exactly what you see in the video."
The signal: Agent infrastructure is moving toward configuration by conversation rather than code. Creator Mode makes Harness more adaptable, but the announcement alone says little about reliability or adoption — the test is whether users trust it enough to run model-written code on real systems.
Read more: Bloomberg
Image credit: Generated with Gemini
