ByteDance has introduced the Coze desktop client, designed for work scenarios. New features include cloud storage and a deep screenshot function, allowing agents to interact directly with the local computer.
In addition to simplifying the reading and writing of local files, the system automatically recognizes Codex CLI and Claude Code integrations as agents, significantly improving collaborative work among multiple agents. The client integrates the recently launched GLM 5.3 API and supports video models such as Seedance 2.5 and MiniMax H3.
The author demonstrated using Coze Desktop to create all promotional material for a new book in a single session: the agent read the manuscript outline on the local device, generated a website, a PPT presentation, and a promotional podcast. All created files are automatically saved to Coze's cloud storage and synchronized in real time between the computer, phone, and web interface. The agent was able to work with the latest versions of the files, bypassing manual uploading and downloading.
The deep screenshot feature, activated in settings, allows a Mac user to capture the current window and readable text by pressing the Option key (left or right), which is then placed into the input field. This screenshot contains not only an image but also the structural text of the interface, which is useful when debugging web pages.
The built-in PPT skill, invoked by the command /slash-ppt, enables the agent to create presentations from both local and online materials. Although images are artificially generated, they can be edited: double-clicking allows elements to be deleted or text to be colored using Seedream's image editing capabilities. A podcast skill is also available, which creates audio material with a natural voice, allowing the agent to override Coze's standard voice.
After installation, Coze Desktop adds Coze cloud storage to the computer, which functions almost like a local hard drive. All files created during dialogues with cloud agents are saved there. Users can drag local files into this storage so that cloud agents can read them, and all changes are synchronized with the phone. Two-way synchronization occurs: any changes on the local machine are reflected in the cloud, and vice versa, eliminating the need for multiple uploads and downloads; the trash keeps files for 30 days.
For working with personal media, Coze provides official skills, including searching for prohibited words via a partner API, enabling the agent to find relevant content, write articles, and publish blog drafts with a single instruction. Additionally, there is a research assistant template for analyzing scientific papers, and thanks to Seedance 2.5 and MiniMax H3, the user can create derivative videos from a single still frame of a film.