DSH Is No Longer Writing Code—It's Posting on Xiaohongshu: 5 Open-Source Projects Can Assemble a Self-Media Workflow

DeepSeek Harness (hereinafter DSH) was originally a toolchain for writing code and running agents. When Xiao Mo was browsing the community on X, he found that DSH is also being used to run self-media accounts: finding trending topics, writing copy, making videos, and posting to Xiaohongshu and Douyin—people have built open-source projects for all of these. The star counts are not low; the first four projects all have over 10,000 stars.

These projects fit together like a workflow. Let's be clear first: they are basically all third-party community projects, and they do not mean that once installed, everything can be run automatically. MCP, CLI, and Skill need to be configured yourself. Account publishing involves Cookies, so it is advisable to keep manual confirmation. A high star count does not equal out-of-the-box usability. Some netizens complained in the replies that no matter how many stars a plugin repository has, after installation you still have to fill in the gaps yourself.

1. agency-agents-zh: 21.9K Stars

This repository has 328 ready-made AI expert roles, covering 22 departments such as company operations, engineering, design, marketing, product, security, and finance. It includes Xiaohongshu operations, Douyin content planning, marketing, and copywriting roles. You can migrate the relevant roles and workflows into DSH to help with account positioning, topic planning, and content creation.

Project URL: https://github.com/jnMetaCode/agency-agents-zh

The README states it plainly: each agent has its own persona, professional workflow, and deliverables, unlike generic prompt templates. The project also supports 20 AI coding tools, including Claude Code, Cursor, and Copilot. For ordinary users, you can think of it as a role library that saves the time of writing prompts from scratch.

2. xiaohongshu-mcp: 17K Stars

MCP is a protocol that lets agents call external tools. This project lets agents search Xiaohongshu notes, fetch comments, research popular content, and also supports publishing image-text posts and videos. For example, every day you can have DSH collect popular notes in a certain field and then organize promising topics to write about.

Project URL: https://github.com/xpzouying/xiaohongshu-mcp

The project documentation lists several practical details: titles should be no more than 20 characters, body text no more than 1,000 characters, and image-text posts usually get better traffic than videos and plain text. There is also an easy pitfall: the same Xiaohongshu account is not allowed to log in on multiple web clients. After logging in to MCP, do not log in again on another web client, or MCP will be kicked offline. You are responsible for publishing permissions and account risks.

3. AutoClip: 10K Stars

If you make video content, take a look at this. Give it a video link, and it can automatically edit long videos, generate subtitles, covers, and publishing copy, and adapt them for Douyin, Xiaohongshu, TikTok, Reels, YouTube Shorts, Bilibili, and YouTube. Interviews or podcasts can be split into clips suitable for short-video platforms, and it can be connected to agent workflows via CLI or MCP.

Project URL: https://github.com/zhouxiaoka/autoclip

Measured data from the source material: for Jensen's 1-hour-43-minute video, converting English to Xiaohongshu, 10 clips were generated in 7.5 minutes, with a text-model cost of about ¥0.09; for the TIM × Luo Yonghao 2-hour-52-minute video, 10 clips were generated in 29.5 minutes, at about ¥0.20. Costs are estimated based on the text-model usage of qwen-plus at the time and do not include cloud speech recognition, AI image generation, or submission service fees.

The video above is an interview-style finished clip from the AutoClip case library; the original footage and more cases are on the project's official website.

4. social-auto-upload: 16.2K Stars

Mainly solves multi-platform automatic publishing. It currently supports platforms such as Douyin, Xiaohongshu, Bilibili, WeChat Channels, Kuaishou, TikTok, and YouTube. The project provides a CLI and Agent Skill, allowing DSH to continue executing publishing and scheduled distribution tasks after content production is complete.

Project URL: https://github.com/dreammis/social-auto-upload

Again, it's best to have a human review everything before publishing. You are responsible for Cookies, account permissions, and platform risk controls.

5. dsh-media-desk: 329 Stars

The star count is not as high as the previous ones, but it is a self-media workbench developed specifically for DSH. You can directly record ideas, generate scripts, covers, accompanying images, and storyboards in the DSH sidebar, and also arrange a content calendar and retrospectives. It is equivalent to adding a work interface for content creation to DSH.

Project URL: https://github.com/woniuniuniu/dsh-media-desk

If you really want to get this set of things running, the order may be more important than the checklist. Don't rush to hand publishing over to the Agent; first use dsh-context to inspect context composition, cache hit rate, and Token consumption. Project URL: https://github.com/bowenliang123/dsh-context. Some netizens mentioned that when the cache hit rate is below 60%, extracting repeated system prompts and tool descriptions into templates is more effective than switching to a cheaper model.

If you want to use cross-session memory, you can look at MemOS. It has 12K stars, supports integration with DSH, and has a local version. Project URL: https://github.com/MemTensor/MemOS. Some netizens reminded that before launching cross-session memory, you should first whitelist sensitive paths; otherwise, long-term memory may also recall temporary keys and local paths into the context, which actually costs more Tokens.

Then use the role library to decide topics, use the Xiaohongshu MCP to find popular content, use AutoClip to make videos, use social-auto-upload to distribute, and use dsh-media-desk to schedule the content calendar. You don't have to install everything; choose according to your own content format.

In the original article, Xiao Mo said that what he himself is more interested in tinkering with is combining the Xiaohongshu MCP with multi-platform publishing tools, letting DSH handle finding trending topics, organizing topics, and generating content, and finally handing it over to the corresponding tools for publishing. This approach is relatively practical. When using DSH for self-media, the real difficulty is managing the context, materials, and publishing process. It is not hard for a model to write a few more paragraphs of copy. Content judgment still has to come from people. Star counts will change, and you should read the README's version compatibility and security notes. Don't hand your account over to an automation setup that hasn't been proven to work.