Codex
RedNote, Douyin and TikTok MCP: let your AI agent collect social data through your own browser
The MediaClaw MCP server is a standard MCP (STDIO) server with 40+ tools. It collects RedNote (Xiaohongshu), Douyin and TikTok posts, videos, comments, profiles and search results, then extracts video transcripts and image text. Every page visit goes through the MediaClaw extension in your own browser, using the account you are already logged into, so the agent never sees your cookies. Eight agent hosts are supported today, including Codex, Claude Code and Cursor.
What tools the MCP server provides
40+ tools in five groups. The agent reuses data you have already collected before it starts anything new.
Collect
Start from a link, keyword, profile or search and get the raw data.
- Post and video details, batch link collection
- Comments and replies
- Creator profiles and their posts
- Keyword search results and account scans
Extract
Turn videos and images into text the agent can read.
- Timestamped video transcripts (confirmed before running)
- Text extracted from post images
- Post body and engagement metrics
Research
Ready-made research actions instead of prompts written from scratch.
- Keyword topics and long-tail expansion
- Benchmark account discovery and hit analysis
- Single-post breakdowns
Read saved data
Anything already stored in the extension is queryable.
- Local collection records and data pool queries
- Saved analyses and style profiles
- Task progress and connection status
Comment outreach
Message users picked from collected comments, only after you confirm the final send.
- Prepare recipients and message copy
- Set send intervals and a daily cap
- Stop remaining sends at any time
Which agents are supported
All three hosts use the same package, the same MCP tools and the same content research skill. Only installation and updates differ.
| Host | Install | Auto-update | Guide |
|---|---|---|---|
| Codex | Paste the install prompt from the extension; Codex installs everything | Codex integration | |
| WorkBuddy | Same as Codex, prompt-driven install | WorkBuddy integration | |
| Doubao for Work | Create a STDIO connector by hand; the agent prepares the values | Doubao for Work integration | |
| Claude Code | Same as Codex; the agent installs it with Claude Code's plugin commands, then you open a new session | Setup guide | |
| Cursor | Prompt-driven; the agent adds MediaClaw to Cursor's MCP config, shared by the editor and cursor-agent CLI | Setup guide | |
| Grok Build | Prompt-driven; the agent adds a local MCP connector to the Grok config | Setup guide | |
| Grok Bot (cloud computer) | The extension and agent both run on Grok's cloud computer; it approves pairing for you, you only log in to the platform | Setup guide | |
| DeepSeek Harness | Prompt-driven; the agent adds a stdio connection to the current profile | Setup guide |
WorkBuddy
Doubao for Work
Claude Code
Cursor
Grok Build
Grok Bot (cloud computer)
DeepSeek Harness
MediaClaw MCP vs. open-source RedNote MCP vs. writing your own scraper
Each approach fits a different job. This table only covers getting the data, based on each project's public documentation.
| MediaClaw MCP | Open-source xiaohongshu-mcp | Own scraper / MediaCrawler etc. | |
|---|---|---|---|
| Platforms | RedNote, Douyin, TikTok | RedNote | Depends; MediaCrawler covers RedNote, Douyin, Kuaishou, Bilibili and more |
| Login | Your everyday Chrome session; the agent never gets cookies | Scan a QR code in its own headless browser; cookies saved locally | QR code, cookie or phone login saved by the script |
| Setup | Browser extension plus one prompt to your agent | Download a binary or run Docker, then configure the MCP client | Python environment, Playwright and other dependencies; coding required |
| Publish posts / like | No, collection, extraction and analysis only | Yes, publishes image and video posts, comments and likes | Usually no |
| Transcripts and analysis methods | Built-in video transcripts, image text extraction and 12 research methods | Not included | Not included; wire up your own |
| Cost | Free extension; agent mode needs a membership, transcripts use credits | Free and open source (Apache-2.0) | MediaCrawler states learning and research use only; a paid Pro edition exists |
Platforms
Login
Setup
Publish posts / like
Transcripts and analysis methods
Cost
Five things the agent must ask you first
Approving the pairing only lets the agent connect. It does not approve every future collection.
| Action | What you see first | After you confirm |
|---|---|---|
| Collect an account | A plan: target, fields, scope, batches and risk level | Runs only the batches in the plan |
| Video transcripts | Which videos it will transcribe and how long each one is | Extraction starts; items already extracted are not extracted again |
| Batch detail fill | The scope to fill | Opens detail pages to collect |
| Delete data | Impact: how many records, comments and transcripts | Deletes |
| Comment outreach | Full recipient list, sending account and message | Sends one by one |
Collect an account
Video transcripts
Batch detail fill
Delete data
Comment outreach
FAQ
Yes. The MediaClaw MCP server covers RedNote, Douyin and TikTok and works with Codex, Claude Code, Cursor, WorkBuddy, Doubao for Work, Grok Build, Grok Bot and DeepSeek Harness. On Codex there is no config file to write: copy the install prompt from the extension (System settings → Agent access), send it to Codex, and approve the pairing once in the extension.
A scraper logs in and pages through results on its own, so you maintain its login state and runtime and get raw data back. With MediaClaw the agent views pages through the extension in your own browser and account, then can go on to transcripts, comment analysis and topic ideas. Before any new collection it shows you a plan and waits for your approval.
Yes for Claude Code and Cursor. Eight hosts are supported: Codex, Claude Code, Cursor, WorkBuddy, Doubao for Work, Grok Build, Grok Bot (cloud computer) and DeepSeek Harness, each with its own install and pairing flow. Other MCP clients are not supported yet; new hosts will be announced here and in the changelog.
Yes. The MCP server cannot run on its own. Every page visit happens in the MediaClaw extension in your browser, and the agent can only use capabilities you approved there.
Collection uses normal page visits in your own browser, and the extension paces them. Before large jobs the agent states the batches and risk level and waits for you. No tool can guarantee zero risk, so follow platform rules and avoid long bursts of continuous collection.
No. The MediaClaw MCP server only collects, extracts and analyzes. It does not publish posts, like or auto-comment. For automated publishing, look at open-source projects such as xiaohongshu-mcp.
The browser extension is free. Agent access needs an active membership, and paid features such as video transcripts use credits with a quote before running. Analysis and writing on the agent side do not use MediaClaw credits.
Give your agent real RedNote, Douyin and TikTok data
Install the free extension, send the install prompt to your agent, and it does the rest.
Updated