mrkeyoor.com_
Tue 03 Nov 02:45 UTC
AI Toolsevaluationupdated 29 Sept 2026

xiaohongshu-mcp review

The primary documentation is in Chinese, with a separate English README available. xiaohongshu-mcp gives an MCP client logged-in access to Xiaohongshu for searching, reading posts, publishing media, and handling interactions through browser automation.

+140stars / 7d
Verdict

Our run built xiaohongshu-mcp in 40 seconds and passed all 16 tests, so the code is easier to verify than the browser account it controls. Use it for a supervised Xiaohongshu workflow where search, publishing, and replies belong in one MCP service. Skip it if a broken selector, expired cookie, or wrong-target reply would create an unacceptable publishing incident.

We ran it

Lab card: what happened when we ran xiaohongshu-mcpScreenshot of xiaohongshu-mcp (www.haha.ai/xiaohongshu-mcp)
Install✓ · 19s67 packages
Build✓ · 40s
Tests✓ · 12s16 passed · 0 failed of 16 (go test)
Repo153 files~13,578 lines of source · 83.8 MB · 5 CI workflows · Dockerfile

Answers from our run

Does xiaohongshu-mcp build from source?

Dependencies installed in 19 seconds (67 packages), and the build succeeded in 40 seconds. We cloned commit a5c8f77 into a clean Debian container with 3 CPUs and no project-specific setup.

Do xiaohongshu-mcp's tests pass?

Yes. The project's own test command finished in 12 seconds and reported 16 passed, 0 failed (go test).

Who should not use xiaohongshu-mcp?

Teams that require an official Xiaohongshu API contract: the project drives the website, and open issue 860 documents reply targeting and emoji confirmation errors after page behavior changed.

What are the alternatives to xiaohongshu-mcp?

MediaCrawler, Browser Use, Firecrawl MCP Server. Our run built xiaohongshu-mcp in 40 seconds and passed all 16 tests, so the code is easier to verify than the browser account it controls.

Setup3/5Build is quick; login, browser state, and media mounts need care
Docs4/5Detailed Chinese guide plus English README and deployment paths
Community5/516,038 stars and active issue and pull request work in September
Maturity3/5v2.5.5 is active, but website changes still break browser actions

Who it’s for

Developers who want a local MCP bridge between an AI client and a Xiaohongshu account.
Content teams willing to keep a dedicated browser session and review posts before they go live.
Go developers who need HTTP endpoints as well as MCP tools.
Claude Code users who are comfortable operating an authenticated local service.

Who it’s NOT for

Teams that require an official Xiaohongshu API contract: the project drives the website, and open issue 860 documents reply targeting and emoji confirmation errors after page behavior changed.
Operators who cannot dedicate one web login to the service: the README warns that signing into the same account in another web browser invalidates the MCP session.
Mac Intel or Linux ARM64 users expecting a supplied binary: the v2.5.5 release lists neither platform.
Anyone planning unsupervised mass posting: the README warns against redirected traffic and copied content, while account enforcement remains outside the project's control.
English-only teams that need the full troubleshooting trail in English: an English README exists, but the main guide and many issue reports are Chinese.

Setup reality

Our sandbox installed 67 Go packages in 19 seconds at commit a5c8f77. The build passed in 40 seconds, and all 16 tests passed in 12 seconds. The checkout held 153 files, about 13,578 source lines, and used 83.8 MB.

A usable session still needs a Xiaohongshu login and a saved cookie. The first binary run downloads a headless browser of about 150 MB. Docker mounts cookie data and media directories, while remote deployments should set the optional AUTH_TOKEN because access control is off by default.

Released binaries cover macOS Apple Silicon, Windows x64, and Linux x64. Browser state is the bigger operational constraint: another web login can evict this session, and local video publishing requires an absolute file path.

The 11 tools turn a Xiaohongshu session into an agent interface

xiaohongshu-mcp wraps a logged-in browser rather than an official platform API. Its tools can check login state, publish image or video posts, search, read the recommended feed, fetch post details and comments, inspect profiles, reply, like, and save posts. That is a useful set for an assistant doing research and account work. It also means every useful action inherits the state and fragility of the website underneath it.

The service exposes MCP over HTTP on port 18060, and the README includes setup for Claude Code, Cursor, and other clients. There are plain HTTP endpoints too, which helps when an existing workflow does not speak MCP. Local image paths are recommended for publishing, while video input must be a local absolute path. A remote container therefore needs deliberate media mounts, not only a network route and a model client.

A 19-second install does not include the account session

Our fresh Debian sandbox installed 67 Go packages in 19 seconds at commit a5c8f77. The repository had 153 files, roughly 13,578 lines of source, and occupied 83.8 MB. That is modest for a service with browser automation, an HTTP layer, and MCP handlers. Five CI workflow files and a Dockerfile also give maintainers visible places to inspect the build path.

Getting a useful result takes more than compiling. The first run downloads a headless browser of about 150 MB, then a person must complete the Xiaohongshu login flow and preserve the resulting cookie file. Docker Compose mounts a data directory for cookies and another for images. If you expose port 18060 beyond localhost, set AUTH_TOKEN through the environment. Authentication is disabled unless you configure it.

What happened when we ran it

Our run built the project in 40 seconds after the 19-second install. Go test then finished in 12 seconds with 16 passed and 0 failed. We used an unprivileged container with 3 CPUs, 8 GB of RAM, Go 1.24 on Debian, and no secrets. The successful suite shows that commit a5c8f77 was internally consistent under those conditions.

We did not sign into Xiaohongshu, download the browser, publish a post, or exercise live search. The 16 tests cannot establish that current page selectors match the live site or that an account will remain authenticated. They also are not a speed test for browsing or uploads. The result is still useful: installation, compilation, and the supplied test path were green before any account-specific work began.

One web login must remain dedicated to the service

The README warns that a Xiaohongshu account cannot stay signed into multiple web sessions. Logging into another browser can kick out the MCP session, although using the mobile app is allowed. In practice, this asks for a dedicated account routine: one persisted browser profile, a known cookie location, and a human who can repeat the login when the cookie expires. Treat that state as an operating dependency.

Release v2.5.5 supplies binaries for macOS on Apple Silicon, Windows x64, and Linux x64. Mac Intel and Linux ARM64 are not included. Source builds and Docker cover some of that gap, but the browser still needs compatible system libraries and writable data storage. The README points Docker users to a long troubleshooting issue, which is more candid than pretending every headless environment behaves alike.

Website changes can break one action while the service stays healthy

Recent issues show the failure mode clearly. Issue 813 reported search requests timing out after roughly 60 seconds while the recommended feed still returned, and issue 839 proposed waiting for search data instead of a quiet DOM. Release v2.5.5 includes later search fixes. A green health check or valid login therefore does not prove that every individual action works against the current page.

Issue 860 is more serious for publishing operations. It reports replies falling back to another comment by the same user and emoji-containing replies being reported as unconfirmed. Two fix pull requests were active on September 27, and the repository was pushed on September 28. That activity is reassuring, but an active repair queue does not make reply automation safe without preview, audit logs, and a way to stop the agent.

The project is active, but its contract belongs to the website

GitHub showed 16,038 stars and 102 open issues and pull requests on September 29, 2026. The last release, v2.5.5, shipped on September 22, and the last push followed six days later. Those dates, plus current fixes for search, login, and comments, show an active project dealing with active breakage rather than an abandoned wrapper.

The Apache-2.0 license is friendly for internal integration, and the split between MCP and HTTP makes the service adaptable. The hard limit is external: browser automation cannot promise the stability of a supported API. Our 16 passing tests make xiaohongshu-mcp a credible project to trial. They do not remove the need for one-account session discipline and human approval before any post or reply reaches a public profile.

Alternatives

ProjectWhat it isPick it when
MediaCrawler gh↗A multi-platform crawler for public social content, including Xiaohongshu.pick this instead when collection across several Chinese platforms matters more than MCP-based publishing and interaction.
Browser Use gh↗A general browser automation framework built for AI agents.pick this instead when you need custom browser workflows across many sites rather than Xiaohongshu-specific tools.
Firecrawl MCP Server gh↗An MCP server for web search, crawling, and structured extraction.pick this instead when public web research is the job and you do not need a logged-in Xiaohongshu session.

Sources

  1. xiaohongshu-mcp README
  2. Release v2.5.5
  3. Issue 813: search timeout
  4. Issue 860: reply targeting and emoji confirmation

More ai tools reviews

kornia · skills · anarlog · pr-agent · agent-sprite-forge · KiroCrew · the whole board →