Agent Xiaohongshu Image-and-Text Workbench

A locally controlled agent workflow that turns Xiaohongshu account positioning, trend research, topic selection, original drafts, brand illustrations, and human-confirmed publishing into one pipeline.

Stars
★ 176
Last updated
2mo ago
License
MIT
Primary language
JavaScript

At a glance

How it runs
Desktop appSelf-hosted service
Works with
Platform-specificCodex
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Node.js >= 22.13Codex CLI (logged in)ChromeOpenCLI Browser Bridge extensionXiaohongshu login sessionShell / CLINetwork accessLocal filesystem
Typical use
Creators running multiple Xiaohongshu image-and-text accounts who need per-account positioning, brand visuals, trend caches, and storylines
Not a fit if
  • Teams that want to plug in other model APIs instead of a logged-in Codex CLI
  • Operators seeking fully automated publishing without human confirmation
  • Workflows needing video-note trend research or cross-account data sync
Maintenance
Commits in the last 6 monthslast commit 2026-09-07
Source review
79/100 · Good

What does this agent do, and when should you use it?

Agent Xiaohongshu Workbench is a local desktop application (Electron) for creators running multiple image-and-text content accounts, published under the MIT license in the EthanYoQ/agent-xiaohongshu-workbench repository. It organizes account positioning, trend research, topic selection, original drafting, brand illustration, full preview, and human confirmation into inspectable, editable, pausable steps. It does not connect to third-party model APIs: reasoning and browser operations run locally through a Codex CLI invocation from PATH (default model gpt-5.6-terra), with the OpenCLI Browser Bridge extension operating inside the user's own Chrome session. The workbench never reads, copies, or persists cookies, passwords, or platform API keys. Each content account keeps its own positioning, brand persona, visual language, trend cache, drafts, and storyline, while sharing a single research-only browser session. Trend research keeps only pure image-and-text viral notes with ≥300 likes, ≥100 saves, or ≥400 combined, and publishing accounts are unbound by default — users must manually switch accounts and acknowledge risk before the publish action is enabled.

The workbench runs this pipeline: after creating or switching a content account, the user uploads a local brand subject image or has Codex Agent's built-in image_gen capability design a brand persona from the account positioning, locks its visible traits, and generates 6 previewable series illustrations. Trend research uses the Codex Agent via the OpenCLI browser session to retrieve image-and-text viral Xiaohongshu notes per account vertical (excluding videos and low-engagement notes), cached only for the current account. It then produces exactly 5 editable topic directions, uses the built-in Lingzao Skill (.agents/skills/lingzao/) to extract transferable title, pacing, and expression mechanics without copying original text or images, and polishes a raw draft through the built-in Chinese de-AI-flavor Skill (.agents/skills/humanized-chinese-writing-polisher/) into an editable final draft. The user selects 1–6 illustrations, each previewable. Every round's output is written to a versioned local output/ directory; without a bound publishing account the flow stops at confirmation, and after binding the user chooses 'Publish now' or 'Pause and leave' (only the exactly-matched '暂存离开' button is ever clicked), with manual 'published' marking writing into that account's storyline.

  1. Creators running multiple Xiaohongshu image-and-text accounts who need per-account positioning, brand visuals, trend caches, and storylines
  2. Writers who want to convert Xiaohongshu trends into original content matching their account's voice rather than copying viral notes
  3. Teams with a brand subject (person, animal, mascot, object) who need 6 series illustrations with consistent visual identity
  4. Chinese writers sensitive to 'AI flavor' who need removal of clichés, translationese, and mechanical parallelism
  5. Risk-conscious operators who require full preview and human confirmation before publishing, with login state kept entirely in their own Chrome session

How do you install or deploy this agent?

Installer path: Windows x64 users download Agent-XHS-Workbench-Setup-*.exe from GitHub Releases; Apple Silicon users download the unsigned/un-notarized Agent-XHS-Workbench-0.2.1-arm64.dmg or .zip from the v0.2.1 Release. Installers do not include Codex CLI, Chrome, Browser Bridge, or a Xiaohongshu login session.

From source (prerequisites: Node.js >= 22.13, a logged-in Codex CLI, Chrome with your own Xiaohongshu session, the OpenCLI Browser Bridge extension):

git clone https://github.com/EthanYoQ/agent-xiaohongshu-workbench.git
cd agent-xiaohongshu-workbench
npm install
npm run setup
npm run dev

macOS local build (must run on macOS; .icns is generated by iconutil):

npm ci
npm run package:mac:arm64

Windows packaging:

npm run package:win

Verify Codex CLI before first run:

codex --version

How do you use this agent?

  1. Install and enable the OpenCLI Browser Bridge extension in Chrome, log in to Xiaohongshu with your own account, keep the Chrome profile open; run npm exec opencli -- doctor to troubleshoot if needed.
  2. After npm run dev, open the local address (default http://127.0.0.1:4173); installer users launch from the Start menu/desktop.
  3. Create a content account, enter its positioning, and start image-and-text trend research.
  4. Proceed through the flow: edit the 5 topics → review the Lingzao breakdown → generate and edit the raw and de-AI-flavored drafts → upload a brand subject master or generate a brand persona → choose 1–6 illustrations.
  5. Confirm drafts and all illustrations in full preview; before publishing, manually switch to the target publishing account in the current browser session, read the risk notice, and explicitly enable. Reasoning effort: medium for trend retrieval, illustrations, and publishing; high for Lingzao breakdown, drafts, and de-AI-flavor polish. Verification commands: npm run verify:runtime, npm test, npm run build.

What are this agent's strengths and limitations?

Pros
  • Multi-account isolation: each content account stores its own positioning, brand, trend cache, drafts, and storyline; switching accounts never shares trends — suited to matrix operations
  • Explicit originality boundaries: Lingzao breakdown transfers only structure and mechanics, never original sentences, personal stories, images, or exact layouts; reader-visible images never contain prompts or agent reasoning
  • Publishing guardrails: publishing accounts are unbound by default and require human confirmation; 'Publish now' counts as success only with a verifiable note ID/URL, and 'pause' only clicks the exact-text '暂存离开' button
  • Credentials never persisted: the workbench does not read cookies, passwords, or API keys; browser research happens in the user's own Chrome session
Limitations
  • Hard dependency on Codex CLI and its built-in image_gen capability; no third-party model API support, and image generation availability depends on the user's Codex environment
  • The macOS package is unsigned and un-notarized, triggering Gatekeeper warnings on first open; official distribution requires separate Apple Developer ID signing and notarization
  • Multiple self-prepared external dependencies: Node.js >= 22.13, Chrome, the OpenCLI Browser Bridge extension, and a Xiaohongshu login session — none ship with the installers
  • Fixed viral threshold (≥300 likes, ≥100 saves, or ≥400 combined, image-and-text only) — video or low-engagement content cannot be researched; comments are never read
  • macOS .icns icon generation depends on iconutil, so cross-building the macOS package from Windows is not possible

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Agent Xiaohongshu Image-and-Text Workbench This agent 79 · Good Desktop appFree + model costs ★ 176 2mo ago JavaScript Codex
Xiaohei Hand-Drawn Illustrations Skill 61 · Some gaps Agent plugin / skillFree + model costs ★ 13k 25d ago — Codex
md2wechat: AI-Agent CLI for WeChat Official Account Publishing 70 · Some gaps CLIFreemium ★ 3.8k 1mo ago Go Codex · Claude Code
LinkedIn Skills for Claude Code and Codex 66 · Some gaps Agent plugin / skillFree + model costs ★ 6.1k 2d ago Python Codex · Claude Code · Claude.ai

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Good
79/ 100 5-point scale 4.0 / 5
Trust 27/29
Reliability 11/14
Adaptability 14/18
Convention 13/18
Effectiveness 9/13
Verifiability 5/8
Why each dimension lost points
Trust27 / 29 · 4.7/5

Evidence shows publish binding is off by default, requires full preview plus a second in-UI confirmation; save_draft is restricted to exact-text '暂存离开' with server-side guards and tests; explicit no Cookie/password/API-key handling with runtime dirs gitignored; skills carry license/source attribution with THIRD_PARTY_NOTICES.md and skills-lock.json. Deductions: dependencies are pinned but no audit or vulnerability-mitigation statement exists, so dependency_security is 2; rollback relies on versioned output/ and safe clearing without explicit recovery guidance, rollback is 2.

Reliability11 / 14 · 3.9/5

Evidence shows the 1–6 image-count constraint is enforced in server state normalization, cross-validating README/AGENTS.md consistency claims; migration conservatively clears stale legacy data. Deductions: reliance on Codex CLI, Chrome and an extension is documented but availability/version compatibility is only runtime-checked, dependency_availability 2; error-handling examples are scattered (doctor, envelope codes) with no unified failure-message policy, failure_messages 2.

Adaptability14 / 18 · 3.9/5

Evidence shows a clearly stated target audience (multi-account graphic-content creators), staged workflow tables with executor/constraints, and concrete trigger phrasing in SKILL.md. Deductions: capability boundaries (image-gen depends on Codex environment) are mentioned but not systematic; environment constraints (unsigned macOS, no cross-packaging) are disclosed only for some scenarios, so capability_boundaries and environment_fit are 2.

Convention13 / 18 · 3.6/5

Evidence shows project structure map, prerequisites, npm scripts, license and third-party notices all present; unsigned/notarized macOS limitations are repeatedly and explicitly disclosed, so known_limitations is full. Deductions: no FAQ or worked examples (only screenshots and a flow diagram), examples_and_faq 1; version 0.2.1 and release links exist but no CHANGELOG, versioning_changelog 1; unverified individual publisher with no explicit maintenance commitment or update path, maintenance_responsibility 1.

Effectiveness9 / 13 · 3.5/5

Evidence shows per-round versioned local output/, editable raw/humanized drafts, and publish success only on a verifiable note ID. Deductions: no in-packet evidence of actual output quality or effectiveness of the virality thresholds, so marginal_value and cost_benefit are 2; setup cost is nontrivial (Codex CLI, Chrome, extension) and static evidence cannot establish net benefit.

Verifiability5 / 8 · 3.1/5

Evidence shows README/AGENTS.md/server code/test scripts cross-corroborate the virality thresholds, image-count rule and publish guards; external skill provenance is traceable via lock file and notices. Deductions: only test filenames are visible in the packet, most claims remain self-reported; the 'no copying' originality boundary is prompt-enforced and not statically provable, so claim_traceability and cross_source_corroboration are 2; documentation largely separates claims from inference (explicitly disclaiming platform data), fact_inference_separation 2.

Risks and how to mitigate them
  • macOS installers are unsigned and unnotarized; users accept Gatekeeper risk and should verify hashes themselves.
  • Publishing drives the user's real logged-in browser session, with platform risk-control and misfire risk; confirm every preview and second confirmation.
  • The 'no copying' originality boundary for viral-note breakdowns is prompt-enforced and not statically provable; human review of originality is still needed.
  • Publisher is an unverified individual maintainer with no stated maintenance commitment; monitor security advisories for pinned dependencies yourself.
  • Only test filenames are visible in the packet; test coverage and quality are not independently verified.
Evidence confidence: Low Reviewed Nov 21, 2026 Reviewed revision 6e58278984cb
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/agent-xiaohongshu-workbench.svg)](https://followagents.com/en/agents/agent-xiaohongshu-workbench)
<a href="https://followagents.com/en/agents/agent-xiaohongshu-workbench"><img src="https://followagents.com/badge/agent-xiaohongshu-workbench.svg" alt="FARS review"></a>

FAQ

Do I need to pay or configure a model API key?
The software itself is MIT-licensed and free. It does not connect to third-party model APIs; it invokes a logged-in Codex CLI from PATH (default model gpt-5.6-terra), so inference and image-generation costs depend on your Codex environment. No model API key is configured inside this repository.
Where are my Xiaohongshu credentials and data stored?
Login state stays entirely in your Chrome session; the workbench never reads, copies, or persists cookies, passwords, or platform API keys. Runtime data (.data/, output/, public/generated/, etc.) lives locally and is Git-ignored.
Can it publish notes automatically?
Not unattended. Publishing accounts are unbound by default; you must manually switch to the target account in the current browser session, read the risk notice, and explicitly enable. Every publish requires draft and illustration preview plus on-screen confirmation, and success requires a verifiable note ID or URL.
What content does trend research cover?
Only pure image-and-text notes — videos, mixed media, unverifiable media types, and notes missing engagement metrics are excluded. The viral threshold is ≥300 likes, ≥100 saves, or ≥400 combined; comment counts alone never qualify, and comment content is not read.
Which operating systems are supported?
There are installers for Windows x64 and macOS Apple Silicon (M-series, unsigned/un-notarized); macOS local builds must run on macOS. No Linux installer is documented.
Is Agent Xiaohongshu Image-and-Text Workbench open source?
Yes. The source code is on GitHub (EthanYoQ/agent-xiaohongshu-workbench) under the MIT licence.
Is Agent Xiaohongshu Image-and-Text Workbench safe to use?
FollowAgents' FARS review (Nov 21, 2026) scored Agent Xiaohongshu Image-and-Text Workbench 27/29 on trust and safety. None of the eight safety controls the review checks for was missing outright.
Is Agent Xiaohongshu Image-and-Text Workbench still maintained?
Development has slowed: the latest commit on GitHub is from 2026-09-07, within the last six months but not the last 30 days.
Who makes Agent Xiaohongshu Image-and-Text Workbench?
Agent Xiaohongshu Image-and-Text Workbench is published on GitHub by the account EthanYoQ. FollowAgents has not verified which person or company is behind that account.
Was this page helpful?
View on GitHub ↗ Install ↓

Host ecosystems

Related concepts