QianWen AI Skills - Setup Guide¶

Source: qianwen-ai/qianwen-ai via skills.sh - 56.2K combined installs across 11 indexed listings; snapshot 2026-10-28 (Oct 28 evening sweep)

GitHub: QianWen-AI/qianwen-ai - 115 stars, Apache-2.0, branch main, last pushed 2026-10-10; 11 SKILL.md directories under skills/<category>/<name>/, one per indexed listing, with no listing that lacks a directory

Category: Multimodal AI API skills: text, vision, media generation, account and billing operations

Quality Tier: 🟡 Beta - maintained by the QianWen-AI organization and install heavy, 115 GitHub stars, 11 indexed listings matched one to one by 11 SKILL.md directories, and two of the six sampled skills carry an auditor warning

QianWen AI Skills is a skill pack from the QianWen-AI organization that wraps the Alibaba Qwen platform in eleven agent skills. One install covers text generation and function calling, vision and OCR, image generation and editing, video generation and editing, text to speech, model selection with pricing, API key management, usage and billing queries, recharge, skill discovery, and a version checker that the other ten skills call. The repository is Apache-2.0, installs through the skills.sh CLI, and ships one SKILL.md per skill with a Chinese primary README and an English mirror.

The install numbers are high against the repository's footprint: 56.2K combined installs across eleven indexed listings against 115 GitHub stars, with the smallest listing at 2,718. It is a client for a commercial API rather than a free tool, so every skill spends against your own QianWen account, and two of the eleven reach money and credentials directly. The README advertises ten skills while the repository tree and the install index both carry eleven; the extra one is the update checker, which the README's skill table omits.

Installation¶

The README requires Node.js 18 or newer. One command installs every skill, and the picker wants a for all of them followed by Enter:

npx skills add QianWen-AI/qianwen-ai

The README's other documented route hands the job to the agent. This is the English mirror of that prompt, trimmed to the steps that matter:

Please install QianWen AI Skills for me:
1. Check if Node.js is installed, install it if not
2. Run: npx skills add QianWen-AI/qianwen-ai
3. If a skill selection prompt appears, select all skills and confirm
4. Once installed, list what skills I got and walk me through what I can do

There is no plugin marketplace or MCP route in the README, so the skills.sh CLI is the only documented installer. For a Hermes profile, the manual copy keeps the category folders out of the skills directory, since skills live at skills/<category>/<name>/SKILL.md:

git clone https://github.com/QianWen-AI/qianwen-ai.git
cp -R qianwen-ai/skills/*/qianwen-* ~/.hermes/profiles/corpusiq/skills/

Credentials come after the skills. qianwen-ops-auth loads a key from the environment first (DASHSCOPE_API_KEY, with QIANWEN_API_KEY as an alias that wins when both are set), then from a .env file in the working directory or the repository root. Keys are created in the QianWen console at https://platform.qianwenai.com/home/api-keys.

The account skills carry a second dependency that the README does not mention. qianwen-usage, qianwen-model-selector and qianwen-payment all depend on the QianWen CLI, which is a separate global install and reports its version with qianwen version:

npm install -g @qianwenai/qianwen-cli

The usage skill pins a CLI baseline of 1.3.0 and tells the agent not to call the billing and subscription groups below that version.

What It Provides¶

Skill	Installs	Use For
qianwen-update-check	5769	Comparing the installed pack version in `version.json` against the GitHub copy with semver, returning `has_update`, and rate limiting the check to once every 24 hours through `.agents/state.json`.
qianwen-image-generation	5754	Text to image and image editing on Wan and Qwen Image models: style transfer, subject consistency, text rendering inside images, and interleaved text and image output.
qianwen-vision	5743	Reading images and video with Qwen vision models: OCR, chart and table extraction, screenshot understanding, multi-image comparison, and visual reasoning.
qianwen-audio-tts	5692	Turning text into speech with Qwen TTS voices for narration, voiceover and audiobook batches. Speech recognition is explicitly out of scope.
qianwen-text	5623	Text generation, chat, code writing, reasoning and function calling on Qwen text models.
qianwen-video-generation	5591	Text to video, image to video, first and last frame, reference-based role play, and VACE video editing on Wan and HappyHorse models.
qianwen-model-selector	5568	Advisory model choice: mapping a task and a budget to a Qwen model and parameters, with pricing and cost estimation for pay-as-you-go keys.
qianwen-usage	5528	Account and spend queries through the `qianwen` CLI: auth status, free tier quota, usage summary and per-model breakdown, billing history, call logs, subscriptions and team seats.
qianwen-ops-auth	5519	API key and endpoint setup: the credential priority chain, key type detection by prefix, 401 troubleshooting, and non-plaintext key handling.
qianwen-find-skills	2724	Discovering, comparing and optionally installing published Agent Skills from QianWen SkillHubs.
qianwen-payment	2718	Balance checks, fixed-amount recharge through Alipay, recharge history, recharge page guidance, and Token Plan quotes and purchases.

All 11 indexed listings clear the 500-install threshold and appear in the table above; the tail below the threshold is empty.

Why This Matters for Hermes Agents¶

Most agent work is multimodal in practice: a screenshot to read, a PDF to summarize, a voice line to synthesize, a poster to render. Each of those normally drags in a separate provider, key, SDK and quota story. This pack puts all of them behind one DASHSCOPE_API_KEY on Alibaba's Qwen and Wan models, so a Hermes profile gets text, vision, image, video and speech from a single credential and a single failure mode to debug.

Routing is the second gain. Every SKILL.md frontmatter carries a TRIGGER and DO NOT TRIGGER block that names its neighbour skill, so an agent holding both qianwen-vision and qianwen-image-generation will not confuse reading a chart with drawing one. The blocks also state the negative case: text to speech refuses speech recognition, image generation refuses video, and vision refuses generation.

The account skills are what make unattended runs defensible. qianwen-model-selector prices a pay-as-you-go call before it happens, qianwen-usage returns free tier quota and per-model spend as JSON, and qianwen-ops-auth reports key state without ever printing the key. The corollary is to install the whole pack rather than a subset, because the update checker is a shared dependency that the other ten skills delegate to.

Usage¶

You say	What happens
"Summarize every PDF in this folder with the cheapest text model"	model-selector picks a Qwen text model that fits the budget, then text writes and runs the batch call over the folder.
"Extract the text from these 200 screenshots into one spreadsheet"	vision runs OCR through Qwen vision models, handles the multi-image comparison, and returns rows ready to write out.
"Make 20 product posters: ink wash, flat and 3D"	image-generation builds a Wan or Qwen Image batch, one call per style, with subject consistency across the set.
"Turn each of these five product shots into a short showcase clip"	video-generation runs image to video on Wan or HappyHorse, one call per asset, with an optional first and last frame.
"Read these 30 chapters aloud in a warm female voice"	audio-tts batches Qwen TTS synthesis per chapter and reports which files landed.
"How much free quota is left on qwen3.6-plus, and what did last month cost"	usage runs `qianwen usage free-tier` and the billing summary with `--format json`, then reports the parsed numbers rather than a guess.
"The account is out of balance, top it up so the batch can finish"	payment reads the balance snapshot, confirms the amount with you, and returns the Alipay recharge route from the CLI rather than constructing a URL.

Verification¶

All eleven raw paths below returned 200 on branch main when this guide was written on 2026-10-28. The layout is one directory per skill under a category folder, so probe before installing a subset:

for p in ops/qianwen-update-check image/qianwen-image-generation vision/qianwen-vision \
         audio/qianwen-audio-tts text/qianwen-text video/qianwen-video-generation \
         models/qianwen-model-selector usage/qianwen-usage ops/qianwen-ops-auth \
         meta/qianwen-find-skills ops/qianwen-payment; do
  curl -sL -o /dev/null -w "%{http_code} $p\n" \
    "https://raw.githubusercontent.com/QianWen-AI/qianwen-ai/main/skills/$p/SKILL.md"
done

Read these two before installing, because one decides how the key is handled and the other decides how money is reported:

https://raw.githubusercontent.com/QianWen-AI/qianwen-ai/main/skills/ops/qianwen-ops-auth/SKILL.md
https://raw.githubusercontent.com/QianWen-AI/qianwen-ai/main/skills/usage/qianwen-usage/SKILL.md

After install, confirm what actually landed. Expect eleven directories and a frontmatter name that matches each directory:

ls ~/.hermes/profiles/corpusiq/skills | grep -c '^qianwen-'
head -5 ~/.hermes/profiles/corpusiq/skills/qianwen-text/SKILL.md

For the account side, qianwen auth status --format json should return an authentication payload rather than a 401, and qianwen version should be 1.3.0 or newer before any billing command is trusted. The update checker can be run on demand with python3 <skill>/scripts/check_update.py --print-response.

Security¶

skills.sh verdicts for sampled skills (verified 2026-10-28) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
qianwen-update-check	Pass	Pass	Warn W011/W012
qianwen-image-generation	Pass	Pass	Pass
qianwen-vision	Pass	Pass	Pass
qianwen-audio-tts	Pass	Pass	Pass
qianwen-text	Warn	Pass	Pass
qianwen-video-generation	Pass	Pass	Pass

Two of the six sampled listings carry a warning: qianwen-update-check on Snyk with codes W011 and W012, and qianwen-text on Gen Agent Trust Hub. The other four pass all three auditors, and this snapshot carries no verdict for the remaining five listings.

The credential rule is written into the pack rather than left to the agent. qianwen-ops-auth forbids asking for a key in chat, forbids printing one, and allows writing a real key into .env only when the user insists; its status checks report set or not set, valid or invalid, and an HTTP status code, never the value. Keep that behaviour when you wire the pack into a profile: the key belongs in the environment, not in a conversation or a log.

Two skills change state on the account. qianwen-payment can create a recharge order and buy Token Plan credits, which is real money through Alipay, and qianwen-usage can log out and revoke the session server side. Both skills require confirmation before acting, so leave those two out of unattended cron runs and keep them in an interactive session where a human confirms the amount.

What the install itself runs is small by count: the clone holds 125 files, one SKILL.md per skill plus a scripts/ folder that is mostly Python helpers calling the Qwen API, and Node preflight and polling modules for the payment routes. None of it executes on clone. The risk sits in what the skills then do, which is call a paid API with your key and, for the account skills, operate the QianWen CLI against your balance.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-10-28: 56229 combined installs across 11 indexed listings; 115 GitHub stars; Apache-2.0 license; last pushed 2026-10-10. Counts drift over time.
The primary README.md is Chinese and the English mirror is README_en.md. Skill frontmatter and SKILL.md bodies are English, so an agent reads them fine while a human reviewing the repository may not.
The README advertises 10 skills while the tree and the install index carry 11. The extra is qianwen-update-check, a version checker that the other ten call, and the README's skill table does not list it. No indexed listing lacks a directory.
qianwen-payment and qianwen-ops-auth touch billing and credentials: recharge runs through Alipay, Token Plan purchases spend credits, and ops-auth reads keys from the environment and can write them into a .env file.
Every skill bills your own QianWen account. The README states that costs are charged to your account, that generated content is not guaranteed accurate, and that the project offers no availability or stability guarantee.
Node.js 18 or newer is required for the install, and qianwen-payment states 18.18 or newer for its own routes. The account skills also need the separate @qianwenai/qianwen-cli global package at 1.3.0 or newer, the update checker needs Python 3.9 plus curl, and nothing here works offline.

Pro Tips¶

Install all eleven. The update checker is what the other ten call before they run, and a partial install leaves them looking for a skill that is not there.
Put DASHSCOPE_API_KEY in the profile environment rather than in a message. Set one variable only: QIANWEN_API_KEY takes priority when both are present, which is a quiet way to run on a key you forgot you exported.
Know which key type you hold. sk-sp- is a Token Plan key with a limited model catalog that must be given an exact supported model, while sk-ws- is a standard pay-as-you-go key with the full catalog and per-token billing.
Ask for the price before the batch: model-selector estimates pay-as-you-go cost, and qianwen usage free-tier says whether quota covers the job before anything is spent.
Use --format json on every qianwen command and parse that. The table format carries ANSI codes and Unicode box drawing, and parsing it will break.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ