Workflow & Core Tools · daymade/claude-code-skills

llm-icon-finder

Name: llm-icon-finder
Author: daymade/claude-code-skills

Finding and accessing AI/LLM model brand icons from lobe-icons library. Use when users need icon URLs, want to download brand logos for AI models/providers/applications (Claude, GPT, Gemini, etc.), or request icons in SVG/PNG/WEBP formats.

L3 Networked · Declare domain & key scopeV1 Commit-PinnedNetwork AccessWrite FileDual-pass Audit

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Extremely simple implementation principle: The skill neither bundles nor mirrors any icon files; its entire capability boils down to string concatenation of "fixed CDN path template + icon name", allowing the agent to generate URLs on the fly for users.

llm-icon-finder/SKILL.md

# PNG https://raw.githubusercontent.com/lobehub/lobe-icons/refs/heads/master/packages/static-png/{light|dark}/{icon-name}.png

What is happening here: SKILL.md uses 4 URL templates to solidify the static directory conventions of the lobe-icons repository (format segment static-svg|static-png|static-webp + theme segment light|dark + icon name), meaning this directory contains only 3 markdown files with zero binary assets.

The default output approach is "providing links rather than files": Unless the user explicitly asks to download, it only returns URLs; downloads are handled by having the host Bash execute curl.

llm-icon-finder/SKILL.md

1. Identify icon name (usually lowercase company/model name, hyphenated if multi-word) 2. Determine format (default: PNG) and theme (default: dark) 3. Construct CDN URL using pattern above 4. Provide URL to user 5. If download requested, use Bash tool with curl 6. Include web viewer link: `https://lobehub.com/icons/{icon-name}`

The six-step workflow sets three default values: PNG format by default, dark theme by default, and lobehub.com/icons/{name} web viewer links attached whether downloaded or not.

Naming conventions are the sole knowledge ontology: lowercase + hyphen, distinguishing multiple variants via suffixes (-color for colored, -text for text versions), with Chinese names mapped to English icon names first.

llm-icon-finder/SKILL.md

**Icon naming convention**: Lowercase, hyphenated (e.g., `claude`, `chatglm`, `openai`, `huggingface`)

Above the same section, there is also 'Color variant (append -color to icon-name)'; testing confirmed that -color.png / -text.png variants do exist on the CDN (claude-color.png returned 200), showing the suffix convention matches the upstream directory.

Chinese queries rely on a built-in mapping table (e.g., 智谱 → chatglm, 月之暗面 → moonshot) and proactively provide related icons from the same family.

llm-icon-finder/SKILL.md

**Chinese AI models**: Support Chinese queries (e.g., "智谱" → `chatglm`, "月之暗面" → `moonshot`)

In the examples section, the same mapping appears again ('User: "找一下智谱的图标" → Identify: 智谱 = ChatGLM → icon name: chatglm', alongside hints to mention related icons like zhipu, codegeex)—an approach that embeds domain vocabulary directly into the prompt.

Common icon rosters are externalized to references/icons-list.md, grouped into four categories: Models / Providers / Applications / Chinese AI, leaving SKILL.md purely for routing.

llm-icon-finder/SKILL.md

**Common icons**: See `references/icons-list.md` for comprehensive list organized by category (Models, Providers, Applications, Chinese AI)

The list contains 57 lines and 44 deduplicated icon names (Models 12 + Providers 17 + Applications 11 + Chinese AI table 17 lines; the Chinese table introduces 4 new names not seen in the first three categories: ai360 / aimass / aihubmix / codegeex). It is the only data file outside SKILL.md and serves as the candidate pool for 404 troubleshooting.

404 troubleshooting is a three-step procedure rather than retries: try the -color suffix → try aliases (company vs product, such as chatgpt/gpt, google/gemini) → guide the user to browse the gallery or search the repository.

llm-icon-finder/SKILL.md

If URL returns 404: 1. Try `-color` suffix variant 2. Check alternate naming (e.g., `chatgpt` vs `gpt`, `google` vs `gemini`) 3. Direct user to https://lobehub.com/icons to browse 4. Search repository: https://github.com/lobehub/lobe-icons

The ironic part: The aliases chatgpt/gpt provided in the troubleshooting steps actually do not exist on the CDN (see discrepancy item in verification.second_pass); candidate names in this step carry risks of obsolescence.

A developer integration path is listed separately in reference: npm package names (@lobehub/icons along with its static-svg/png/webp, rn variants) and React component usage.

llm-icon-finder/references/developer-info.md

Icons can be installed as npm packages:

This path targets users who want to "install icons into their own projects", serving as an independent route from the primary workflow (URL concatenation); all 5 npm packages were verified to exist on the registry (HTTP 200).

The skill itself contains no validation/fallback code: Whether a URL is accessible can only be discovered after the fact if the user or agent runs curl and gets a 404; the skill contains no probing logic.

llm-icon-finder/SKILL.md

- Browse https://lobehub.com/icons - Try variations (e.g., company name vs product name: `alibaba` vs `alibabacloud`) - Check for `-color` variants if standard URL fails

The 'Uncertain names' section leaves uncertainty handling entirely to human or model judgment without any catalog validation scripts—the main difference compared to similar skills that provide built-in catalog validation (such as hyperframes registry-type skills).

2Core Capabilities

01Generate directly referenceable CDN URLs based on icon names (SVG/PNG/WEBP × light/dark, plus -color variants)

02Built-in roster of 44 unique common AI/LLM icon names (57 lines, organized across Models / Providers / Applications / Chinese AI)

03Chinese name to icon name mapping (智谱 → chatglm, 月之暗面 → moonshot, 零一万物 → zeroone/yi, 通义千问 → qwen, etc.)

04Controlled download: Host Bash executes curl -o to save files locally only when requested by the user (examples include openai.svg / anthropic.svg)

05404 troubleshooting guidance: variant suffix → alias candidate → official gallery/repo search

06Developer integration guidance: 5 npm packages (React / static-svg / static-png / static-webp / React Native) with React component examples

07Always attaches a web viewer link to facilitate manual visual confirmation of icons

3External Dependencies

Kind	Dependency / Endpoint
network	lobe-icons 静态资源 CDN（PNG，实测 200）
network	lobe-icons 静态资源 CDN（WEBP，实测 200）
network	lobe-icons 静态资源 CDN（SVG；文档写的 {light\|dark} 路径实测 404，实际位于 static-svg/icons/）
cli	curl（宿主 Bash 执行的下载动作）
network	图标画廊（人工浏览/确认外观）
network	上游仓库（排障检索入口）
package	@lobehub/icons（React 组件包；实测 npm registry 200）
package	@lobehub/icons-static-svg / @lobehub/icons-static-png / @lobehub/icons-static-webp / @lobehub/icons-rn（实测 4 个包 registry 均 200）
network	开发者文档站（图标库文档）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Declare allowed domains and key scope

For Administrators

Least privilege keys, retain request logs

Evidence & Triggers (File Paths & Excerpts)

网络端点: Declared external network access: https://raw.githubusercontent.com/lobehub/lobe-icons/refs/heads/master/pack
本地工作目录: File I/O and artifact output: Only when explicitly requested by user: host Bash runs `curl -o <icon>.<ext>` to write to agent/user-specified path; SKILL.md specifies no fixed path

Audit Notes: Evaluated against Tier 3 of the 6-tier paradigm ('Routine network invocation via public API/official CLI with predictable targets'): The skill involves no script execution, credential access, or local disk writing (unless the user explicitly requests a download). Its sole external action is reading from the fixed public endpoint raw.githubusercontent.com/lobehub/lobe-icons with a static path prefix and no outbound user data—avoiding credentials or remote package execution, so it is not Orange. However, it does initiate network requests and guides local file saving, introducing a runtime dependency beyond pure-text skills, which precludes a Green rating. The real concerns lie in reliability rather than security: 1) SKILL.md's SVG path template (packages/static-svg/{light|dark}/) conflicts with current upstream structure (verified 404; upstream is static-svg/icons/), leading to broken links if copied directly; 2) Four names in the roster (chatgpt / gpt / llama / lobechat) do not exist on the CDN (verified 404; upstream names are openai / meta / lobehub); 3) URLs track upstream master without pinned versions, meaning upstream directory refactoring can instantly invalidate this skill's documentation (which was observed during this evaluation).

5Second-Pass Verification

[ok] PNG / WEBP 端点可用性（skill 的主路径） — curl 实测：packages/static-png/dark/claude.png = 200、packages/static-png/dark/claude-color.png = 200、packages/static-png/light/claude.png = 200、packages/static-webp/dark/claude.webp = 200，不存在的名字（notarealicon123.png）返回 404 —— 主路径与文档模板一致，404 也可正常反映为 404。
[discrepancy] SVG 路径模板 vs 上游真实布局 — SKILL.md 写 `packages/static-svg/{light|dark}/{icon-name}.svg`，实测 packages/static-svg/dark/claude.svg = 404、.../light/claude.svg = 404、.../dark/claude-color.svg = 404；GitHub Contents API 显示 packages/static-svg 下只有 icons/ 子目录（无 light/dark 段），tree API 中 static-svg 路径 911 条、形如 packages/static-svg/icons/claude.svg，curl 实测 200。即 SVG 分支的模板已过期（PNG/WEBP 仍是 light|dark 布局，两条路径规则已不一致）。
[discrepancy] icons-list.md 名录逐名验证 — 把名录去重后的 44 个图标名逐个 GET static-png/dark/<name>.png：40 个 200，4 个 404 —— `chatgpt`、`gpt`、`llama`、`lobechat`；这 4 个名字在 static-svg/icons 的仓库树里同样不存在（tree API 过滤结果为空）。可用替代名实测均 200：openai（对应 chatgpt/gpt）、meta（对应 llama）、lobehub（对应 lobechat）。即名录是『上游旧名 + 别名』的混合体，需要 agent 现场纠偏（讽刺的是 SKILL.md 的 404 排障恰恰建议试 chatgpt/gpt 这两个已失效的名字）。
[ok] npm 包名真实性（developer-info.md） — curl registry.npmjs.org/<pkg>/latest：@lobehub/icons、@lobehub/icons-static-svg、@lobehub/icons-static-png、@lobehub/icons-static-webp、@lobehub/icons-rn 五个均返回 200，包名与用途注释对应无误；这些是给用户在自己项目里安装的指引，skill 运行期不会自动安装。
[ok] 画廊与仓库入口可达性 — https://lobehub.com/icons 返回 302（跳转，属正常站点行为）；https://github.com/lobehub/lobe-icons 为公开仓库且 tree API 可读（master 分支可访问，packages/ 下有 react-native、static-avatar、static-png、static-svg、static-webp 五个包）。文档给的四个外链（画廊/仓库/文档/npm 搜索）均非死链。
[ok] 功能声明 vs 实际能力（夸大检查） — SKILL.md/README 声称 '100+ icons'，实际上游规模远大：单 packages/static-png/dark 目录即含数百个 PNG（含 -color/-text 变体），static-svg/icons 有 911 个 svg 路径，且涵盖名录之外的品牌（如 adobe、aws、cursor、vercel）——声明属保守而非夸大。反面：declared 'Comprehensive' 的名录只 44 名、且含 4 个失效名，'comprehensive' 一词偏乐观。
[ok] 『支持中文查询』的映射来源 — SKILL.md 的 '智谱 → chatglm、月之暗面 → moonshot' 与 icons-list.md 的中文列表格（| `chatglm` | 智谱清言 | ChatGLM |、| `moonshot` | 月之暗面 | Moonshot (Kimi) | 等 17 行）一致，映射有仓内依据；但『智谱=ChatGLM』这类厂商别名词典属领域知识，仓内无法证真伪，只能确认两侧文本自洽。
[discrepancy] 下载示例可照抄性 — SKILL.md 的 Multi-icons 示例里 URL 用了省略号：`curl -o openai.svg "https://raw.githubusercontent.com/lobehub/lobe-icons/.../dark/openai.svg"`，and 完整可用的路径需要补 `packages/static-png`（SVG 情况还要改成 static-svg/icons）。照字面复制该示例会得到无效/404 请求；同一文档上方的模板段才是可执行形式。

6Verdict & Scenarios

Ultra-low cost with zero dependencies: 3 markdown files are all it takes; directly readable by any host without package installation or attached scripts.

Defaults to providing links rather than saving files, making it the cleanest integration method for READMEs/docs/demos while avoiding clutter in user directories.

Comprehensive format dimensions: three formats × light/dark × -color variants, with a single template covering common referencing needs.

44-item common roster + Chinese mapping covering mainstream models/vendors/applications and domestic vendors, allowing Chinese queries to resolve directly to icon names.

Complete developer path: 5 verified npm packages + React usage examples, facilitating installing icons directly into engineering projects instead of merely linking them.

Best for: Suitable for: 1) Those writing READMEs, docs, presentations, or landing pages who want fast direct links to brand icons like Claude, OpenAI, and Gemini; 2) Teams wanting to reference icons without bloating repositories (PNG/WEBP direct links are most reliable); 3) Chinese-speaking users needing Chinese name search (智谱, 月之暗面, 通义千问); 4) Developers looking to use React components rather than image links in frontend projects (via developer-info.md npm path).

Not for: Not suitable for: 1) Scenarios requiring icons to have permanently stable, break-proof URLs—upstream uses master branch links, and tests show the SVG template is currently broken with 4 listed names returning 404; 2) Offline/intranet environments (all icons reside on raw.githubusercontent.com); 3) Scenarios needing icon metadata, licensing information, or brand usage terms—this skill only gives direct image URLs without licensing/compliance details; 4) Those expecting the skill to automatically validate whether an icon name exists (it performs no active probing; 404s must be discovered manually).

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload llm-icon-finder.tar.gz

sha256: 8c8f311d4974d206…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit d5c4678cb5.