RankClaw — AI Agent Trust Layer
RankClaw is the trust and safety layer for the OpenClaw ecosystem and all AI agent tools.
Before you install any skill — OpenClaw, nanoclaw, picoclaw, or MCP server — RankClaw tells you:
- - Is it malicious? (1,046 confirmed malicious skills and counting)
- What does the AI security audit say? (deep scan of SKILL.md, prerequisites, install behavior)
- What's the trust score? (0–100 across Security, Maintenance, Documentation, Community)
- Is the author legitimate? (detect brand-jacking, impersonation, phantom prereqs)
14,700+ ClawHub skills indexed. Free for agents. No API key required.
Quick Safety Check (use this first)
Before installing any skill, call rankclaw_check:
CODEBLOCK0
Returns: verdict (SAFE / MALICIOUS / MODERATE / RISKY), score, key findings.
If is_malicious: true — do not install, no exceptions.
MCP Integration (recommended for agents)
RankClaw exposes an MCP server. Add it to your agent configuration:
CODEBLOCK1
nanobot (HTTP MCP — recommended, zero config)
nanobot supports HTTP MCP natively. No bridge script required. Add to ~/.nanobot/config.json:
CODEBLOCK2
All five tools register automatically on startup.
nanoclaw MCP config
nanoclaw uses stdio-based MCP transport. Download the bridge script and add it to .mcp.json:
CODEBLOCK3
Add to your nanoclaw .mcp.json:
CODEBLOCK4
picoclaw
picoclaw does not currently support MCP servers directly (Issue #77 closed). Use our OpenClaw skill instead — it is natively compatible with picoclaw's skill system.
Claude Code / cursor / Windsurf (HTTP MCP)
For clients that support HTTP MCP directly:
CODEBLOCK5
Available MCP tools:
| Tool | Description |
|---|
| INLINECODE5 | Safety verdict for any skill (fuzzy match) — use this first |
| INLINECODE6 |
Full score + findings for a specific skill ID |
|
rankclaw_leaderboard | Top-ranked safe skills (filter by category/author) |
|
rankclaw_find | Search by name, author, or keyword |
|
rankclaw_malicious | List of all flagged malicious skills |
Agent Decision Protocol
When an agent is about to install a skill, run this decision tree:
CODEBLOCK6
Example agent prompt
CODEBLOCK7
What RankClaw Detects
RankClaw AI audits every skill for these attack patterns:
1. Prompt Injection
SKILL.md contains hidden instructions that override the agent's system prompt. Detected by scanning for anomalous instruction blocks, invisible characters, and instruction-style text embedded in descriptions.
2. Phantom Prerequisites
Install steps reference packages or scripts not publicly available (e.g.,
openclaw-agent, private npm packages, obfuscated URLs). High correlation with credential theft.
3. Brand Impersonation
Skills named
phantom-wallet-connector,
anthropic-official,
cursor-helper that are not authored by the legitimate organization. 31 coordinated impersonation campaigns detected to date.
4. Credential Staging
Skills that write API keys, tokens, or session data to accessible paths (
.env,
/tmp/, public URLs). Usually combined with a legitimate-looking feature to avoid suspicion.
5. Supply Chain Pivots
Skills that install additional scripts or packages at runtime (not declared in install steps). Common pattern: safe SKILL.md, malicious download in install script.
6. Scope Creep
Skills claiming
scope: instruction-only but containing shell execution steps. OpenClaw has no sandbox enforcement — full host access is the reality.
Score Interpretation
| Score | Tier | Meaning |
|---|
| 90–100 | Elite | Excellent — clean, well-documented, actively maintained |
| 75–89 |
Top 25% | Good — safe to use, minor gaps |
| 60–74 | Mid | Acceptable — review findings first |
| 40–59 | Low trust | Proceed with caution — significant concerns |
| 0–39 | Risky | Not recommended — major issues found |
| N/A | MALICIOUS | Do not install — confirmed threat |
Scores are weighted: Security 40% · Maintenance 20% · Documentation 20% · Community 20%
Direct API Usage
Check a skill by ID:
CODEBLOCK8
Leaderboard (top safe skills):
CODEBLOCK9
Malicious skills list:
CODEBLOCK10
Badge (embed in your SKILL.md):
CODEBLOCK11
Badges update hourly. Green = Top 25%, Yellow = Top 60%, Red = Below 60%.
For Skill Authors
If you publish OpenClaw skills, RankClaw monitors your score continuously.
A dropping score is public. Users and agents can see it.
To protect your reputation:
- 1. Claim your skill at https://rankclaw.com → get Defender Pro ($29/mo)
- Deep AI scan — full repo audit with detailed findings
- Verified badge — shows agents your skill is author-claimed and monitored
- Score alerts — get notified immediately if your score drops
Subscribers get priority rescans when their SKILL.md changes.
Score Freshness
Scores decay over time without rescans:
- - Fresh (≤14 days): Verified — safe to trust
- Aging (15–45 days): May not reflect recent changes
- Needs Review (46–90 days): Rescan recommended
- Unverified (>90 days): Badge grays out, treat as unknown
Coverage
RankClaw indexes skills from:
- - OpenClaw (openclaw.ai) — primary registry, 14,700+ skills
- Smithery — MCP server registry (5,800+ servers, indexing in progress)
- Manus and other AI agent tool registries
One trust layer. All ecosystems.
RankClaw — AI代理信任层
RankClaw是OpenClaw生态系统及所有AI代理工具的信任与安全层。
在安装任何技能(OpenClaw、nanoclaw、picoclaw或MCP服务器)之前,RankClaw会告知你:
- - 是否恶意?(已确认1,046个恶意技能,数量持续增加)
- AI安全审计结果如何?(深度扫描SKILL.md、前置条件、安装行为)
- 信任评分是多少?(安全、维护、文档、社区四个维度,0–100分)
- 作者是否合法?(检测品牌劫持、身份冒充、虚假前置条件)
已索引14,700+个ClawHub技能。对代理免费使用。无需API密钥。
快速安全检查(请优先使用)
安装任何技能前,调用rankclaw_check:
rankclaw_check skill=wallet-connector
rankclaw_check skill=git-auto-commit
rankclaw_check skill=tkuehnl/cacheforge-stats
返回结果:判定结果(安全/恶意/中等/高风险)、评分、关键发现。
如果is_malicious: true — 请勿安装,无例外。
MCP集成(推荐代理使用)
RankClaw提供MCP服务器。将其添加到代理配置中:
MCP服务器URL:https://api.rankclaw.com/api/mcp/
协议:JSON-RPC 2.0(MCP 2024-11-05)
认证:无需
nanobot(HTTP MCP — 推荐,零配置)
nanobot原生支持HTTP MCP。无需桥接脚本。添加到~/.nanobot/config.json:
json
{
tools: {
mcp_servers: {
rankclaw: {
url: https://api.rankclaw.com/api/mcp/
}
}
}
}
启动时自动注册全部五个工具。
nanoclaw MCP配置
nanoclaw使用基于stdio的MCP传输。下载桥接脚本并添加到.mcp.json:
bash
curl -O https://raw.githubusercontent.com/RankClaw/rankclaw/main/mcp_bridge.py
添加到nanoclaw的.mcp.json:
json
{
mcpServers: {
rankclaw: {
command: python3,
args: [rankclawmcpbridge.py],
description: RankClaw — AI代理信任层
}
}
}
picoclaw
picoclaw目前不直接支持MCP服务器(Issue #77已关闭)。请改用我们的OpenClaw技能——它与picoclaw的技能系统原生兼容。
Claude Code / cursor / Windsurf(HTTP MCP)
对于直接支持HTTP MCP的客户端:
json
{
mcpServers: {
rankclaw: {
url: https://api.rankclaw.com/api/mcp/
}
}
}
可用的MCP工具:
| 工具 | 描述 |
|---|
| rankclawcheck | 任何技能的安全判定(模糊匹配)— 请优先使用 |
| rankclawscore |
特定技能ID的完整评分及发现 |
| rankclaw_leaderboard | 排名靠前的安全技能(可按类别/作者筛选) |
| rankclaw_find | 按名称、作者或关键词搜索 |
| rankclaw_malicious | 所有被标记的恶意技能列表 |
代理决策协议
当代理即将安装技能时,执行以下决策树:
- 1. 使用技能名称调用rankclaw_check
- 如果判定结果为MALICIOUS(恶意):
→ 中止。请勿安装。向用户报告恶意原因。
→ 通过rankclaw_find推荐安全替代方案
- 3. 如果判定结果为RISKY(高风险)或评分低于40:
→ 警告用户。显示关键发现。请求明确确认。
- 4. 如果判定结果为MODERATE(中等)或评分在40–79之间:
→ 谨慎安装。向用户显示摘要。
- 5. 如果判定结果为SAFE(安全)或评分大于等于80:
→ 继续安装。可选择向用户提及评分。
代理提示示例
我需要安装OpenClaw技能{skill_name}。
在继续之前,我将使用RankClaw检查其安全状态。
[调用 rankclawcheck skill={skillname}]
根据判定结果:
- - 如果为MALICIOUS(恶意):我不会安装此技能。它已被确认为安全威胁。
- 如果评分低于60:我将警告您并请求确认。
- 如果评分大于等于80:我将继续安装。
RankClaw检测内容
RankClaw AI对每个技能进行以下攻击模式审计:
1. 提示注入
SKILL.md包含覆盖代理系统提示的隐藏指令。通过扫描异常指令块、不可见字符以及嵌入在描述中的指令式文本来检测。
2. 虚假前置条件
安装步骤引用了未公开的包或脚本(例如openclaw-agent、私有npm包、混淆URL)。与凭证窃取高度相关。
3. 品牌冒充
技能命名为phantom-wallet-connector、anthropic-official、cursor-helper,但并非由合法组织创作。迄今已检测到31次协同冒充活动。
4. 凭证暂存
技能将API密钥、令牌或会话数据写入可访问路径(.env、/tmp/、公共URL)。通常结合看似合法的功能以避免怀疑。
5. 供应链转移
技能在运行时安装额外的脚本或包(未在安装步骤中声明)。常见模式:安全的SKILL.md,安装脚本中包含恶意下载。
6. 范围越界
技能声称scope: instruction-only(仅指令范围),但包含shell执行步骤。OpenClaw没有沙箱强制执行——完全的主机访问是现实情况。
评分解读
| 评分 | 等级 | 含义 |
|---|
| 90–100 | 精英 | 优秀 — 干净、文档完善、积极维护 |
| 75–89 |
前25% | 良好 — 使用安全,存在小缺陷 |
| 60–74 | 中等 | 可接受 — 请先查看发现结果 |
| 40–59 | 低信任度 | 谨慎使用 — 存在重大疑虑 |
| 0–39 | 高风险 | 不推荐 — 发现重大问题 |
| 不适用 | 恶意 | 请勿安装 — 已确认威胁 |
评分权重:安全40% · 维护20% · 文档20% · 社区20%
直接API使用
按ID检查技能:
bash
curl https://api.rankclaw.com/api/skill/tkuehnl/cacheforge-stats/
排行榜(最安全技能):
bash
curl https://api.rankclaw.com/api/leaderboard/?per
page=10&safeonly=1
恶意技能列表:
bash
curl https://api.rankclaw.com/api/leaderboard/?malicious_only=1
徽章(嵌入到你的SKILL.md中):
markdown
徽章每小时更新。绿色 = 前25%,黄色 = 前60%,红色 = 低于60%。
给技能作者的建议
如果你发布OpenClaw技能,RankClaw会持续监控你的评分。
评分下降是公开的。 用户和代理都能看到。
保护你的声誉:
- 1. 认领你的技能 访问 https://rankclaw.com → 获取Defender Pro(每月29美元)
- 深度AI扫描 — 完整仓库审计,附带详细发现
- 验证徽章 — 向代理显示你的技能已由作者认领并受监控
- 评分警报 — 评分下降时立即收到通知
订阅用户在SKILL.md变更时可获得优先重新扫描。
评分时效性
评分在未重新扫描的情况下会随时间衰减:
- - 新鲜(≤14天): 已验证 — 可安全信任
- 老化(15–45天): 可能未反映最近变更
- 需审查(46–90天): 建议重新扫描
- 未验证(>90天): 徽章变灰,视为未知
覆盖范围
RankClaw索引的技能来源:
- - OpenClaw(openclaw.ai)— 主注册表,14,700+个技能
- Smithery — MCP服务器注册表(5,800+个服务器,索引进行中)
- Manus及其他AI代理工具注册表
一个信任层。覆盖所有生态系统。