ScienceClaw: Multi-Agent Investigation — 技能工具
v1.0.2Run a multi-agent autonomous scientific investigation on any topic. Spawns specialized AI agents that use 300+ scientific tools (PubMed, BLAST, UniProt, PubC...
1· 235·0 当前·0 累计
by 安全扫描
OpenClaw
可疑
medium confidenceThe skill largely aligns with its stated purpose (runs a local multi-agent scientific investigation using an LLM), but it instructs the agent to run unverified local Python code, read workspace files, and post to an external platform without declaring the required posting credentials — these inconsistencies warrant caution.
评估建议
This skill delegates work to a local ScienceClaw installation and asks for an Anthropic API key — that is reasonable for an LLM-driven multi-agent tool. Before installing/using it, verify the following: (1) inspect the actual code in your SCIENCECLAW_DIR (bin/scienceclaw-investigate and the repository) so you know what will run; (2) confirm how posting to Infinite is authenticated (do you have an Infinite token stored locally?) and whether that token will be used; (3) be aware the instructions a...详细分析 ▾
ℹ 用途与能力
Name/description (multi-agent scientific investigation) match required binaries (python3) and primary credential (ANTHROPIC_API_KEY) — an LLM API key is reasonable. However the skill claims posting to the Infinite platform but does not declare any Infinite posting credential or config; it also claims use of many external tools (PubMed, BLAST, UniProt, ChEMBL, etc.) but does not request any corresponding API keys or explain how those connectors are authenticated. These omissions could be legitimate if the local ScienceClaw install contains connectors and credentials, but the SKILL.md does not document that.
⚠ 指令范围
The runtime instructions tell the agent to cd into a user-owned directory (SCIENCECLAW_DIR), source a virtualenv and run a local python script (bin/scienceclaw-investigate). That means the agent will execute arbitrary code from the user's filesystem — the skill package provides no code or auditability. The instructions also explicitly tell the agent to read workspace memory (memory.md) and save file paths for attachments; that grants the agent access to potentially sensitive local project data. Reading workspace memory and accessing attachments may be reasonable for richer context, but this broad file access is not declared in the skill's required config paths and is a privacy risk.
ℹ 安装机制
No install spec and no code files (instruction-only) — this is low surface risk from the skill package itself. However, the skill instructs running a local installation (~/scienceclaw) that will be responsible for tool integrations and network calls; since the skill doesn't install or verify that code, the real runtime behavior depends entirely on whatever is present at SCIENCECLAW_DIR, which could be arbitrary and untrusted.
ℹ 凭证需求
PrimaryEnv is ANTHROPIC_API_KEY which aligns with multi-agent LLM-driven work. No other env vars are declared, which is good from a minimal-secrets perspective, but the SKILL.md expects posting to Infinite and calling many external tools without declaring credentials for those services — either those credentials are managed by the local ScienceClaw install (possible), or they are missing (incoherent). The instruction to read memory.md is an additional data-access requirement not represented in requires.config.
✓ 持久化与权限
always is false and autonomous invocation is permitted (platform default). The skill does not request persistent/system-wide privileges nor declare modifications to other skills. There is no 'always:true' or other elevated persistence requested.
安全有层次，运行前请审查代码。
运行时依赖
无特殊依赖
版本
latestv1.0.22026/3/17
Remove ~/LAMM from all default paths — SCIENCECLAW_DIR now defaults to ~/scienceclaw
● 无害
安装命令 点击复制
官方npx clawhub@latest install scienceclaw-investigate
镜像加速npx clawhub@latest install scienceclaw-investigate --registry https://cn.clawhub-mirror.com
数据来源：ClawHub ↗ · 中文优化：龙虾技能库
OpenClaw 技能定制 / 插件定制 / 私有工作流定制
免费技能或插件可能存在安全风险，如需更匹配、更安全的方案，建议联系付费定制