菜单
Run Bumblebee supply-chain inventory and exposure scans on macOS/Linux to detect compromised packages, extensions, and MCP host configs.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/phoroth/AGENTIC --skill bumblebee命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | bumblebee |
| description | Run Bumblebee supply-chain inventory and exposure scans on macOS/Linux to detect compromised packages, extensions, and MCP host configs. |
| category | security |
| risk | safe |
| source | community |
| source_repo | mycelos-ai/bumblebee-skill |
| source_type | community |
| date_added | 2026-05-27 |
| author | stefan-kp |
| tags | ["security","supply-chain","incident-response","npm","pypi","tooling"] |
| tools | ["claude"] |
| license | MIT |
| license_source | https://github.com/mycelos-ai/bumblebee-skill/blob/main/LICENSE |
Bumblebee (https://github.com/perplexityai/bumblebee) is a read-only inventory collector that surfaces package, extension, and developer-tool metadata on developer endpoints. It answers a focused supply-chain question: when an advisory names a package or version, do any matches exist on this machine right now?
This skill drives a single Bumblebee scan from start to finish:
bumblebee binary.baseline, project, or deep).Communicate with the user in the language they used (German for Stefan). Code, commit messages, and on-disk file contents stay in English to match existing project conventions.
Use this skill when an advisory, incident report, or exposure catalog names compromised packages, developer tools, browser/editor extensions, or MCP host configuration that may exist on a local macOS or Linux developer endpoint.
Use it for read-only inventory and exposure checks. Do not use it to patch, uninstall, quarantine, or otherwise mutate the scanned machine.
Before running anything, confirm two things with the user via AskUserQuestion, unless the message already pins them down:
baseline (global package roots), project (specific dev folders like ~/code), or deep (explicit --root paths, including $HOME for incident response).project and deep profiles, ask which directories to scan. deep is the only profile that accepts a bare-home root.If the user has an advisory or exposure-catalog file ready, also ask whether they want to pass it via --exposure-catalog. The skill does not ship its own catalogs — point them at threat_intel/ in the Bumblebee repo if they ask where to find ready-made ones.
Skip the questions for one-liner asks like "lauf mal ne Baseline-Scan" — just run a baseline.
Run command -v go && go version in bash. Three outcomes:
brew install go (or download from https://go.dev/dl/).sudo apt install golang-go only as fallback.sudo dnf install golang or the official tarball.After installation, the user must ensure $GOBIN (or $HOME/go/bin) is on $PATH so bumblebee is found later.
Run command -v bumblebee && bumblebee version. If missing:
go install github.com/perplexityai/bumblebee/cmd/bumblebee@latest
Then re-check bumblebee version. If the binary still cannot be located, the user's GOBIN/PATH is likely misconfigured — surface the resolved go env GOPATH and go env GOBIN so they can fix it. Do not fall back to running the binary by absolute path silently; explain what is happening.
Once installed, also run bumblebee selftest as a sanity check. A non-zero exit means the local install is broken and the scan should not proceed.
All scans write NDJSON to a file. Use the workspace folder for output so the user can open the results afterwards.
Output filenames (use the user's workspace path; the example below assumes $OUT is set):
bumblebee-<profile>-<UTC-timestamp>.ndjson — raw records.bumblebee-<profile>-<UTC-timestamp>.report.md — Markdown report (generated in Step 5).Pick a sensible --max-duration so a runaway scan does not hang the session. Reasonable defaults:
baseline: 5mproject: 10mdeep: 15m (warn the user that scanning $HOME can still take longer; offer to raise the limit)Always stream stderr to a sibling .log file — Bumblebee emits diagnostic NDJSON there that helps explain partial scans.
bumblebee scan --profile baseline \
--max-duration 5m \
> "$OUT/bumblebee-baseline-$TS.ndjson" \
2> "$OUT/bumblebee-baseline-$TS.log"
Optional: scope to specific ecosystems if the user only cares about, say, npm and PyPI:
bumblebee scan --profile baseline --ecosystem npm,pypi ...
Each --root must be an existing absolute path. Reject bare $HOME for this profile (Bumblebee will reject it too — surface the message clearly).
bumblebee scan --profile project \
--root "$HOME/code" \
--root "$HOME/Developer" \
--max-duration 10m \
> "$OUT/bumblebee-project-$TS.ndjson" \
2> "$OUT/bumblebee-project-$TS.log"
Used for incident response — broad roots are allowed but should be paired with an exposure catalog and --findings-only whenever possible, so the output stays focused.
bumblebee scan --profile deep \
--root "$HOME" \
--exposure-catalog "$CATALOG" \
--findings-only \
--max-duration 15m \
> "$OUT/bumblebee-deep-$TS.ndjson" \
2> "$OUT/bumblebee-deep-$TS.log"
If the user has no catalog, run deep without --findings-only but warn them that the NDJSON file can grow large (hundreds of MB on dense developer machines).
Run the bundled helper to turn the NDJSON into a human-readable report. Resolve
the helper from the installed Bumblebee skill directory; never run a
workspace-relative scripts/render_report.py from the scanned project.
BUMBLEBEE_SKILL_DIR="/absolute/path/to/the/bumblebee-skill-directory"
test -f "$BUMBLEBEE_SKILL_DIR/scripts/render_report.py"
python3 "$BUMBLEBEE_SKILL_DIR/scripts/render_report.py" \
"$OUT/bumblebee-<profile>-$TS.ndjson" \
"$OUT/bumblebee-<profile>-$TS.report.md"
The helper groups records by type and ecosystem, lists every finding record with its catalog entry and severity, and embeds the scan_summary for traceability. It is dependency-free Python 3 — no pip install needed.
If render_report.py exits non-zero (malformed NDJSON, missing summary), surface stderr to the user instead of silently producing an empty report.
End the turn with:
computer:// links to both the NDJSON and the Markdown report so the user can open them directly..log file indicate skipped roots or read errors, mention it and link the log too.Do not paste large chunks of NDJSON into the chat — it is noisy and not where the user will read it.
npm uninstall actions from inside this skill; the user runs remediation themselves once they know what is affected.env blocks. Bumblebee does not emit those values, but the .log file may still contain paths to sensitive config files. Treat the output files as containing inventory data and do not upload them to third-party services without the user's explicit consent (DSGVO-relevant).bumblebee with elevated privileges (sudo). It is meant to inspect the current user's developer enprojectnment, not the whole system.bumblebee: command not found after go install → almost always a PATH/GOBIN problem. Show go env GOPATH GOBIN PATH to debug.refusing to scan bare home with profile baseline → use deep for $HOME, or pick a subdirectory for project.--root set, scope with --ecosystem, or raise --max-duration. Do not loop and retry blindly.schema_version and entries keys (bare top-level arrays are rejected) and that schema_version is one Bumblebee understands.See scripts/render_report.py for the report layout. Bumblebee's own documentation lives at https://github.com/perplexityai/bumblebee — consult docs/inventory-sources.md, docs/transport.md, and docs/state-model.md when a question goes beyond what this skill covers.
Bumblebee is developed by Perplexity (https://github.com/perplexityai/bumblebee, Apache-2.0). All scan logic, output formats, and exposure-catalog semantics belong to that project. This repository is just a thin Claude-skill wrapper around the official bumblebee CLI; the wrapper itself is MIT-licensed (see LICENSE).