bbc-skill · Bilibili Comment Collector

中文文档 · Online Docs

Built for Bilibili UP主 (content creators): fetch every comment on your own
videos and feed them to Claude Code / Codex / Gemini / any agent for
sentiment / keyword / audience analysis.

• 🐍 Zero dependencies — Python 3.9+ standard library only, no pip install
• 💬 Complete — top-level + nested + pinned comments, nothing skipped
• 📊 Video metadata — title, view/like/coin/favorite counts, tags included
• 🤖 Agent-native CLI — stable stdout JSON envelope, NDJSON stderr
  progress, distinct exit codes, dry-run, schema introspection
• 🧑‍🎤 Batch mode — fetch every video of a UP主 sequentially (one at a time,
  5-10s randomised cooldown between videos)
• 🔐 Delegated auth — human logs in once, agent just consumes the cookie;
  no browser automation
• ♻️ Resumable — re-running the same BV skips completed pages; --since
  for incremental monitoring
• 📁 Analysis-friendly — comments.jsonl + summary.json + raw/ archive

Multi-Platform Support

Follows the Agent Skills spec. Works with every
major AI coding agent:

⚠️ Responsible Use

Please read and accept these guidelines before using this tool.

• ✅ Personal, low-volume, legal use only: analyze comments on your own
  videos, or assist another creator with their explicit authorization.
• ✅ Respect the built-in throttling: 1s per request, 5-10s random
  cooldown between videos in batch mode. Do not patch these out.
• ❌ Do NOT use for:
  • Mass-scraping strangers' videos
  • Building derivative data products for resale or public redistribution
  • Bypassing rate limits, spoofing User-Agents, using proxy pools to evade
    anti-bot systems
  • High-frequency automation (e.g. daily full re-scans of the same channel)
  • Harassment, doxxing, coordinated attacks, or targeting specific users
• 📜 Comply with Bilibili's ToS and robots.txt.
  For commercial/organization use, switch to the official
  Bilibili Open Platform APIs.
• 🔒 Data minimization: fetch → analyze → delete. Do not retain raw data
  long-term, and do not share files containing user IDs or IP locations.
• 🎯 The tool is designed for one-shot analyses, not long-term
  surveillance.

This project is not affiliated with bilibili.com. Any account-level risk
control, bans, or legal consequences are the user's responsibility. When
in doubt about whether a specific use case is allowed — don't run it.

Install

Claude Code

# Global install (available in all projects)
git clone https://github.com/Agents365-ai/bbc-skill.git ~/.claude/skills/bbc-skill

# Project-level install
git clone https://github.com/Agents365-ai/bbc-skill.git .claude/skills/bbc-skill

OpenAI Codex

git clone https://github.com/Agents365-ai/bbc-skill.git ~/.agents/skills/bbc-skill
# Project-level
git clone https://github.com/Agents365-ai/bbc-skill.git .agents/skills/bbc-skill

OpenClaw / ClawHub

# Via ClawHub
clawhub install bbc-skill

# Manual
git clone https://github.com/Agents365-ai/bbc-skill.git ~/.openclaw/skills/bbc-skill

Opencode

git clone https://github.com/Agents365-ai/bbc-skill.git ~/.config/opencode/skills/bbc-skill
# Or reuse an existing ~/.claude/skills/bbc-skill — Opencode reads that path too

Hermes Agent

git clone https://github.com/Agents365-ai/bbc-skill.git ~/.hermes/skills/data/bbc-skill

SkillsMP

skills install bbc-skill

Standalone CLI (no skill)

git clone https://github.com/Agents365-ai/bbc-skill.git && cd bbc-skill
./scripts/bbc --help
# Or add to PATH
export PATH="$PWD/scripts:$PATH"

Installation paths summary

Quick start

Step 1 · Export your Bilibili cookie

Bilibili's comment API rate-limits and returns thin data for unauthenticated
requests. For full UP主 analysis you must authenticate with a cookie.

Recommended: the open-source Chrome extension
Get cookies.txt LOCALLY
— runs entirely locally, uploads nothing.

1. Install Get cookies.txt LOCALLY from the Chrome Web Store.
2. Visit https://www.bilibili.com and confirm you
   are logged in (avatar visible top-right).
3. Click the extension icon → Export → download
   www.bilibili.com_cookies.txt.
4. Save it somewhere convenient, e.g. ~/Downloads/bilibili_cookies.txt.

Other options:

• Firefox: cookies.txt
  add-on.
• Edge: the same Chrome extension works.
• Manual: DevTools F12 → Application → Cookies → copy the SESSDATA value,
  then export BBC_SESSDATA="<value>".

Do not share SESSDATA — it authorizes full account access.

Step 2 · Verify the cookie works

./scripts/bbc cookie-check --cookie-file ~/Downloads/bilibili_cookies.txt

Expected:

{"ok": true, "data": {"mid": 441831884, "uname": "探索未至之境", "vip": true, "level": 5, ...}}

If it fails: confirm you are currently logged in at bilibili.com, re-export
the cookie, and retry.

Step 3 · Fetch comments

./scripts/bbc fetch BV1NjA7zjEAU \
  --cookie-file ~/Downloads/bilibili_cookies.txt

URLs are accepted too:

./scripts/bbc fetch "https://www.bilibili.com/video/BV1NjA7zjEAU/"

Output lives in ./bilibili-comments/BV1NjA7zjEAU/:

bilibili-comments/BV1NjA7zjEAU/
├── comments.jsonl      # flat JSONL — one comment per line
├── summary.json        # video meta + aggregated stats + top-N
├── raw/                # archived API responses
└── .bbc-state.json     # resume / incremental state

Environment variables

export BBC_COOKIE_FILE="$HOME/Downloads/bilibili_cookies.txt"
./scripts/bbc fetch BV1NjA7zjEAU

Or pass SESSDATA directly:

export BBC_SESSDATA="<value from DevTools>"
./scripts/bbc fetch BV1NjA7zjEAU

Analysis workflow with Claude Code

After fetch completes, ask Claude something like:

Read ./bilibili-comments/BV1NjA7zjEAU/summary.json first — give me the
overall picture: video stats, comment distribution, top 20 liked. Then
I'll direct what to analyze next.

Claude follows this path:

1. Read summary.json first (a few KB) — video title, stats, time
   distribution, IP distribution, top-N comments.
2. Sample comments.jsonl on demand — each line is a flat JSON record;
   Grep for keywords, head/tail for chronology, sort by like for
   hot-comment analysis.
3. Typical analyses:
   • Sentiment: positive / negative / neutral ratio
   • Keyword frequency (excluding stopwords)
   • UP interaction audit: filter is_up_reply=true, see which threads you
     replied to vs. missed
   • Geographic breakdown from ip_location
   • Feedback evolution: bucket ctime_iso by week/month
   • Super-fan detection: group by mid, rank by comment count
   • Negative-review triage: high like + negative keywords

Commands

bbc fetch <BV|URL>

--max N                 Cap top-level comments (default: all)
--since <date>          Only fetch comments newer than this (ISO, e.g. 2026-04-01)
--output <dir>          Output directory (default ./bilibili-comments/<BV>/)
--cookie-file <path>    Netscape cookie file
--browser <name>        auto / firefox / chrome / edge / safari
--format json|table     stdout format
--dry-run               Preview request plan, no network calls
--force                 Ignore resume state, refetch everything

bbc fetch-user <UID> (coming soon)

Batch fetch across a UP主's entire video catalog.

bbc summarize <dir>

Rebuild summary.json from an existing comments.jsonl.

bbc cookie-check

Validate the cookie and print the logged-in user.

bbc schema [command]

Return JSON schema for a command (param types, exit codes, error codes).

Exit codes

Output schemas

comments.jsonl record

{
  "rpid": 296636680849,
  "bvid": "BV1NjA7zjEAU",
  "parent": 0,
  "root": 0,
  "mid": 71171081,
  "uname": "user nickname",
  "user_level": 4,
  "vip": false,
  "ctime": 1776521119,
  "ctime_iso": "2026-04-18T06:25:19+00:00",
  "message": "...",
  "like": 1,
  "rcount": 0,
  "ip_location": "河北",
  "is_up_reply": false,
  "top_type": 0,
  "mentioned_users": [],
  "jump_urls": []
}

• parent=0 → top-level; otherwise rpid of the parent comment
• top_type: 0=normal, 1=UP pinned, 2=editor pinned
• is_up_reply: true if the comment was authored by the video owner

summary.json fields

• video: title, description, stats, tags, cover URL, owner
• counts: total, top-level, nested, pinned, unique users, UP replies,
  completeness ratio
• time_distribution: earliest/latest timestamps, daily histogram
• top_liked: top-N comments by like count
• top_replied: top-N top-level comments by reply count
• ip_distribution: histogram of IP provinces

See references/agent-contract.md for the full schema.

Limits & caveats

• Read-only — never posts, edits, or deletes. Safety tier: open.
• Rate — 1s between top-level requests, 0.5s for nested. ~5000 comments
  takes 10-15 minutes.
• Anti-bot — HTTP 412 triggers exponential backoff (3 retries).
• Completeness — the completeness field in summary.json compares
  fetched vs. declared counts; values below 1.0 indicate deleted comments
  or API inconsistency.
• Anonymous not supported — UP主 analysis requires a valid cookie.

References

• SKILL.md — skill trigger + usage guide for Claude
• references/api-endpoints.md — Bilibili
  API fields used
• references/agent-contract.md — envelope /
  exit code / schema contract

Contributing

Suggestions, bug reports, and pull requests are all welcome. If you have
ideas — new analysis workflows, better anti-bot defaults, additional
platform support, documentation fixes — feel free to
open an issue or
submit a PR directly.

This skill is community-friendly: every contribution, no matter how small,
helps make it better for everyone.

License

MIT

Support

If this skill helps you, consider supporting the author:

WeChat Pay

Alipay

Buy Me a Coffee

Author

Agents365-ai — building open-source skills for AI coding agents.

• Bilibili: https://space.bilibili.com/441831884
• GitHub: https://github.com/Agents365-ai
• Skills: drawio-skill · asta-skill · paper-fetch · more →