Bibi

BibiGPT — video summarizer agent skill (summarize / transcript / chapters)

This file is a discovery stub, not the usage guide. It tells you which mode you are in and where the live docs are. The live sources always match the current product; anything copied into this file would go stale.

This skill is watch-video only. Sibling skills in the same repo (same CLI, same account): bibi-library (saved videos / notes / collections), bibi-feed (subscribe / latest / mark seen), bibi-vision (frames / mind map). Landing: https://bibigpt.co/mcp · Human install: https://bibigpt.co/agent Quota: Plus 100/day, Pro 300/day on the agent-skill channel; extra calls use API balance; upgrade https://bibigpt.co/shop

1. Detect Mode

Run scripts/bibi-check.sh if present, or check in order:

Neither available → install the desktop app (brew install --cask jimmylv/bibigpt/bibigpt on macOS, winget install BibiGPT on Windows, curl -fsSL https://bibigpt.co/install.sh | bash on Linux) or get an API token. Details: references/installation.md.

2. Load Live Docs (per mode)

CLI mode — the CLI is the always-current doc source:

bibi check-update            # once per session; run `bibi upgrade` if outdated
bibi --help                  # full command surface, grouped, with examples
bibi <command> --help        # progressive help — every --help includes examples
bibi commands                # re-fetch server-defined commands (new capabilities
                             # appear here without a binary update)

API mode — the machine-readable spec is the source of truth:

https://bibigpt.co/api/openapi.json      # all endpoints, schemas, auth

references/api.md has curl examples but the spec wins on conflict.

MCP mode — connect and list tools; they are self-describing:

https://bibigpt.co/api/mcp               # Streamable HTTP, OAuth 2.1

Latest docs without reinstalling — this repo is served raw from GitHub:

https://raw.githubusercontent.com/JimmyLv/bibigpt-skill/main/skills/bibi/SKILL.md
https://raw.githubusercontent.com/JimmyLv/bibigpt-skill/main/skills/bibi/references/<name>.md
https://raw.githubusercontent.com/JimmyLv/bibigpt-skill/main/skills/bibi/workflows/<name>.md

If a local workflows/ or references/ path below is missing (embedded-only install via bibi skill), fetch it from the raw URL above instead.

3. Intent Routing

Disambiguation: intent matches more than one workflow → ask one clarifying question first. Matches none → ask what they want; do not guess. Bare URL with no context → default to workflows/quick-summary.md.

4. Quick One-Liners (CLI mode)

For single-command requests that don't need a full workflow — discover the rest via bibi --help:

bibi summarize "<URL-or-local-file>"     # quick summary (local: .mp4 .mp3 .m4a ...)
bibi summarize "<INPUT>" --chapter       # chapter summary
bibi summarize "<INPUT>" --subtitle      # transcript only
bibi me                                  # account, plan, remaining minutes

URLs containing ? or & must be quoted. API mode has no local-file upload — guide the user to a public URL (OSS/S3) first, see references/supported-platforms.md.

5. Payment Fallback (HTTP 402)

If no auth is set and the user has an Alipay account, BibiGPT may respond with HTTP 402 Payment Required + Payment-Needed header (AI 收 protocol). The bibi CLI prints a stable marker line [HTTP/402 Payment Required] to stderr before any human-readable prompt. When either signal appears, route to references/billing-aipay.md instead of treating the call as failed — the agent can resolve payment automatically via `@alipay/agent-payment` or a one-off QR purchase.