| slug | pixiv-cli |
| version | 0.7.0 |
| displayName | Pixiv CLI |
| summary | Safely operate Pixiv with the pixiv-cli binary for discovery, account actions, and downloads. |
| license | MIT |
| homepage | https://github.com/FlanChanXwO/pixiv-cli |
| tags | ["pixiv","cli","mcp"] |
| name | pixiv-cli |
| description | Operate Pixiv through the pixiv-cli binary — search illustrations, novels, and users; inspect Pixiv artwork or user IDs/URLs; view rankings and recommendations; manage bookmarks/follows; and download works. Load only when the user explicitly mentions Pixiv or pixiv-cli, provides a pixiv.net URL or ID in a clear Pixiv context, or requests a specific Pixiv operation or `pixiv` command. Do not trigger for generic illustration, artist, image-search, or download requests without Pixiv context. Verify current syntax with `pixiv <cmd> --help`. |
pixiv-cli Operator
Teaches an agent to drive the pixiv CLI correctly. This skill encodes workflow
orchestration, safety rules, and semantic traps — flag details always defer to
the installed binary's pixiv <cmd> --help output.
Preflight and account checks
- Run
pixiv version --json. The binary itself is the only environment probe:
if it is missing or not executable, report that blocker. Install only when
the user explicitly asked for installation; then read
references/install.md and use the official platform script. Otherwise do
not install or guess an installation method.
- Do not enumerate local accounts on every session. Run
pixiv auth list --json only when authentication, account selection, or an
anonymous-fallback decision actually requires it. Presence in this list does
not prove a credential is currently valid; use the networked
pixiv auth check [UID] --json only when validation is needed. Treat both
{"accounts": null} and {"accounts": []} as an empty account list.
Hard rules
- Refresh tokens and authentication export bundles contain secrets. Except
for the explicit bare-stdout export case below, do not echo, log, summarize,
or reproduce them in commentary or results. Use the controlled workflows in
references/auth.md when the user's task requires moving a credential.
- A bare
auth export is allowed only when the user explicitly asks to
receive or see its raw token or bundle for that invocation. Before running
it, explain that the secret necessarily enters tool output/transcript and
may be retained there, then obtain the user's explicit confirmation. After
execution, do not repeat, transform, parse, or log the secret. This is the
sole exception to rule 1; otherwise use --output or a direct same-command
pipeline to the intended consumer.
- If the user already disclosed a refresh token in the conversation and
explicitly asks to import it, the agent may run the positional
pixiv auth import 'RFT' form. Do not create an extra file or repeat the
secret in the result. The token will still be copied into the tool call,
process arguments, shell history, and process context; disclose that before
execution, and explain that an already-disclosed token cannot be erased.
- For a token not already disclosed, keep it out of chat and command
arguments. Use the hidden
pixiv auth import prompt only when the runtime
gives the user a terminal they can type into directly. Do not start it in a
standard agent PTY with no user-input channel and leave it waiting; instead,
give the command to the user for their private terminal, or use an authorized
secret-manager-to-stdin pipeline. Non-TTY input is read automatically; there
is no --stdin flag. This is an environment constraint, not a command ban.
- Run interactive
pixiv auth login only when the user explicitly asks and
is present to complete browser OAuth.
- Do not accept or forward cookies (
PHPSESSID, refresh_token=... cookie
strings) as credentials — the CLI rejects them by design; only raw Pixiv App
API refresh tokens work.
Operation tiers
| Tier | Commands | Behavior |
|---|
| Credential transfer | auth import auth export | Execute only for the user's explicit import/export task; follow references/auth.md so secret input/output is not exposed accidentally |
| Read | search novel search search-options detail ranking recommended user * config get/path version update --check | Execute when the user's task requires it |
| Account diagnosis | auth list/check | List only for authentication/account/fallback decisions; check only when network validation is needed |
| Write | bookmark add/remove follow add/remove | State the target (illust/user ID) in one line before executing |
| Disk | download | Confirm target directory and exact targets (IDs or supported Pixiv URLs) before each invocation; a user URL expands every visual work, so state that scope explicitly; approval never carries over; see references/download.md |
| Interactive credential | auth login | Run only on explicit request while the user is present for browser OAuth |
| Account/config state | auth use/remove config set/unset update (actual install) | Ask for explicit confirmation each time; approval does not carry over |
| MCP server | mcp | Run only when the user explicitly asks to start it; it is a long-lived stdio JSON-RPC server, not a data command—do not auto-probe, auto-wait, or include it in preflight |
Output & token control (in priority order)
- Reduce at the source (preferred): pass a positive
--limit N only when
that command's help exposes it. In the audited binary these commands are
search, novel search, ranking, recommended, user search, user artworks, user bookmarks, and
user following. Add --page, --type, --rating, or other flags only
when that specific command's help exposes them.
--limit 0 requests all results, so never use it unless the user explicitly
asks for everything. A result that reaches --limit N is not proof that
only N matches exist; follow the pagination and completeness contract in
references/discover.md.
- Small result, display only: use the default human-readable output and
relay it. JSON carries field names and metadata — it is larger than the
table for display purposes.
- Programmatic processing (extract IDs, filter, chain into a next
command): use
--json. If the result is large, redirect to a temp file and
use built-in Grep/Read tools to search and read in segments.
- Opportunistic tooling: probe once for
jq; if present, prefer
--json + jq for field selection. If absent, fall back to tier 3
silently — never ask the user to install anything.
- Check status before parsing JSON:
--json controls successful output;
it does not guarantee that usage, validation, flag, or authentication errors
are JSON. Inspect the exit code first. On failure, expect stdout may be empty
and report the plain stderr error rather than treating it as malformed JSON.
Command cheat sheet
Verify flags with --help before use; this list is orientation, not a contract.
pixiv auth list --json # only when an account decision needs it
pixiv auth check [UID] --json # validate token, shows user_id/username
pixiv auth import # hidden TTY prompt, or automatic non-TTY stdin
pixiv auth import --file PATH # restore a versioned bundle offline
pixiv auth export UID --output PATH # write one private versioned bundle
pixiv auth export --all --output PATH # write all accounts to a private bundle
pixiv auth use [UID] # switch default account (confirm first)
pixiv config path # print config.toml location
pixiv config get download_path # read one effective setting
pixiv config set download_path ./downloads # config write; confirm first
pixiv search "WORD" --limit 10 --json # illustration search
pixiv search "WORD" --rating sfw --type illust --ai-mode exclude
pixiv search "WORD" --resolution high --aspect-ratio landscape --draw-tool "CLIP STUDIO PAINT"
pixiv novel search "WORD" --rating sfw --min-text-length 1000 --limit 10 --json
pixiv search-options "WORD" --json # authenticated dynamic tool choices
pixiv detail ILLUST_ID_OR_URL --json # single artwork detail
pixiv ranking --mode day
pixiv recommended illust --limit 10 # kind is REQUIRED; needs auth
pixiv recommended all --limit 10 # request all supported kinds; needs auth
pixiv user search "WORD" --limit 10 --json # source labels App search vs related-author fallback
pixiv user detail USER_ID --json # full public profile (USER_ID required)
pixiv user artworks [USER_ID] --limit 20 # omit USER_ID = current account
pixiv user bookmarks [USER_ID] --tag TAG --limit 20
pixiv user following [USER_ID] --limit 20
pixiv bookmark add ILLUST_ID --tag TAG # --tag repeatable; write op
pixiv bookmark remove ILLUST_ID # write op
pixiv follow add USER_ID # write op
pixiv follow remove USER_ID # write op
pixiv download TARGET... [--pages 1,3-5] [--quality original|regular|small|thumb|mini] [--download-path DIR]
pixiv update --check --json # read-only update check
Common per-command flags on Pixiv data commands: --uid UID (pick local account),
--json, --proxy URL / --no-proxy (this command only, never persisted).
Common config keys include download_path, filename_template, https_proxy,
and web_fallback_enabled; inspect a key with pixiv config get KEY instead of
assuming its effective value.
Critical semantics (traps — read before assuming a bug)
- No silent fallback, by design. With a refresh token present, App API
errors are surfaced as-is and NEVER auto-fall back to the anonymous web API.
Anonymous fallback happens only when no token exists anywhere AND
web_fallback_enabled=true, and only for search / detail / ranking /
download. Authenticated detail, pages, and ugoira metadata therefore
come from App API; a missing App page resource is a real error, not a reason
to scrape or retry Web.
recommended requires a kind. Choose one of the kinds shown by
pixiv recommended --help; it requires authentication and does not work
anonymously.--limit is command-specific. It is available on search, ranking,
recommended, user search, user artworks, user bookmarks, and user following;
do not attach it to detail, search-options, auth/config commands, or
other commands whose help omits it. Where supported, a positive value sets
the maximum result count and 0 requests all results. --page requires a
positive --limit.- Search flags are command-scoped. Verify
--search-by, --period,
--start-date, --end-date, --sort, --rating, --type, --ai-mode, --aspect-ratio,
--resolution, --draw-tool, --bookmark-min, and --bookmark-max against pixiv search --help; do not infer
undocumented aliases or attach illustration-only filters to other commands.
novel search supports --search-by, --sort, --period, --rating,
--min-text-length, --max-text-length, and --original-only; it does not
support illustration AI, drawing-tool, resolution, aspect-ratio, or type filters.
- Anonymous restricted search fails explicitly. Web fallback uses only
reliable illustration-search filters.
r18, r18g, mature, --search-by tag-title-caption,
bookmark-count bounds, and search-options
require App authentication; do not present the failure as an empty result
or add a Cookie workaround. novel search is App-only and requires authentication.
Bookmark count is a public bookmark total, never a like count.
- Extended rankings need authentication. Valid modes are
day,
day_male, day_female, week, week_original, week_rookie, month,
day_manga, week_manga, month_manga, week_rookie_manga, day_r18,
day_male_r18, day_female_r18, week_r18, week_r18g. The final nine
must not be replaced with an anonymous day ranking.
- Empty filtered batches are skipped. With local filters, search continues past leading empty upstream batches to the first non-empty logical batch or true end;
--limit N fills logical results and --limit 0 walks all filtered results. Do not invent request caps.
- No like-count field. Do not invent or label bookmark totals as likes.
update --json is only valid with --check. The actual install never
emits JSON.- Proxy is per-command. The browser's system proxy is NOT inherited.
Persist with
pixiv config set https_proxy URL; override per command with
--proxy / --no-proxy (mutually exclusive). With an HTTP(S) proxy,
resource downloads deliberately use HTTP/1.1; App API, OAuth, and Web
metadata requests retain normal protocol negotiation.
- Long downloads may legitimately take time. Do not impose an arbitrary
timeout or kill the process merely because it is slow; wait for completion,
user cancellation, or a real error.
- Tag search has query grammar.
user bookmarks --tag TAG filters
bookmark listings; bookmark add --tag TAG adds a repeatable bookmark tag.
search has no --tag flag — put the tag expression in its required WORD.
For a reliable boolean tag query, use --search-by tag-exact: tagA tagB
requires both complete tags, and uppercase tagA OR tagB accepts either.
Literal AND is not an operator. The default tag-partial also accepts the
verified uppercase OR syntax, but its fuzzy/alias/translated matches are
not a strict exact-tag AND. title-caption and App-only tag-title-caption have no boolean-tag contract;
no literal-uppercase-OR escape syntax is verified.
- Direct URLs are intentionally narrow.
detail accepts only an artwork
ID or a pixiv.net/www.pixiv.net HTTPS /artworks/{id} URL (an optional
locale, query, or fragment is harmless). download also accepts /users/{id}
and /users/{id}/artworks, which expand every illust, manga, and ugoira for
that user — no novels and no implicit limit. User URL downloads require App
OAuth and never use Cookie, WebView, or anonymous fallback. Do not suggest
short links, old URLs, FANBOX, Pixivision, Sketch, or HTML scraping.
Routing
| Task | Read |
|---|
Explicitly install or repair the missing pixiv binary | references/install.md |
| Import, export, back up, or restore authentication | references/auth.md |
| Find works/artists (search → filter → detail chains) | references/discover.md |
| Download workflows (single, batch, ugoira) | references/download.md |
| Errors: auth failures, network/proxy, empty results | references/troubleshooting.md |
Image delivery for agents
After pixiv download or MCP download, share local file paths via the host attachment API. If the host cannot attach files, share the artwork url only and do not claim an image was sent.
