An unofficial MyFigureCollection.net API: a Python
library plus a local MCP server, so agents can read figure data, collections,
lists and clubs.
MFC has no working public API. Its own API club (#349)
has been promising "API v4 coming soon" for eight years. This package reads the
public HTML instead, and turns it into typed models.
New here? The short version
MyFigureCollection is the largest catalog of anime figures and merch on the
internet, and for years there has been no good way for a program to read it.
This project fixes that on your own computer. You install it once, and an AI
assistant like Claude can then look things up on MFC for you: "how much is this
figure at AmiAmi?" or "list every Chiikawa item releasing in October."
sequenceDiagram
participant You
participant Claude
participant api as mfc-api, on your computer
participant MFC as myfigurecollection.net
You->>Claude: How much is the Alice Carroll figure?
Claude->>api: get_partner_listings(287)
api->>MFC: fetch the item's Buy window
MFC-->>api: HTML
api-->>Claude: AmiAmi · Available · ¥11,980
Claude-->>You: AmiAmi has it in stock for ¥11,980
What is MCP? The Model Context Protocol is a standard plug for giving AI
assistants new abilities. An MCP server is a small program on your machine that
an assistant is allowed to call. This one gives your assistant 12 MFC abilities,
from item search to barcode lookup. Adding it takes one snippet in a config
file, shown in Use it as an MCP server.
Why not use MFC's official API? There isn't one. The site's own API club
has been waiting for it since 2018.
Why isn't this a website I can visit? Cloudflare guards MFC and decides who
gets in partly by IP reputation. It trusts connections from home computers and
distrusts servers, so a hosted version would get blocked within days while a
copy on your own machine keeps working. That constraint shaped the whole
design.
Do I need to know Python? Two terminal commands to install (below). After
that your assistant does the driving.
Why this exists (and why the obvious approach fails)
MFC is behind Cloudflare, which blocks on TLS fingerprint — not user-agent.
requests, httpx and aiohttp all get a 403 "Just a moment..." challenge no
matter what headers you send. That is why the older tenji scraper stopped
working; its parsers never even received HTML.
This package uses curl_cffi with
impersonate="chrome", which reproduces a real Chrome handshake and gets a 200.
That choice is load-bearing. Swapping in a normal HTTP client will break
everything.
It is also why this ships as a local tool rather than a hosted service:
Cloudflare weights IP reputation, so a scraper on datacenter IPs gets challenged
and burns the shared address for everyone. Run it on your own machine.
flowchart LR
A["requests / httpx / aiohttp<br/>(any headers you like)"] -->|"403 Just a moment..."| CF{Cloudflare}
B["Your browser"] -->|200| CF
C["mfc-api via curl_cffi<br/>(Chrome TLS handshake)"] -->|200| CF
CF --> MFC[("myfigurecollection.net")]
Install
This needs Python 3.10+. The package is on PyPI:
pip install myfigurecollection-api
On a Mac, python3 is often Xcode's 3.9, which will not run this code, so be
explicit about the interpreter (e.g. pip3.12). Working from a clone instead,
use pip3.12 install -e ..
Then check it landed:
If mfc-api isn't found, your shell's python3 bin directory isn't on PATH. Find the
script with python3.12 -c "import sysconfig; print(sysconfig.get_path('scripts'))" and
use that absolute path.
Use it as an MCP server
mfc-api with no arguments starts the MCP server on stdio. Add this to your MCP config —
~/Library/Application Support/Claude/claude_desktop_config.json for Claude Desktop, or
.mcp.json in your project for Claude Code:
{
"mcpServers": {
"myfigurecollection": {
"command": "uvx",
"args": ["myfigurecollection-api"]
}
}
}
This needs uv installed; uvx fetches the package
from PyPI on first run, so there is nothing else to set up.
No uv? Point command at the installed script by its absolute path, e.g.
/Library/Frameworks/Python.framework/Versions/3.12/bin/mfc-api. The path must be
absolute because MCP clients launch servers with a minimal PATH that does not
include your shell's Python bin directory — a bare "command": "mfc-api" will fail
even though it works in your terminal.
| Tool | What it returns |
|---|
get_item(item_id) | Full item: category, picture, origins, characters, companies, artists, releases (date / edition / price / barcode), scale, height, rating, owner counts |
search_items(query, page, category_id) | 50 items per page, with pagination |
search_by_barcode(barcode) | Item lookup by JAN/UPC — the join key for matching an item across stores |
get_partner_listings(item_id) | Which partner shops sell an item, with availability and price |
get_shop(shop_id) | Homepage, location, shipping, rating, review and item counts |
search_shops(keywords, page, average_score, partners_only) | The shop directory, 10 per page |
get_profile(username) | Join date, hits, rank, about fields, per-category collection counts |
get_collection(username, status, page) | 50 items per page; status is owned, ordered, wished or favorites |
get_user_lists(username, page) | A user's published item lists |
get_list(list_id, page) | A list's owner, description, tags and items |
get_club(club_id) | Description, threads, recent comments, staff, members |
get_club_members(club_id, page) | The full member roster |
get_club is new — tenji never had it.
Two things to know about store data
Prices only exist where MFC has a barcode. get_partner_listings returns every
partner shop, but MFC can only check real stock for items whose JAN it recorded. For an
item with no barcode, all 32 partners come back as maybe_available with no price.
That means "not checked", not "in stock". Filter on availability == "available".
search_by_barcode is a lookup, not a search. A known barcode returns the full item
under item with matched: true. An unknown one returns matched: false — which
includes barcodes that are real but that MFC simply never recorded, common for Western
releases.
Configuration
| Environment variable | Default | Meaning |
|---|
MFC_API_RATE_LIMIT | 1.0 | Seconds between requests |
MFC_API_CACHE_TTL | 3600 | Cache lifetime in seconds; 0 disables |
MFC_API_CACHE_DIR | ~/.cache/mfc-api | Where cached pages live |
MFC_API_LOG_LEVEL | WARNING | Server log level (logs go to stderr) |
Use it as a library
from mfc_api import MFCClient, CollectionStatus
with MFCClient() as mfc:
item = mfc.get_item(287)
print(item.name)
print(item.releases[0].price)
print([c.name for c in item.characters])
collection = mfc.get_collection("Climbatize", status=CollectionStatus.OWNED)
print(collection.stats.owned, collection.pagination.total_pages)
for page in mfc.iter_collection("Climbatize", max_pages=3):
print(page.pagination.current_page, len(page.items))
Barcode in, prices out:
from mfc_api import MFCClient
with MFCClient() as mfc:
match = mfc.search_by_barcode("4543341130624")
if match.matched:
for listing in mfc.get_partner_listings(match.item.id).available:
print(listing.shop_name, listing.price, listing.currency)
Use it from the shell
mfc-api barcode 4543341130624
mfc-api shops Japan --partners
mfc-api collection Climbatize --status owned --page 2
Everything prints JSON on stdout; logs go to stderr. mfc-api --help lists every
subcommand. mfc-api-cli is kept as an alias.
Archiving an item page to the Wayback Machine
A price or a listing you read today is uncitable tomorrow: the page will have
moved on, and nobody has to take your word for what it said. archive=True
files the public item page with the Internet Archive's Save Page Now and
attaches the result to the item.
with MFCClient() as mfc:
item = mfc.get_item(287, archive=True)
print(item.releases[0].price, item.archive.wayback_url)
Three things fall out of one hook: a price-history series accumulates as a
byproduct of ordinary reads, every row becomes auditable by someone who does not
trust you, and the data survives the site. That last one is not hypothetical —
tenji died, and Mandarake went dark in September 2026.
It never breaks a read. A failure is a row, not an exception:
item.archive.status is one of success, already (a recent capture existed,
so the Archive declined to make another — still a preservation success),
skipped, pending, or failed with a reason. If the Internet Archive is
down you still get the item.
Nothing about a person is ever archived
Most of MFC is people. is_public_item_url is a whitelist of /item/<id>
with no query string, so profiles, collections, user lists and club rosters are
refused before any request is made:
mfc.snapshot("https://myfigurecollection.net/item/287")
mfc.snapshot("https://myfigurecollection.net/profile/someone")
A whitelist rather than a blacklist on purpose: a new personal URL shape cannot
leak through something that only lists what to keep.
It is opt-in for a reason
Reads default to archive=False, and archiving is meant to run as its own
deliberate pass over the items you actually care about — not inline in a
latency-sensitive loop:
- Weekly cadence. A JSON ledger records when each URL was last captured;
anything captured in the last 7 days is skipped without a request. Weekly is
plenty of resolution for a price curve and much gentler on both the Archive
and MFC. The submit also carries
if_not_archived_within=7d, so the Archive
enforces the same window even if the ledger is lost.
- Serialized, one every 6 seconds, with a per-run cap of 50, and
Retry-After honoured on a 429.
| Setting | What it does |
|---|
MERCH_ARCHIVE=0 | Kill switch — disables every snapshot in the process. Shared with the sibling merch clients, so one variable turns archiving off across a bulk run. |
MERCH_ARCHIVE_LEDGER | Ledger path, shared across sibling clients so one pass keeps one honest cadence per URL. |
MFC_ARCHIVE_LEDGER | Ledger path for this package only. Default: <cache dir>/archive-ledger.json. |
IA_ACCESS_KEY / IA_SECRET_KEY | Internet Archive S3 keys. Free, from archive.org/account/s3.php. On macOS they are also read from the Keychain services ia-s3-access / ia-s3-secret. |
Without credentials the result is status="failed" with a reason naming the
fix — not a crash, and not a silent no-op.
Build one Archiver per pass and reuse it (the client does this for you), so
the rate limiter and the per-run cap apply across the whole batch rather than
resetting each time.
Being a good citizen
- Rate-limited to 1 request/second by default, process-wide.
- Responses are cached on disk for an hour, so repeated agent calls cost nothing.
- Read-only. There are no login, write, or vote operations, and none are planned.
- Wayback snapshots are opt-in, weekly per URL, capped per run, and restricted
to
/item/<id> pages — never a profile, collection, list or club page.
If you scrape MFC hard enough to be noticed, Cloudflare will start challenging
your IP, and you will have made the site worse for everyone. Don't.
Tests
The default run is entirely offline — parser tests read saved HTML fixtures
captured on 2026-07-30. Tests that hit the live site are marked live and
deselected by default:
python3 -m pytest -m live
Run those when MFC's markup may have changed; they are the canary.
Credits
The parser architecture — a parser class per page, given HTML, returning a
model — is ported from tenji by
Nate Shoffner (MIT). The selectors are not: MFC's templates moved on since
tenji's last release. Two endpoints tenji used are gone entirely
(itemlists.v4.php now 404s), and the collection, profile, list and shop pages
all render different markup than it expects.
One piece of tenji did survive verbatim: the Buy window is still a form POST to
/item/{id} with commit=loadWindow&window=buyItem, answering with JSON. That
envelope is tenji's discovery and it still works in 2026.
tenji's MIT notice is reproduced in full in NOTICE, along with a
precise account of what was and wasn't derived from it.
Disclaimer
Unofficial. Not affiliated with, endorsed by, or supported by
MyFigureCollection.net or Tsuki Board. It scrapes public pages and will break
whenever MFC changes its HTML. Item data belongs to MFC and its contributors.
MIT licensed — see LICENSE.
Built by Sara Kay.
mcp-name: io.github.ssskay/myfigurecollection-api