A co-op game tracker for groups of friends, with a personal-library mode that reuses the same backlog and history tools without diluting the co-op defaults. Sign in with Steam, create a group, rate games in a queue with the trailer playing, suggest new ones by pasting a store link, and settle "what next" with a group score shown next to how many people voted for it — and with a check on whether your whole group can even fit in the lobby.
Multi-tenant: any number of co-op groups or personal workspaces, each with its own library. One person can belong to several.
Running cost: €0/month on Cloudflare's free tier, which is not a trial and does not sleep. Rough headroom is in Scaling.
 Fast, focused voting. One game at a time, with keyboard-friendly controls. |
 A shortlist that fits. Choose tonight's crew and the backlog re-ranks immediately. |
 Statistics backed by real data. Empty charts simply stay out of the way. |
The screenshots use the realistic g4m1ng fixture with publishable placeholder
identities; no member account, avatar, invite code or private comment is shown.
The public landing page is also the one canonical, indexable URL. Its search and
sharing metadata lives in public/index.html; public/robots.txt keeps API routes
out of crawlers, and public/sitemap.xml lists the canonical page. Keep all three
on https://coopqueue.app/ unless the production domain changes, then update
them together. Do not add estimated sitemap dates or review scores: search metadata
must describe facts the app can substantiate.
Rate. A queue of everything you personally haven't scored yet, one card at a
time, with the Steam trailer embedded. It is split into two modes, because it asks
two different questions: Played ("how good was it?") and Want to play
("how much do you want to play this?"). Played comes first — it is the smaller set
and the more perishable, since memory fades. Keyboard-driven: 1–9 (0 = 10),
← → to nudge by half a point, Enter to save and advance, S to skip.
Forty unrated games takes about three minutes.
Play next. The backlog ranked, with attendee chips directly above it. Changing who is playing immediately recomputes the score from only those people's ratings and checks verified player caps against that crew. Unknown caps stay visibly unknown instead of emptying the list; verified games that are too small move to a separate block. Steam header art keeps both lists recognizable at a glance.
Library. Every game, every score, sortable — including by "most disagreement" and "fewest votes", which is how you find what needs attention. Filterable by whose votes a game already has, by group score, and by whether anyone has logged hours. An optional one-question-at-a-time Complete our history card finds missing first-played, last-played and group-hours facts for games the group has actually played. It asks likely participants first, offers clearly marked Steam/release-date approximations when there is honest evidence, and lets each person answer, defer or dismiss a prompt without hiding it from everyone else. It starts disabled and each member can opt in from Settings when they want it. Personal libraries collapse the group-only player, member, average and vote-count columns into one Rating column.
Find games. The most-reviewed online co-op games on Steam, as a global catalogue, defaulting to a Recommended for us order built from your group's own ratings — and it names the tags it based that on, so you can disagree with it. A Co-op only filter is on by default. Search results and catalogue cards link to the official Steam page in a new tab, so you can check the details before adding anything. Add the ones your group might play to its library, and only library games ever appear in anyone's rating queue. That opt-in step is the whole point: a catalogue of thousands would otherwise bury the handful of votes that decide what you actually play. Games whose player cap nobody has verified say so rather than guessing, and the "only ones we can all play" filter leaves them out instead of claiming they fit.
Coverage. Your own unrated count sits above the rating queue, and the Group tab ranks everyone by how much of the library they have scored. Rating coverage is the thing that decides whether any of the scoring means anything, so it is not hidden in a stats page.
Stats. A compact full-width library-status summary followed by paired genre and rating profiles, yearly starts and finishes, time-versus-rating, and a dot plot of your biggest arguments. Missing hours remain unknown and every history chart shows its date coverage. Panels with no underlying data stay hidden. Personal workspaces add a completion track record, rating distribution, playtime buckets, backlog check-in, personal records and an achievement cabinet while omitting comparisons that require another person.
Steam ownership and trophies. A three-state ownership badge uses the public libraries already synced from the Group tab: owned, absent from a sufficiently recent sync, or unknown. Private and stale data are never relabelled "not owned". Each game's detail panel also has a Trophy relay: opening it automatically refreshes the signed-in member's stale or missing progress for that one game, then the group sees its combined progress, fresh unlocks and achievements that are one person short of a clean sweep. Personal workspaces turn it into a Trophy trail with playful progress ranks. A refresh button remains for retries.
Personal workspaces. Create one directly during onboarding, or convert an existing one-person group from Settings without losing games, ratings or history. They have no invite link or group-only comparisons, use your Steam playtime in history and Stats, and retain the co-op-first catalogue and verified-cap metadata. Settings includes a selective Steam import for both the owned library and public wishlist. The same review is available to co-op groups. It previews every candidate before writing, with source and Co-op / Multiplayer / Solo filters, select/deselect-all controls, import actions above and below the list, and owned-only or wishlist-only shortcuts. New personal and co-op workspaces open this review immediately; personal libraries enable every game type while co-op groups begin with Co-op selected. Cancelling an empty import goes to Find games instead of an empty rating queue. Until a game is added, Rate, Play next, Library and Stats each offer direct routes back to the Steam importer and Find games. Source filters load their complete candidate set, and larger reviewed selections are submitted in bounded batches. Played games begin as Sampled; unplayed and wishlisted games begin as Backlog. Exact lifetime minutes are rounded to one decimal hour. Verified catalogue player caps carry across; unknown caps remain visibly unknown. Store art is filled in through the shared metadata cache in small resumable batches. Playtime sync is part of the import; metadata enrichment begins automatically in bounded passes and remains resumable from Settings. Personal workspaces also start bounded achievement analysis automatically. It can save the earliest unlock as an approximate first-played boundary. It shows Likely finished only for a small curated set of explicit ending achievements, and the player must confirm every status change.
Opening a workspace normally leads to Rate while it has unrated games. Once the queue is clear it opens the Library instead, or Find games when the library itself is empty. Workspace names only need to be unique within one person's own switcher; unrelated private workspaces may use the same name.
Group / Settings. Members, invite links you can rotate or disable, direct invitations to registered CoopQueue accounts, the headcount that drives the viability check, and owner-only JSON or per-table CSV exports — including history provenance — before a group is permanently deleted. Direct invitations stay outside the group until the named account explicitly accepts; recipients can accept or decline from their onboarding screen or the Invites button in the app.
The group's score is the plain average, shown next to the number of votes.
That is a deliberate choice rather than a shortcut. An earlier version calibrated each rating against that person's own history and shrank thin results toward the group mean. It was removed: a rating is a judgement, not a measurement needing correction — somebody who has only ever given 8s and 9s may simply have played good games, and may well give the next one a 4. And a score you cannot reconstruct from the column beside it is a poor way to settle an argument.
Thin evidence is shown rather than corrected for: a 9.0 with one vote sits next to a 1 in the Votes column, and the Library sorts by it.
Every game carries max_players — the maximum simultaneous online co-op
count, which is frequently not what a store page advertises. Compared against
the group's own headcount, a game is viable, blocked, or a workaround (couch
co-op that works over Remote Play Together, or one-copy games like Keep Talking
and Nobody Explodes). Steam's own co-op tags are stored as a hint but never
treated as the truth, because they routinely are not.
There is no registration. No email, no password, no confirmation link — nothing to store and nothing to reset. Everyone signs in with Steam OpenID, which is free, needs no approval from Valve, and is the same "Sign in through Steam" flow people already know from Humble Bundle and SteamDB.
What actually happens on first sign-in:
- The person clicks Sign in through Steam and lands on Valve's own login page. They type their Steam password on steamcommunity.com, never here — this site never sees a credential of any kind.
- Steam redirects back with a claim, which the server hands straight back to
Steam (
mode=check_authentication). Nothing in that redirect is trusted until Valve confirms it. This matters: without the callback check, OpenID is trivially forgeable by editing the URL. - On success we store one row — their SteamID64, display name and avatar — and set a signed session cookie good for 90 days.
That account is just an identity. It grants access to nothing until they either
create a co-op group or personal workspace, join one with an invite link, or
accept a direct in-app invitation. Groups are private by
default: a non-member asking for /api/g/some-group/games gets a 404, not a
403, so a private group can't even be confirmed to exist by guessing the URL.
So for your four: you sign in, claim the "Ama" identity, then either send the Group tab's invite link or search for friends who have already used CoopQueue. A direct invite creates no membership until that exact Steam account accepts it. Either join path still applies a pre-assigned imported identity and its ratings automatically.
Using Steam rather than email has one more payoff. Because a SteamID64 is a real Steam account, playtime and ownership come from Steam's public Web API keyed on the ID already stored. Achievement completion uses that API first and Steam's public Community profile as a fallback when its stats method refuses an otherwise valid key. The Community view can expose a game's achievement list while hiding every personal unlock, so an all-zero fallback remains unknown rather than becoming a false 0%. Both paths respect Steam privacy: hidden or stale data is never guessed.
If you want the instance private while you try it out, set ALLOWED_STEAM_IDS
and only those Steam accounts can sign in at all. Leave it unset to let anyone
create their own groups.
One command does the whole thing:
npm install
npm run setupIt will open a browser once for the Cloudflare login, then create the Pages
project and the D1 database, write the database id into wrangler.toml, apply
the schema, offer to import the 92-game seed data, generate and store
SESSION_SECRET, prompt for an optional Steam API key, and deploy. It prints
your live URL at the end.
The script is safe to re-run — every step checks for existing state first, so if something fails you can fix it and run it again without duplicating anything. It never applies the destructive fresh-install schema when tables already exist.
You need a free Cloudflare account (no card) and, optionally, a free Steam Web API key from https://steamcommunity.com/dev/apikey. The key is used for display names and avatars, manual library/playtime sync and per-game achievement sync. Without it the tracker and ratings still work, but those Steam-backed features do not.
For a real product you probably want pushes to deploy themselves. After running
npm run setup once, go to Workers & Pages → coopqueue → Settings → Builds,
connect the GitHub repo, set the build command to nothing and the output
directory to public. Every push to main then deploys, and pull requests get
preview URLs. The D1 binding comes from the committed wrangler.toml; secrets
stay in the dashboard.
If you would rather run the steps yourself
npx wrangler login
npx wrangler pages project create coopqueue --production-branch main
npx wrangler d1 create g4m1ng # paste the id into wrangler.toml
# Only against the brand-new, empty database created on the line above:
npx wrangler d1 execute g4m1ng --remote --file=./schema.sql
npx wrangler d1 execute g4m1ng --remote --file=./seed.sql # optional
openssl rand -base64 32 | npx wrangler pages secret put SESSION_SECRET --project-name coopqueue
npx wrangler pages secret put STEAM_API_KEY --project-name coopqueue
npm run deployImported games have names but no Steam links, so the queue has no trailers for them yet. Sign in, open your group, and:
curl -X POST https://coopqueue.app/api/g/g4m1ng/enrich \
-H "Cookie: g4_session=..." # copy from browser devtoolsEight games per call — Steam throttles at roughly 200 requests per 5 minutes —
and only unambiguous name matches link automatically. Repeat until remaining
is 0. Anything in needsReview you link by hand: open the game in the Library
and paste its store URL. Being conservative here is deliberate; a wrong appid
means the wrong trailer plays in the queue.
For a custom domain, add it under Workers & Pages → coopqueue → Custom domains. Cloudflare charges nothing for the hosting; you just pay a registrar for the domain (~€10/year).
cp .dev.vars.example .dev.vars # then edit SESSION_SECRET
npm run db:init:local
npm run devSteam OpenID won't redirect back to localhost, so for UI work use the test
harness instead — it serves the real frontend against an in-memory copy of the
schema with auth stubbed:
node test/harness.mjs # signed in, co-op and personal fixtures
SCENARIO=anon node test/harness.mjs # landing page
SCENARIO=fresh node test/harness.mjs # signed in, no groups
SCENARIO=claim node test/harness.mjs # first run in a groupnode test/isolation.mjs # 170 tenancy / permission / scoring assertions
node test/shots.mjs # renders every screen, fails on any console errorisolation.mjs imports the real route handlers and runs them against a D1
shim over SQLite, so it tests shipped code rather than a reimplementation. It
covers the things that would actually hurt: cross-group reads, cross-group
writes by guessed id, invite-code and direct-invite validation, owner-only actions,
recipient-only acceptance, and the last-owner-can't-leave rule.
functions/
_middleware.js who are you — session cookie → users row
api/
auth/steam.js start Steam OpenID (carries an invite code)
auth/callback.js verify with Steam, set the signed cookie
me.js current user + groups + pending invitations
invites.js accept or reject a direct invitation
groups/index.js GET my groups · POST create one
groups/join.js POST an invite code
steam/search.js store search for the add-a-game box
g/[slug]/
_middleware.js THE TENANCY BOUNDARY — membership or 404
games/index.js GET the group library with scores · POST add
games/[id].js PATCH · DELETE · PUT to link a Steam appid
history.js shared-history cleanup prompts
invites.js owner search, invite and cancellation
steam-import.js selective library/wishlist import
play-night.js attendee-specific backlog ranking
export.js owner-only JSON and CSV downloads
achievements.js cached per-game Trophy relay
rate.js skip.js stats.js claim.js members.js enrich.js
lib/
auth.js HMAC-signed cookie sessions, optional signup allowlist
groups.js slugs, invite codes, membership lookup
history.js missing-fact prompts and bounded automatic suggestions
stats.js shared library-shape, genre, timeline and playtime aggregations
score.js plain-mean scoring + viability
ownership.js three-state Steam-library ownership
steam.js Steam Web API and store endpoints, cache-first
http.js JSON responses, body parsing, score clamping
public/ index.html · app.js · styles.css — no build step
scripts/setup.mjs one-command Cloudflare provisioning
schema.sql · seed.sql · recommendations.sql
test/ harness.mjs · isolation.mjs · shots.mjs
No framework, no bundler. Edit a file, npm run deploy, done.
Every group-scoped route lives under functions/api/g/[slug]/, and
_middleware.js there is the only place that decides whether you're allowed in.
It looks up your membership and, failing that, returns 404 rather than 403 —
so the existence of a private group can't be probed by guessing URLs.
Below that boundary, every single statement still filters on
data.group.group_id. Game ids are global and sequential, so
PATCH /api/g/my-group/games/5000 must not touch group 2's game 5000 just
because you're a legitimate member of my-group. test/isolation.mjs asserts
exactly this for read, rate, patch, delete, and the rating queue.
Sessions are HMAC-SHA256-signed cookies (HttpOnly, Secure, SameSite=Lax,
90-day expiry) with no server-side session table. Steam OpenID responses are
verified by handing every parameter back to Steam with
mode=check_authentication; nothing in the callback URL is trusted before that
returns is_valid:true. Every database call uses bound parameters. All
user-supplied strings are escaped before reaching the DOM. Invite codes are
10 characters from a 32-symbol alphabet (~50 bits) and can be rotated or
disabled by the owner at any time.
Soft caps exist so one account can't exhaust the free tier: 20 owned groups per person, 64 members per group, 2000 games per group. The root middleware also limits Steam searches to 30/minute, Steam-backed writes to 20/5 minutes, and other writes to 60/minute per signed-in account. JSON bodies are capped at 32 KB, Steam calls time out after 10 seconds, and static responses carry a restrictive CSP and the standard browser hardening headers.
Free tier, per day: 100,000 Function requests, 5 million D1 rows read, 100,000 D1 rows written, 5 GB storage. Static assets and bandwidth are unmetered.
A rough sense of the shape: one person opening the app and rating twenty games is on the order of a hundred requests and a few thousand row reads, because the library loads in one call and everything after that filters client-side. A hundred active groups is comfortably inside the free tier; a thousand is where you'd start watching the dashboard. Storage is a non-issue — the whole original spreadsheet is under 100 KB in D1.
If you outgrow it, Workers Paid is $5/month and includes 10 million requests and 25 billion D1 row reads per month. That is the entire cost step; there is no tier in between and no per-seat pricing. Realistically the site would need to be genuinely popular before you'd notice.
The actual constraint is Steam, not Cloudflare. The store endpoints throttle
at roughly 200 requests / 5 minutes per IP, and Cloudflare Workers egress from
a shared egress pool. That's why steam_apps exists: metadata is cached globally by
appid with a 14-day TTL, so the thousandth group to add Terraria costs zero
Steam requests. Search can't be cached (unbounded query space) and is the thing
that would break first under load. If it does, the fix is a debounce on the
client plus a KV-backed cache of the last N search terms.
One thing to know before opening it up: D1's free tier has no automatic
backups (run wrangler d1 export on a cron — also free). The application now
rate-limits Steam search, Steam-backed writes, and ordinary writes per signed-in
account; the existing group, member, and game caps remain the longer-term guardrails.
docs/OPERATIONS.md— the runbook for a deployed instance: secrets, deploying, migrations, backups and recovery, catalogue maintenance, and a troubleshooting table of everything that has actually gone wrong.AGENTS.md— orientation for coding agents: the hard rules, the layout, and which constants are measured rather than chosen.CLAUDE.md— the long form of all of the above, including the numbers behind every invariant.
Steam playtime/ownership sync, attendee filtering, personal workspaces, group exports, automatic per-game achievement progress, Complete our history, comments, archiving and the recommender are done. Still open, roughly in value order:
- Scheduled D1 backups via a cron trigger running
d1 export. The free tier has no automatic backups, and Time Travel only reaches back 30 days. - Verified player caps at scale. Only 55 of 337 catalogue rows have one, so
most cards read "player count unverified".
npm run measure:capsmeasures Steam app matching andmultiplayer_modes.onlinecoopmaxcoverage in IGDB without writing a single cap. It still needs confidential Twitch developer credentials and a real coverage run before any candidate is manually verified and migrated. - A weekly nudge (Discord webhook) listing games with fewer votes than the group has members. Rating coverage is the app's core problem — in the original spreadsheet it ranged from 89% to 27% per person.
- Public read-only group pages, also the cheapest growth loop the site has.