The single identity provider that MAC's apps trust. Users sign in with Google or Microsoft (passwordless), and this service mints short-lived, asymmetrically-signed JWTs (EdDSA / Ed25519) that MAC apps verify locally against a published JWKS — no app calls this service per request.
- Domain:
https://auth.monashcoding.com - Stack: Node 22 · TypeScript (ESM) · Express · Better Auth
^1.6.9· Drizzle ORM · Postgres 16 - Deploy: self-hosted Dokploy on the shared MAC Oracle VM, behind Dokploy's Traefik
(which terminates TLS via Let's Encrypt). See
DEPLOY-dokploy.md. - Building an app that needs login? Skip to Integrating your app — the full copy-paste recipe.
The service is a Dokploy Compose stack of three containers
(docker-compose.dokploy.yml):
| Service | Image | Role |
|---|---|---|
auth |
built from ./auth |
This service (Better Auth on Express). Runs DB migrations on boot, then listens on :3000 |
postgres |
postgres:16 |
Self-hosted identity store (user/session/account/verification/jwks) |
backup |
postgres:16 |
Nightly pg_dump with rotation into ./backups (no host cron, no third-party image) |
Ingress is handled by the shared Traefik that Dokploy already runs on the box — this
service does not run its own reverse proxy and does not publish host ports. Traefik
routes auth.monashcoding.com → auth:3000 over the dokploy-network.
The JWT plugin generates and stores its Ed25519 signing keypair itself in the jwks
table on first boot — there are no manual openssl steps. Public keys are served at
/api/auth/jwks.
Canonical identity: the JWT carries macUserId, which is the Better Auth user.id.
Keep this as the canonical user identifier across all MAC apps (a later migration will
preserve MonMap's existing user.id values as these IDs).
Full step-by-step (dashboard access, Git deploy key, environment, domain, verification, and
the Traefik/dokploy-network gotcha) is in DEPLOY-dokploy.md.
In short: it's a Dokploy Compose service pointing at this repo with Compose Path
docker-compose.dokploy.yml, the .env pasted into Dokploy's Environment tab, and a Domain
(auth.monashcoding.com, port 3000, Let's Encrypt) added in the Domains tab.
Check it's healthy:
curl -s https://auth.monashcoding.com/health # -> {"status":"ok"}
curl -s https://auth.monashcoding.com/api/auth/jwks # -> JWKS with an Ed25519 keyauth.monashcoding.com must resolve to the VM (a direct A record → the VM's public IP).
Traefik obtains and renews the Let's Encrypt cert automatically once the Domain is added and
DNS resolves — there are no manual cert steps.
Copy .env.example to .env for local use, or paste the values into Dokploy's Environment
tab for deployment. Never commit .env — secrets live in the committee password manager
under projects@monashcoding.com.
| Variable | Notes |
|---|---|
BETTER_AUTH_URL |
Public base URL, e.g. https://auth.monashcoding.com. Used as the JWT iss. |
BETTER_AUTH_SECRET |
Long random secret. Generate with openssl rand -base64 32. |
POSTGRES_USER / POSTGRES_PASSWORD / POSTGRES_DB |
Self-hosted Postgres credentials. Keep the password URL/shell-safe — letters+digits only ($ @ : / # break env interpolation and the assembled DATABASE_URL; use openssl rand -hex 24). |
GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET |
Google OAuth client. The ID must end in .apps.googleusercontent.com (watch for truncation when pasting). |
MICROSOFT_CLIENT_ID / MICROSOFT_CLIENT_SECRET |
Microsoft OAuth client (tenant common). |
TRUSTED_ORIGINS |
Comma-separated extra origins allowed to start auth flows. Every https *.monashcoding.com subdomain is trusted automatically, so this is only for off-domain origins (e.g. http://localhost:3000 in dev). |
JWT_AUDIENCE |
JWT aud claim — mac-suite. |
NOTION_TOKEN / NOTION_ROSTER_DB_ID |
Notion integration token (owned by projects@, shared with notioncal-to-gcal) + the committee-roster database id. Drives the derived committee/exec/team claims. If unset, the roster sync is skipped and those claims stay empty — logins still work. |
NOTION_VERSION |
Notion API version. Match notioncal-to-gcal (default 2022-06-28). |
ADMIN_EMAILS |
Comma-separated infra superusers granted admin (NOT from Notion). Also gates POST /api/admin/sync-roster. |
FORCE_ROSTER_SYNC |
Set to 1 for ONE sync to bypass the >50%-removal guard (recruitment churn), then unset. |
Register these redirect URIs exactly:
- Google:
https://auth.monashcoding.com/api/auth/callback/google(Google Cloud Console → APIs & Services → Credentials → OAuth 2.0 Client) - Microsoft:
https://auth.monashcoding.com/api/auth/callback/microsoft(Entra ID → App registrations → your app → Authentication → Web → Redirect URIs)
Microsoft uses tenantId: "common" so personal Microsoft accounts (Hotmail/Outlook/Live)
work as well as work/school accounts.
Apps can rely on exactly these claims (plus standard sub, iat):
{
"macUserId": "<user.id>",
"email": "<user.email>",
"roles": ["member", "committee", "exec"],
"team": "Events",
"ver": 1,
"iss": "https://auth.monashcoding.com",
"aud": "mac-suite",
"exp": 1234567890
}- Algorithm: EdDSA (Ed25519), signed with the key published at
/api/auth/jwks. - Access-token lifetime: 15 minutes.
rolesis the union of the user's base roles (member) and roles derived from the committee roster at mint time:committee(on the roster),exec("Executive"in their Notion Team), andadmin(env allowlist, not from Notion). See Committee roster.teamis the person's first functional team (e.g."Events"), ornullif not on the roster. Informational — gate access onroles, notteam.isMonashis recorded on theuserrow (from@monash.edu/@student.monash.edu) for later eligibility logic — it is not in the token and does not gate signup.
Copy examples/verify.ts into the app (only dependency: jose).
It fetches and caches the JWKS with createRemoteJWKSet and checks iss, aud, and
exp, returning typed { macUserId, email, roles, team, ver } claims. No per-request call
to this service.
Committee membership, team, and exec status are derived from the central Notion committee
roster (the Committee Directory database) and injected into the token — no app keeps its own
membership list. Curate the committee once in Notion; removing someone there revokes their
committee/exec roles everywhere on their next token.
How it flows: Notion → roster tables (Postgres) → claims.
- An hourly
node-cronjob fetches the Notion roster and upserts it into theroster/roster_emailtables (keyed by the stable Notion page id). - The
Committee Directoryalso holds former members (taggedCurrent MAC Role = "Alumni"); the sync skips alumni, so only current committee get roles. Re-tag someoneAlumnito revoke their access on the next sync. - At token-mint time, the login email is matched against
roster_email(any of a person's Student / Preferred / Personal / Work emails) and the roles/team are derived. Nothing is stamped on the user row. - Tokens are always minted from Postgres, never from live Notion — so a Notion outage cannot affect logins. It only means the roster stops updating until Notion recovers (the last-synced roster keeps serving). If the roster read itself ever fails, the mint path falls back to base roles rather than blocking the login.
Roles: committee (present in the roster), exec ("Executive" in the person's Notion
Team multi-select), admin (env ADMIN_EMAILS, independent of Notion).
First-time setup (per environment):
- Set
NOTION_TOKEN,NOTION_ROSTER_DB_ID,NOTION_VERSION, andADMIN_EMAILS(see the Environment table). In Dokploy these go in the service's Environment tab — and they must also be forwarded indocker-compose.dokploy.yml(compose only injects the vars it references; they're already listed there). See DEPLOY-dokploy.md. - Share the
Committee Directorydatabase with the Notion integration that ownsNOTION_TOKEN(theGcalIntintegration, shared withnotioncal-to-gcal), or the sync gets a 404. - Deploy (migration
0001creates the roster tables automatically), then run the first sync so you don't wait for the hourly job — see Apply now below. On boot the logs show[roster-sync] scheduled hourly (0 * * * *).(or a "not set — skipping" warning if the env vars didn't land).
Operating it:
- Apply now (recruitment day, or the first sync): run the manual sync instead of waiting up to
an hour. In production (Docker Swarm):
Locally use
docker exec "$(docker ps -qf name=auth | head -1)" node dist/sync/cli.js # -> [roster-sync] upserted=<n> removed=<n>
npm run sync-roster; there's alsoPOST /api/admin/sync-roster(gated toADMIN_EMAILS, needs an admin session). - Guard rail: a sync that would empty the roster or remove >50% of people is refused (likely a
bad fetch or fat-fingered Notion edit). For genuine churn, run once with
FORCE_ROSTER_SYNC=1, then unset it. - A committee member 403s / has no
committeerole? First check they aren't taggedAlumni. Then check the email they log in with is one of their Notion emails (Student / Preferred / Personal / Work). A typo in an email is the usual cause — fix it in Notion, then re-sync. - Notion token lives with
projects@. If rotated, updateNOTION_TOKENhere and innotioncal-to-gcal.
This is the full recipe for making any MAC app use this service as its login. The app stores
no passwords and no accounts — it trusts JWTs this service mints and keys its own data by
macUserId.
User ─sign in─▶ auth.monashcoding.com ─Google/Microsoft─▶ shared session cookie (.monashcoding.com)
│
Your app ◀─ JWT {macUserId,email,roles,team} ◀─ GET /api/auth/token ◀─┘
└─ verifies the JWT locally (jose + JWKS) — no call back to auth per request
└─ stores/loads its data keyed by macUserId
-
Serve the app on a
*.monashcoding.comsubdomain (e.g.jobs.monashcoding.com). Cross-app single sign-on relies on a cookie scoped to.monashcoding.com, so an app on a different domain won't get silent SSO (sign-in still works, just not shared). Being on a*.monashcoding.comsubdomain also means the app's origin is trusted automatically — no auth-side config or redeploy is needed to onboard it. -
npm i joseand copyexamples/verify.tsinto the app's backend.(Only if the app is served off-domain — e.g. local dev on
http://localhost:3000— add that origin toTRUSTED_ORIGINSin the auth service's Dokploy Environment tab and redeploy.)
Social sign-in is a POST (not a GET link). It returns a URL to redirect the user to:
const res = await fetch("https://auth.monashcoding.com/api/auth/sign-in/social", {
method: "POST",
headers: { "Content-Type": "application/json" },
credentials: "include", // so the session cookie is set
body: JSON.stringify({
provider: "google", // or "microsoft"
callbackURL: "https://jobs.monashcoding.com/", // where to return after login
}),
});
window.location = (await res.json()).url; // → Google/Microsoft consent → back to your appAfter the user returns, the browser holds a session cookie for .monashcoding.com.
const { token } = await fetch("https://auth.monashcoding.com/api/auth/token", {
credentials: "include", // sends the shared cookie
}).then(r => r.json());
// send it to YOUR backend:
await fetch("/api/whatever", { headers: { Authorization: `Bearer ${token}` } });If /api/auth/token returns 401, the user isn't signed in — send them through Step 1.
import { verifyMacToken } from "./verify"; // examples/verify.ts
const auth = req.headers.authorization?.replace("Bearer ", "");
const claims = await verifyMacToken(auth); // throws if invalid/expired
// claims: { macUserId, email, roles, team, ver }Set AUTH_URL=https://auth.monashcoding.com in the app's env (verify.ts reads it).
claims.macUserId is the canonical, stable per-person ID. Use it as the foreign key for the
app's own tables — never store the email as the primary key (emails can change).
claims.roles (e.g. ["member"], ["member","committee","exec"]) comes straight from the token:
// Committee-only feature:
if (!claims.roles.includes("committee")) return res.status(403).end();committee / exec / team are derived from the committee roster (Notion,
synced hourly) — you never manage membership per-app. member is the baseline for any signed-in
user; admin is an infra allowlist. Changes appear on each user's next token.
- Access tokens live 15 minutes. When a call 401s on an expired token, re-fetch a fresh
one from
/api/auth/token(the session cookie lasts much longer) and retry. - Sign out with
POST https://auth.monashcoding.com/api/auth/sign-out(credentials: "include"). This clears the shared session across all MAC apps.
| Endpoint | Method | Purpose |
|---|---|---|
/api/auth/sign-in/social |
POST | Start Google/Microsoft login → returns { url } |
/api/auth/token |
GET | Mint a JWT for the current session → { token } |
/api/auth/get-session |
GET | Inspect the current session |
/api/auth/sign-out |
POST | End the session |
/api/auth/jwks |
GET | Public keys (verify.ts uses this; you don't call it directly) |
Backups are written to ./backups/<db>-<timestamp>.sql.gz (inside the Dokploy compose
working directory on the server) on boot and nightly, keeping the last BACKUP_KEEP_DAYS
(default 14). Each file is a plain-SQL pg_dump, gzipped.
To restore into the running stack (this replaces current data for those tables):
gunzip -c backups/mac_auth-YYYYMMDD-HHMMSS.sql.gz \
| docker exec -i <postgres-container> psql -U "$POSTGRES_USER" -d "$POSTGRES_DB"For a clean restore into an empty database, stop the auth container, drop/recreate the
database, then pipe the dump in as above.
cd auth
npm install
# Point DATABASE_URL at any local Postgres 16, then:
npm run db:generate # drizzle-kit generate (regenerate SQL after schema changes)
npm run db:migrate # apply migrations
npm run dev # tsx watch on :3000Regenerate the Better Auth schema after changing auth.ts (the CLI is the source of truth
for what the installed Better Auth version expects), then reconcile src/schema.ts:
npx @better-auth/cli generateMigrations under auth/drizzle/ are committed (not regenerated at build) so their
timestamps stay stable — otherwise a rebuild makes the drizzle migrator re-run
already-applied migrations against the live DB and crash. After changing schema.ts, run
npm run db:generate and commit the new migration. (auth/dist/ stays gitignored.)
Code quality is enforced by Biome (lint + format) and GitHub Actions. Run the same checks CI runs before you push:
cd auth
npm run lint # biome check (formatting + lint)
npm run lint:fix # auto-fix formatting / imports / safe lint issues
npm run typecheck # tsc --noEmit
npm run ci # lint + typecheck together (what the CI gate runs)Every push and PR to main runs .github/workflows/ci.yml: lint → typecheck → build →
docker build. Since Dokploy auto-deploys from main, a red CI is your signal that the
deploy would ship broken — fix it before merging. noExplicitAny at the Notion API
boundary is intentionally a non-blocking warning. Dependabot opens weekly dependency PRs;
CI verifies each one still builds before you merge.
For a new committee member inheriting this service:
- Access to the Oracle Cloud tenancy and the VM (SSH key added).
- Access to the committee password manager (
projects@monashcoding.com) — this holds the Dokploy admin login,BETTER_AUTH_SECRET, Postgres password, and both OAuth secrets. - Can reach the Dokploy dashboard (SSH tunnel to
localhost:3000, seeDEPLOY-dokploy.md). - DNS control for
monashcoding.com(to keepauth.pointed at the VM). - Owner/editor on the Google Cloud OAuth client and the Microsoft Entra app registration; confirm the redirect URIs above are still registered.
- Verify the stack:
curl .../health,curl .../api/auth/jwks. - Verify backups are being written and test a restore.
- Know how to redeploy: push to
main, then Deploy in Dokploy.
Do not put any secret in this repo. Rotate BETTER_AUTH_SECRET and OAuth secrets when
a committee member with access graduates.
- No email/password auth, email sending, or verification/reset flows.
- No user data migration from MonMap (a separate, later deliverable).
- No OIDC Provider plugin (the JWT plugin is sufficient).
- Eventual direction is to manage this box via Dokploy Cloud (shared MAC org) rather than the self-hosted panel — a later, deliberate migration.