Multi Tenant Architecture

Multi-Tenant Platform Architecture (Cloudflare or Vercel)

• IS: platform choice, domain strategy and PSL, tenant identification, compute and data isolation, hostname routing, tenant context propagation, custom domains and SSL, per-tenant static files, and mapping platform limits to plans.
• IS NOT: general folder structure or module contracts (use codebase-architecture), scaffolding a new repo (use scaffold-nextjs), or the content of per-tenant SEO files once routing serves them dynamically: sitemap entries, canonical URLs, structured data, indexing policy (use seo).

Contents

• Platform dispatch (decide first)
• Reference files
• Workflow (order matters)
• Gotchas
• Output schema
• Pre-commit checklist
• Related skills

Platform dispatch (decide first)

• Pick one platform per product. Fronting a Vercel app with a Cloudflare proxy doubles the TLS and redirect layers and is the usual cause of redirect loops and failed certificate issuance.
• Tenants shipping their own code on Vercel is the multi-project model (one Vercel project per tenant, created with the SDK). It follows the Cloudflare row's isolation reasoning; this skill's Vercel references cover the single-deployment model only.

Reference files

Workflow (order matters)

Copy this checklist to track progress:

Multi-tenant progress:
- [ ] Step 1: Domain strategy and PSL decision
- [ ] Step 2: Tenant identification strategy
- [ ] Step 3: Isolation model (compute and data)
- [ ] Step 4: Deterministic routing
- [ ] Step 5: Tenant context propagation
- [ ] Step 6: Tenant config and least-privilege bindings
- [ ] Step 7: Custom domains and per-tenant static files
- [ ] Step 8: Limits mapped to plans, evidence captured

1. Choose the domain strategy

• Put tenant workloads on a dedicated registrable domain (acme.app for tenants, acme.com for brand). One phishing tenant on x.acme.com puts the whole domain on blocklists, and a tenant cookie with Domain=acme.com reaches your dashboard.
• Keep the dashboard and auth on a different apex from tenant subdomains (app.acme.com for the console, *.acme.app for tenants).
• If tenants publish content or run code on sibling subdomains, submit the label directly above the tenant name (acme.app, or sites.acme.app for <tenant>.sites.acme.app) to the PSL and start now: there is no SLA. Tenant-owned custom domains need no PSL entry. Otherwise record No PSL with the reason.

2. Choose tenant identification (one primary; custom domain as the upgrade path)

• Subdomain tenant.acme.app: wildcard DNS plus wildcard certificate. The default.
• Custom domain tenant.com: the tenant CNAMEs to you. Paying tenants; reputation shifts to them; needs the onboarding lifecycle in step 7.
• Path acme.app/tenant: no per-tenant DNS or certificates, but no cookie isolation and no branding. Choose it only when tenants will never get a hostname.

3. Define the isolation model

• Compute, Cloudflare: one dispatch namespace in untrusted mode; per-invocation cpuMs and subRequests limits per plan; an outbound Worker if tenant code may call the internet.
• Compute, Vercel: one deployment, tenant code never executes. If tenants must ship code, move to Vercel multi-project or Cloudflare rather than sandboxing inside the app.
• Data: shared schema with tenant_id on every tenant-aware table plus RLS is the default; database-per-tenant for regulated or noisy tenants, selectable per plan. See data-isolation.md.

4. Route deterministically (tenants never influence routing or see each other)

• Cloudflare: a single */* route on the SaaS zone to the dispatch Worker; hostname -> tenant record (KV, D1 on miss) -> env.DISPATCHER.get(script); Worker not found -> 404.
• Vercel: proxy.ts (Next.js 16; middleware.ts with runtime: 'nodejs' on 15) reads host, looks the tenant up in Global Config or the database, rewrites into the tenant segment; unknown hostname -> 404, never the brand site.
• Let /.well-known through before any tenant rewrite. Route robots.txt, sitemap.xml, and llms.txt into the tenant segment so they vary per tenant.

5. Propagate tenant context from one authority

• Delete every inbound x-tenant-* header, set x-tenant-id, x-tenant-slug, x-tenant-plan from the resolved tenant, and forward them on the request (NextResponse.next({ request: { headers } })). Server Components read await headers(); route handlers read request.headers. Cloudflare: the dispatch Worker sets headers or passes parameters before fetch.
• The proxy is routing, not authorization. Server Functions, route handlers, and jobs re-derive the tenant from the session and the data layer enforces it (RLS or tenant_id predicates).

6. Bind only what the tenant needs

• Cloudflare: each user Worker gets its own bindings (KV namespace, D1 database, R2 prefix); adding a binding is an explicit redeploy. No shared globals.
• Vercel: Global Config holds only hostname -> { id, slug, plan }; the database is the source of truth and write-through happens when a domain verifies. Feature flags and branding come from the database keyed by tenant id.

7. Support custom domains and per-tenant static files

• Lifecycle to design and record: add domain -> show DNS target -> verify ownership -> certificate issued -> mapping activated -> removal or failure path.
• Cloudflare: Cloudflare for SaaS custom hostname on the SaaS zone, proxied fallback origin, customers.<you>.com CNAME target, http or txt validation, pre-validate before DNS cutover. See cloudflare-platform.md.
• Vercel: projectsAddProjectDomain -> DNS values from the project's domain card -> _vercel TXT only if the domain is already on Vercel -> projectsVerifyProjectDomain -> Let's Encrypt HTTP-01. See vercel-domains.md.
• robots.txt, sitemap.xml, llms.txt are route handlers inside the tenant segment with explicit Content-Type; nothing tenant-specific lives in /public. Their content is seo territory.

8. Surface limits as plans and capture evidence

• Fill the limits-to-plan table from limits-and-quotas.md, re-checking each source URL and dating it; enforce at the routing layer (Cloudflare limits, Vercel plan header plus server checks).
• Nothing long-running in the request path: Cloudflare Queues or Workflows, Vercel background functions or cron.
• Every tenant operation (create tenant, add domain, verify, remove) works over HTTP with the same authority as the UI; if it only works in the dashboard, the platform leaks into the UI.
• Run the evidence commands in the pre-commit checklist and paste results into the output.

Gotchas

• Tenant headers set on the response instead of the request: NextResponse.next({ headers }) sends x-tenant-id to the browser and headers() in Server Components reads nothing. Use NextResponse.next({ request: { headers: requestHeaders } }).
• Forwarding inbound tenant headers: curl -H "x-tenant-id: <other>" then serves another tenant's data. Delete or overwrite x-tenant-* on every path through the proxy, including paths that skip resolution.
• The starter kit matcher '/((?!api|_next|[\\w-]+\\.\\w+).*)' excludes every root file with an extension, so robots.txt and sitemap.xml skip the proxy and every tenant gets the platform's /public copy. Match them, and rewrite them into the tenant segment.
• Next.js 16 renamed middleware.ts to proxy.ts (export proxy, Node.js runtime, a runtime config option throws). npx @next/codemod@canary middleware-to-proxy . migrates. A matcher that excludes a path also skips Server Function POSTs on it, so tenant checks live in the data layer too.
• Global Config (formerly Edge Config) key names must match ^[\w-]+$; tenant_acme.com is rejected. Use a collision-free encoding or hash; replacing dots with underscores can map different hostnames to the same key. Writes propagate in up to 10 s, so a "domain connected" screen that reads Global Config right after the write shows stale state; read the database there. The legacy @vercel/edge-config SDK cannot read stores connected after the rename (they create GLOBAL_CONFIG, not EDGE_CONFIG).
• RLS is bypassed by superusers and BYPASSRLS roles; table owners bypass it unless FORCE ROW LEVEL SECURITY is enabled. An app connecting as the migration role sees every tenant with policies "on". Connect as a separate role, add ALTER TABLE ... FORCE ROW LEVEL SECURITY, and test with SET ROLE app_user.
• SET app.tenant_id = ... outside a transaction on a pooled connection persists into the next request. Use set_config('app.tenant_id', $1, true) inside the transaction; with PgBouncer in transaction mode it is the only safe form.
• Wildcard *.acme.app on Vercel without Vercel nameservers never gets a certificate: DNS-01 needs Vercel to write _acme-challenge. Point ns1.vercel-dns.com and ns2.vercel-dns.com first and re-add MX records.
• /.well-known is reserved on Vercel and cannot be rewritten or redirected; a proxy that rewrites every path into /s/[slug] breaks HTTP-01 and custom-domain certificates never issue. Pass it through first.
• Cloudflare for SaaS: the fallback origin must be a proxied record in the SaaS zone; a custom hostname equal to the zone name is unsupported; _cf-custom-hostname pre-validation does not work when the customer's zone is also on Cloudflare (O2O, marked by cf-connecting-o2o: 1).
• Untrusted dispatch namespaces (default) have no request.cf and no caches.default, so tenant code reading request.cf.country throws. Trusted mode restores them but shares one cache across every tenant Worker in the namespace.
• KV is eventually consistent (up to 60 s, negative lookups cached): a hostname added after the dispatch Worker's first lookup 404s for a minute. Fall back to D1 on miss during onboarding.
• PSL rejects domains with under two years of registration left; the _psl.<suffix> TXT stays in place after merge; browsers ship the list on their own release cycles. Listing also kills Domain=acme.app cookies, including your own cross-subdomain SSO if it lives there.
• Starting path-based with custom domains on the roadmap means URL rewrites, cookie changes, and DNS migration later.
• Domain quotas and charges vary by provider and plan. Put current official limits and their access dates in the plan table before setting pricing.

Output schema

Length follows the decisions: drop any section the project does not face rather than filling it.

# Multi-tenant architecture

## Platform decision
- Platform: Cloudflare | Vercel
- Why this platform:
- Rejected platform and reason:

## Domain map
- Brand domain:
- Tenant domain:
- Tenant subdomains:
- Custom domains:
- PSL decision: Submit (suffix, owner, PR link, _psl TXT date) | No PSL (reason)

## Routing matrix
| Host pattern | Resolver | Destination | Unknown tenant behavior |
|---|---|---|---|

## Tenant context flow
- Authority: proxy.ts | dispatch Worker
- Headers set and stripped:
- Server read path:
- Data-layer enforcement:

## Isolation model
- Compute isolation:
- Data isolation (and per-plan variant):
- Config/binding isolation:

## Custom-domain lifecycle
1. DNS target:
2. Ownership verification:
3. Certificate provisioning:
4. Routing activation:
5. Removal/failure path:

## Limits-to-plan table
| Limit | Source URL / access date | Free | Pro | Enterprise | Enforcement point |
|---|---|---:|---:|---:|---|

## Validation evidence
| Check | Command | Expected | Result |
|---|---|---|---|

Pre-commit checklist

• Platform chosen with rationale; multi-project or Cloudflare chosen if tenants ship code
• Tenant workloads off the brand domain; dashboard on a separate apex; PSL decision recorded
• Identification strategy chosen; custom-domain upgrade path defined
• Isolation model defined for compute and data, including the per-plan variant
• Routing tenant-blind: unknown host -> 404; /.well-known passes through; static files vary per tenant
• Inbound x-tenant-* stripped; context set by the proxy or dispatch Worker only; data layer enforces tenant
• Custom-domain lifecycle defined end to end, including removal
• Limits table dated from official URLs; enforcement points named; long work off the request path

Evidence commands (run against local or preview; mark N/A with a reason):

Related skills

• codebase-architecture: folder structure, module contracts, and the request-context pipeline for the application itself.
• scaffold-nextjs: bootstrap the Next.js turborepo before applying these tenancy patterns.
• seo: content of per-tenant robots.txt, sitemap.xml, llms.txt, canonical URLs, and structured data once routing serves them.

Maintenance only: evals/evals.json contains regression scenarios for changes to this skill; it does not load during a user task.