Build cloudflare email sending

Build Cloudflare Email Sending

Cloudflare Email Service (open beta since 2026-04) sends transactional email through three equivalent paths — a Workers send_email binding, a REST API, and SMTP — all sharing one pipeline, one quota, and automatic DKIM/ARC signing. This skill gets a sending domain onboarded, picks the right send path for the caller's architecture, and executes migrations from other providers without breaking email mid-cutover.

When to use

Use this skill if you are:

• Migrating transactional email off Resend, SES, Postmark, or SendGrid onto Cloudflare.
• Fixing Supabase Auth's rate-limited built-in mailer, or fully rebranding its signup/reset/magic-link emails.
• Wiring a send_email Workers binding, wrangler email sending, or a Worker relay for a backend that runs outside Cloudflare (Railway, Fly, Vercel, a VPS).
• Architecting transactional email from scratch for a stack already on Workers.
• Debugging cf-bounce DNS, DKIM/SPF conflicts, or Email Service quota errors.

This is the sending product (open beta, 2026) — not Email Routing/receiving, not the retired MailChannels path.

First: three facts that shape every decision

1. The sending domain's DNS zone must be on Cloudflare. Onboarding auto-creates bounce MX + SPF + DKIM + DMARC records; it cannot do that on external DNS. If the target domain's nameservers point elsewhere, either move the zone first or onboard a different Cloudflare-hosted domain as an interim sender (working substitute now beats perfect sender later — swap from when the zone moves).
2. API-token creation is dashboard-gated; the Worker binding needs no token at all. For backends running outside Cloudflare, the highest-leverage pattern is a tiny Worker relay: the Worker holds the send_email binding ambiently, the backend calls it over HTTPS with a shared random secret. No Cloudflare credential ever touches the backend.
3. Quota starts at ~1,000 sends/day and Cloudflare raises it reluctantly (practitioner reports: a $20-30k/mo account fought to get 5,000/day). 3,000 sends/month included on Workers Paid ($5/mo floor), then $0.35/1k. If projected volume exceeds 1,000/day, file the limit-increase form at the start of the migration, not the end — treat it as a blocking dependency.

Workflow

1. Verify tooling and account

wrangler --version          # need >= 4.110; older versions 404 on sending endpoints
wrangler whoami             # confirms login; note available account IDs

Multi-account logins fail non-interactively — export CLOUDFLARE_ACCOUNT_ID=<id> before every wrangler email command. Find the account that owns the target zone by probing:

CLOUDFLARE_ACCOUNT_ID=<id> wrangler email sending settings <domain>
# "Could not find a zone" → wrong account; repeat with the next id

2. Onboard the sending domain

dig +short NS <domain>      # *.ns.cloudflare.com → zone is on Cloudflare; else see fact 1
CLOUDFLARE_ACCOUNT_ID=<id> wrangler email sending enable <domain>
CLOUDFLARE_ACCOUNT_ID=<id> wrangler email sending dns get <domain>   # inspect what was created

enable auto-creates, on the cf-bounce subdomain: three bounce MX records, an SPF TXT, a DKIM key at cf-bounce._domainkey, plus _dmarc.<domain>. Verify propagation before the first send:

dig +short MX cf-bounce.<domain>
dig +short TXT cf-bounce._domainkey.<domain> | head -c 40
dig +short TXT _dmarc.<domain>

SPF conflict rule: if the root domain already has an SPF record from another sender, Cloudflare's include:_spf.mx.cloudflare.net must be merged into that one record — a second SPF TXT is invalid and tanks deliverability. The cf-bounce records are namespaced and never conflict. Subdomains (notifications.<domain>) onboard independently — use one per traffic class for reputation isolation.

3. Pick the send path

Caller	Path	Why
Code already on Workers	send_email binding	Zero credentials, zero HTTP hop
Backend outside Cloudflare (Railway/Fly/VPS/Vercel functions)	Worker relay (this skill's template)	Token creation is dashboard-gated; a shared secret to your own Worker is the only credential
Legacy app that only speaks SMTP	smtp.mx.cloudflare.net:465, user api_token, password = a token with Email Sending: Edit	No code changes
One-off / CI / smoke test	wrangler email sending send --from ... --to ... --subject ... --text ...	Fastest proof

For the Worker relay: copy assets/email-sender-worker/ into the repo (e.g. workers/email-sender/), deploy, set the secret, verify:

cd workers/email-sender
CLOUDFLARE_ACCOUNT_ID=<id> npx wrangler deploy
openssl rand -hex 32 | CLOUDFLARE_ACCOUNT_ID=<id> npx wrangler secret put AUTH_TOKEN
# keep the generated value ONLY in the backend's secret store (env var / vault) — never in git

The #1 API landmine: the Workers binding's address-object shape is {email, name} while the REST API's is {address, name}. The bundled Worker template translates; hand-rolled code that passes {address} to the binding fails with "Incorrect type for the 'email' field". The wrangler CLI smoke-send accepts display strings ("Name <addr@dom>").

4. Wire the application with a fallback window

Never hard-cut from the old provider. Route by config: prefer Cloudflare when its env vars are set, fall back to the old provider otherwise — either provider alone keeps email alive, and rollback is unsetting two vars. Keep the old provider's fallback for 2–4 weeks of observed deliverability (the product is beta; no public inbox-placement history exists yet).

Provider-specific mappings, DNS specifics, and code-shape diffs:

• Migrating from Resend (or SES/Postmark — same shape): read references/migrate-from-resend.md
• Replacing Supabase Auth's built-in mailer (confirmation, magic link, recovery emails): read references/supabase-auth-emails.md
• Greenfield — architecting email for a new product on Cloudflare: read references/greenfield-architecture.md

5. Verify end-to-end, then clean up

Every leg gets a live probe — a green deploy is not a delivered email:

# 1 Worker auth gate: wrong token must 401, missing config must 503
curl -s -X POST https://<worker>/send -H "Authorization: Bearer wrong" -d '{}' -o /dev/null -w "%{http_code}\n"
# 2 Real send returns a message_id
curl -s -X POST https://<worker>/send -H "Authorization: Bearer $EMAIL_SENDER_TOKEN" \
  -H "content-type: application/json" \
  -d '{"from":{"address":"welcome@<domain>","name":"Product"},"to":"<real inbox you control>","subject":"cutover probe","text":"probe"}'
# 3 The email actually lands in that inbox (check spam folder too)
# 4 The OLD provider's send log shows nothing new → traffic really moved
# 5 Quota headroom: GET /accounts/<id>/email/sending/limits (or the dashboard)

Delete any test users/rows the probes created. If probes sent to fake addresses, expect them in permanent_bounces — repeated hard bounces auto-suppress recipients, and unsuppression requires a Cloudflare support case, so always probe with a real inbox you control.

Credential hygiene (non-negotiable)

• The Worker AUTH_TOKEN lives in exactly two places: wrangler secret and the backend's secret store. Never in git, wrangler.jsonc, docs, or chat.
• Rotation is a paired operation: wrangler secret put AUTH_TOKEN + update the backend var in the same change window.
• SMTP-path tokens are real Cloudflare API tokens — scope to Email Sending: Edit only, nothing broader.
• When writing examples or reports, show openssl rand -hex 32 as the generation step, never a literal value.

Failure quick-reference

Symptom	Cause	Fix
wrangler email sending → 404	wrangler < 4.110	npm i -g wrangler@latest
"More than one account available"	multi-account login	CLOUDFLARE_ACCOUNT_ID=<id>
"Incorrect type for the 'email' field"	REST shape sent to the binding	{email} for binding, {address} for REST
E_SENDER_NOT_VERIFIED	domain/subdomain not onboarded	wrangler email sending enable + DNS check
429 / E_RATE_LIMIT_EXCEEDED	daily quota (default ~1,000)	limit-increase form; queue + retry with backoff
Lands in spam	DMARC alignment missing	verify _dmarc TXT exists; merged SPF on root
Recipient silently gets nothing	auto-suppressed after hard bounces	Cloudflare support case; stop probing fake addresses
Dashboard shows sends as "dropped"	known Email Routing metrics false-negative	trust the send response's message_id, not that panel