Email Intake & Triage CRM
An inbound enquiry handling system. Someone emails a dedicated address; the system
researches their company, writes a structured record, checks whether they've emailed
before, assigns a priority with reasoning, and surfaces it on a login-gated dashboard
where records can be updated by chatting in plain language.
Live: https://email-intake-crm.vercel.app
Intake address: [email protected]
Stack
- Next.js 16 (App Router, TypeScript, Tailwind) on Vercel
- Supabase — Postgres, Row Level Security, Auth
- Claude (Anthropic API) — company research, priority classification, record-update chat
- Mailgun — inbound email routing
Pipeline
email -> Mailgun MX -> route: forward()
-> POST /api/inbound/mailgun (verify HMAC signature, insert row, return 200)
-> Supabase enquiries table
-> [enrichment + classification, separate path]
-> dashboard
The webhook stays thin on purpose. Mailgun retries any non-2xx response for hours, and
two Claude calls would blow past its timeout — so the handler verifies, inserts, and
returns 200. Enrichment happens off the request path.
Project status
Current state, what is outstanding, costs, and configuration live in
STATUS.md.
Eight of the nine build-order steps are complete and deployed; the chat-based record
update interface is the remaining item.
Debugging log
Every problem hit during this build — with the actual error text, the real cause,
and the fix — is documented in TROUBLESHOOTING.md.
It covers seven issues, including two application bugs that failed completely
silently, a DNS misconfiguration that made inbound mail impossible, and two
occasions where work still in flight was wrongly declared dead. The closing
section draws out what the failures had in common.
Known limitations
- Supabase free-tier projects pause after inactivity. A paused project means the
webhook's insert fails and Mailgun retries. A keep-alive is planned.
- The webhook is a public endpoint that triggers paid Claude calls. HMAC signature
verification means only Mailgun can invoke it, and a hard daily cap (
ENRICHMENT_DAILY_CAP,
currently 25) bounds the spend. The cap is claimed with a single atomic Postgres
statement and fails closed. Measured cost is roughly $0.15-0.25 per enriched enquiry
at medium effort; a full-effort run measured $0.54, almost entirely from web search
results being fed back into context.
- Email confirmation is enabled and uses Supabase's built-in email service.
Delivery works but is slow (measured: ~3 minutes) and the free tier allows
only a few messages per hour, so repeated sign-up testing will hit the limit.
Custom SMTP via Mailgun is the proper fix and remains outstanding.
- Priority quality depends on
src/lib/business-context.ts. That file describes what
a valuable enquiry looks like for this specific business. It is the difference between
real triage and a keyword rule, and it should be edited to match reality.
- Every push to
main deploys to production. The GitHub repo is connected to the
Vercel project, so a broken commit goes live. npm run build passing locally is the
gate before each commit.
Local development
npm install
cp .env.example .env.local # then fill in every value
npm run dev
Database migrations live in supabase/migrations/ and are applied with:
supabase db push --db-url "postgresql://postgres.<ref>:<password>@aws-0-<region>.pooler.supabase.com:5432/postgres"
