User clicks "Connect Bank"

The dashboard opens Plaid Link (US/CA), GoCardless Connect (EU), or Teller Connect depending on the user's region.

User authenticates with their bank

The banking provider's widget handles the bank login (multi-factor auth, account selection). Midday never sees the user's bank credentials.

Provider returns a connection token

The banking widget returns a public token/requisition ID. The dashboard sends it to the API:

POST /trpc/bankAccounts.create
Body: { provider: "plaid", publicToken: "..." }

API exchanges token and stores connection

1. Exchange token — API calls the provider to convert public token → access token
2. Fetch accounts — API calls provider to list available accounts (checking, savings, etc.)
3. Store in DB — bank_connections record (provider, access token encrypted) + bank_accounts records (name, balance, currency, type)

Initial transaction sync triggers

A BullMQ job is queued to fetch historical transactions:

Worker: TransactionSyncJob
  1. Call provider API: getTransactions(last 90 days)
  2. For each transaction:
     a. Store in `transactions` table
     b. Queue AI categorization (match against category embeddings)
     c. Queue enrichment (merchant name, logo via provider data)
  3. Update bank_account balance

Transactions appear in the dashboard

TanStack Query invalidates the transaction list cache. The user sees their bank transactions, auto-categorized and enriched.

User creates an invoice

The invoice editor lets users add line items (invoice_products), set payment terms, choose a template, and select a customer.

Invoice is saved

tRPC: invoices.create → Drizzle INSERT into `invoices` + `invoice_products`

The invoice number is auto-generated (configurable format). The status is set to draft.

User clicks "Send"

The dashboard calls the API to send the invoice:

1. Generate PDF — @midday/invoice uses @react-pdf/renderer to produce a PDF
2. Send email — @midday/notifications sends via Resend with a React Email template (invoice.tsx)
3. Update status — Invoice status changes to sent
4. Create Stripe payment link — If Stripe Connect is configured, a payment intent is created

Customer receives and pays

The customer gets an email with a link to the invoice. The link opens a public page where they can:

• View the invoice
• Pay via Stripe (credit card, bank transfer)
• Download the PDF

Stripe webhook confirms payment

Stripe sends POST /webhooks/stripe
  Event: payment_intent.succeeded
    │
    ▼
  Update invoice status → "paid"
  Update invoice.paid_at timestamp
  Record payment details
    │
    ▼
  Notify user via email (invoice-paid.tsx template)

Recurring invoices (optional)

If the invoice is set to recurring, Trigger.dev schedules the next send:

Trigger.dev: InvoiceSchedulerTask
  Runs on schedule → creates new invoice from template → sends automatically

Receipts arrive via email

The @midday/inbox package syncs from connected Gmail or Outlook accounts:

Worker: InboxSyncJob
  1. Fetch new emails from Gmail API / Microsoft Graph
  2. Filter for receipt-like attachments (PDF, images)
  3. Store in `inbox` table with file in Supabase Storage

Document Intelligence extracts data

Worker: DocumentProcessingJob
  1. Send PDF/image to Azure Document Intelligence (OCR)
  2. Extract: merchant name, amount, date, currency
  3. Store extracted data on inbox record

AI matches receipt to transaction

Worker: InboxMatchingJob
  1. Query `transactions` for similar amount + date range + merchant
  2. Use embedding similarity for fuzzy matching
  3. If confidence > threshold:
     → Auto-link receipt to transaction (transaction_attachments)
  4. If uncertain:
     → Store as suggestion (transaction_match_suggestions)
     → User reviews in dashboard

User reviews matches

Unmatched or low-confidence receipts appear in the Inbox view. The user can manually link them to transactions or dismiss them.

User types a question

For example: "What's my burn rate this quarter?"

The dashboard sends the message to the AI chat endpoint.

AI SDK selects and executes tools

AI SDK (OpenAI / Anthropic / Google):
  1. Parse user intent
  2. Select tool: "get-burn-rate"
  3. Tool executes: Drizzle query against transactions
     → SUM(amount) WHERE date IN quarter, GROUP BY month
  4. Return structured data to the model

Model generates a response

The AI model receives the financial data and generates a natural-language response with the burn rate breakdown. The response is streamed to the user via AI SDK React hooks for instant feedback.

Suggested follow-ups

The assistant suggests related questions: "Compare to last quarter?" or "What are my top 3 expenses?"

User tracks time on a project

The time tracker records entries against a tracker_project:

• Start/stop timer (live tracking)
• Manual entry (date, duration, description)
• Entries stored in tracker_entries

User generates an invoice from tracked time

The dashboard aggregates billable entries for a project and pre-fills an invoice:

1. Sum hours per entry
2. Apply project hourly rate
3. Create line items from entries
4. Link invoice to customer and project

Invoice follows the standard send/pay flow

Same as Workflow 2 — generate PDF, send via email, collect payment via Stripe.