Settlematic MCP for Claude and ChatGPT: dashboard vs AI workflows, setup, and when each wins

Finance software used to mean screens: click Clients, click New, fill twelve fields, save, click Invoices, repeat. That model still works — and for complex line-item editing, it often should. But the operators we talk to in 2026 live in Claude, ChatGPT, and Slack threads. They want to ask 'create a $4,200 invoice for Nexus Digital, Net 14, USDC allowed' and get a draft with a payment link — not alt-tab through five browser tabs before their coffee cools.

Settlematic ships a production MCP (Model Context Protocol) server mapped to the same REST API your engineers use. Connect it to Claude in the browser via Settings → Connectors, or wire ChatGPT Custom GPT actions from our OpenAPI document. This guide compares dashboard workflows vs MCP-assisted workflows, documents every tool surface, and explains how user journeys change when billing becomes conversational — without bypassing permissions.

What MCP is — and what Settlematic exposes

MCP is Anthropic's open standard for giving AI assistants structured access to external systems. Instead of pasting CSV exports into chat, the model calls tools with typed parameters. Settlematic's remote connector authenticates with your org's API key and registers tools for clients, invoices, payments, reports, tax exports, and bulk operations. Claude interprets natural language; Settlematic enforces scopes server-side. The assistant cannot see another org's data or perform writes your key does not allow.

The connector runs at your backend MCP endpoint — no local Node install required for Claude web users. Paste the MCP URL from Settings → API keys, enable the connector in a new chat, and tools appear when relevant. Desktop users can alternatively run our published @settlematic/mcp-server package if they prefer local stdio transport.

Tool surface reference (what you can do from chat)

As of MCP server version 1.3.0, these tools map to production API behavior:

• list_clients / get_client / create_client / update_client — client directory CRUD
• list_invoices / get_invoice — filter by status, full line items and payment history
• create_invoice — resolve client by email, apply org defaults, optional send flag
• send_invoice — email hosted payment link (confirm before production use)
• get_payment_link / get_invoice_pdf — share settlematic.com/pay URLs and PDF downloads
• get_dashboard_summary — outstanding AR and KPI snapshot
• get_payments_received — on-chain receipts with tx hashes for a date range
• get_reports_summary / get_visual_report — revenue summaries plus shareable chart URLs
• get_tax_report / get_tax_report_csv — jurisdiction-aware tax buckets and CSV export
• get_bulk_invoice_template / validate_bulk_invoice_csv / create_bulk_invoices — up to 100 rows
• list_tax_jurisdictions — supported VAT/GST/sales tax guidance

Read tools work with read-scoped keys. create_client, create_invoice, send_invoice, and bulk create require explicit write scopes on the API key — same as REST. If a user asks Claude to send an invoice without invoices:write, the tool returns a clear API error rather than silently failing.

Setup: Claude in the browser (five minutes)

Create a Settlematic account and navigate to Settings → API keys. Generate a key named for its purpose — 'Claude Finance Read' is better than 'key 3'. Start with scopes: clients:read, invoices:read, reports:read. Copy the MCP connector URL shown on the keys page.

In Claude: Settings → Connectors → Add custom connector → paste URL → save. Start a new chat (connectors do not retroactively attach to old threads). Enable the Settlematic connector in the tools menu. Ask a read-only prompt first: 'What is our total outstanding AR?' Validate the number against your dashboard. When comfortable, mint a second key with invoices:write and clients:write for drafting workflows — keep read and write keys separate for blast-radius control.

Setup: ChatGPT Custom GPT

ChatGPT does not use MCP natively today; it uses OpenAPI actions. In GPT Builder → Configure → Actions → Import from URL, paste the OpenAPI document URL from Settings → API keys. Set authentication to API key header (Authorization: Bearer your_key). Test with list endpoints before enabling create_invoice. Document which GPT in your org has write access — treat it like a shared service account.

The user journey differs slightly: ChatGPT action calls hit REST directly, while Claude MCP calls go through the MCP tool layer with richer descriptions for each parameter. Functionally, permissions and data are identical because both authenticate with the same API key model.

User journey: adding a client — dashboard vs MCP

Dashboard path: Sign in → Clients → New client → enter name, email, company, address, tax ID, notes → Save → navigate away. Roughly eight field interactions plus two page loads. Fine when you are already in the app with a CRM export open.

MCP path: 'Add client Marcus Webb, marcus@studio.io, company Webb Design Ltd, UK VAT ID GB123456789.' Claude calls create_client with structured fields, returns client ID JSON. You never left the conversation. Roughly one sentence, one confirmation if you configured the assistant to confirm writes.

UX shift: context stays in language. The assistant can immediately follow up — 'Create their first invoice for $6,000 due in 30 days' — without you re-finding the client row. Risk shift: typos in email addresses become model errors; mitigate with read-back confirmation before create_client on write keys.

User journey: creating an invoice — dashboard vs MCP

Dashboard path: Invoices → New → select client → add line items with qty, unit price, tax, discount → set dates and allowed assets → preview PDF branding → Save draft → Send. Best when the invoice has six line items, mixed tax rates, and partial payment enabled — visual editing wins.

MCP path: 'Create a $3,500 invoice for nexus@digital.com — description March retainer, Net 14, USDC and ETH, draft only.' create_invoice resolves client by email, pulls org defaults for notes and assets, sets due date from netDays, returns invoice number, paymentUrl, and pdfDownloadUrl. You paste the payment link to the client from chat or call send_invoice after explicit approval.

Time comparison from merchant interviews: simple single-line invoices average 4–6 minutes in dashboard vs 30–45 seconds in MCP for experienced users. Complex multi-line invoices still favor dashboard or bulk CSV — MCP create_invoice today optimizes for single-line speed, not full spreadsheet parity.

User journey: AR review — where MCP clearly wins

Dashboard path: Dashboard → note outstanding total → Reports → set date range → export CSV → filter overdue in spreadsheet → cross-reference client names. MCP path: 'List overdue invoices' → list_invoices with statusFilter OVERDUE → 'Which of these are over $10k?' answered in-thread without export.

Follow-up questions compound value: 'Summarize payments received this month by chain' calls get_payments_received with current month bounds. 'Show me a visual tax report for Q2' calls get_tax_report with includeVisualReport true and returns a shareable reportViewUrl with charts — finance forwards the link to a CPA in the same chat window.

This is where user journey changes most: from navigation-heavy exploration to iterative questioning. The dashboard remains authoritative for clicking into invoice INV-1042 to see timeline events; MCP is superior for aggregation and narrative summaries.

Side-by-side comparison table (mental model)

• Simple invoice create: MCP faster; dashboard better for multi-line tax complexity
• Client onboarding: MCP faster when details arrive in email prose
• Branding preview: dashboard wins — live PDF preview with logo and colors
• Sweep destination edits: dashboard only — treasury policy stays visual
• Bulk 50+ invoices: CSV upload or create_bulk_invoices — not manual dashboard clicks
• Payment status drill-down: dashboard timeline UI; MCP get_invoice for hash and URL
• Board-ready charts: MCP get_visual_report link; dashboard reports for interactive filters
• Audit / compliance: dashboard for immutable click path; MCP logs via API access logs

Governance: least privilege for AI operators

Treat MCP keys like service accounts. Read key for founders and advisers. Write key for ops lead only. Never embed write keys in public Custom GPTs. Rotate on schedule. Revoke instantly from Settings if a laptop is lost.

Instruct assistants to confirm before send_invoice — the tool description says so, but system prompts help: 'Always show draft totals before sending.' Claude's connector model asks for user approval on sensitive tool calls when configured; combine platform defaults with your org policy.

Org isolation is non-negotiable: one key, one tenant. Assistants never cross organizations. Same model as REST and webhooks — MCP is not a backdoor around RBAC.

Bulk operations from chat

Month-end agencies sometimes need forty similar invoices. Dashboard CSV upload handles this at scale. MCP offers validate_bulk_invoice_csv then create_bulk_invoices for up to 100 rows when the assistant already has tabular data in context — for example, pasted from a sprint spreadsheet. Workflow: get template spec → validate → create. Clients must exist beforehand (match by client_email); create clients in a prior step if needed.

UX note: chat-based bulk is powerful and dangerous. Require human confirmation showing row count and total amounts before create_bulk_invoices executes. We return per-row errors without partial silent failure so you can fix and retry.

Tax and reporting workflows via MCP

Tax reviews are conversational by nature. 'What did we collect at 20% VAT in Q2?' triggers get_tax_report. The response includes totals by rate, per-invoice breakdown, and optionally reportViewUrl — an interactive page with bar and doughnut charts your adviser opens without Settlematic login. For filing prep, get_tax_report_csv returns filename and raw CSV text to save locally.

Important E-E-A-T boundary: MCP output assists analysis; it does not file taxes. Human sign-off remains mandatory. Read-only reporting keys are appropriate for external accountants you embed in Claude projects.

When to stay on the dashboard

Use the dashboard for branding setup (logo upload, hex color), sweep destination configuration with cooldown policies, conversion flow graph editing, team invites, and TOTP security settings. These surfaces are intentionally visual and infrequent — poor fits for chat.

Use the dashboard when training new hires who need spatial memory of where invoice statuses live. Use MCP when veterans perform repetitive CRUD and reporting. Most mature teams use both in the same day.

Failure modes and how to debug

• 401 errors: key revoked or wrong connector URL — regenerate and update Claude settings
• 403 on create: missing invoices:write — upgrade scopes or use read-only key intentionally
• Client not found by email: create_client first; MCP does not fuzzy-match typos
• Stale connector in old Claude thread: start new chat after enabling tools
• ChatGPT action schema drift: re-import OpenAPI after platform upgrades

Security model: same API, same enforcement

MCP requests authenticate with Bearer API keys — identical to curl and SDK calls. Rate limits apply per key. Session cookies from the browser dashboard do not leak into Claude; the connector never sees your TOTP. Each tool invocation logs server-side with org ID for audit. When security teams ask 'is MCP a shadow IT API,' the honest answer is: it is the same API with a conversational front end — govern keys accordingly.

We recommend naming keys after use case and owner on-call rotation. Example: claude-read-finance@settlematic.com documented in internal wiki with rotation every 90 days. Write keys require manager approval in change-control tickets for regulated merchants.

Example day-in-the-life: agency ops lead

9:00 — Claude with read key: 'Outstanding AR and overdue list.' 9:05 — Slack message to account exec with three client names. 9:20 — Dashboard: fix complex milestone invoice with five line items. 10:00 — Claude with write key: 'Create draft $2,800 invoice for client@email.com, design sprint, Net 30.' Review JSON preview. 10:02 — Approve send_invoice. 11:00 — Webhook logs show payment.confirmed from earlier retainer; ERP updates automatically. No MCP for sweep edits — treasury already configured.

This pattern keeps high-stakes visual work in UI and high-volume language work in chat. Teams that force 100% chat or 100% dashboard both complain; hybrid teams report lowest operational friction in our onboarding surveys.

The hybrid operating model we recommend

Monday morning: MCP read session — overdue summary, payments last week, forward visual report link to leadership. Tuesday: dashboard — edit complex invoice, adjust branding. Wednesday: MCP — create three draft invoices from Slack requests. Thursday: webhooks fire payment.confirmed to ERP automatically. Friday: MCP tax CSV export for adviser.

AI does not replace billing policy. It reduces friction between intent and API calls. Settlematic built MCP-ready infrastructure because finance teams already speak in sentences — we just finally let the software listen without forcing a form for every sentence.

Start read-only this week. Add writes when confirmations are documented. Keep the dashboard one click away for the moments visual precision matters. That hybrid is the user journey change worth adopting — not replacing your entire UI with a chat box, but meeting operators where they already work.