Operate Settlematic from Claude with MCP: a practitioner's guide

Finance teams should not context-switch between chat, spreadsheets, and billing software for routine questions. 'What is our outstanding AR?' should not require five clicks and a CSV export. Settlematic's REST API — and MCP connector for Claude Desktop — let assistants read and act on invoicing data with the same scoped permissions you would give an engineer or contractor.

Model Context Protocol in plain language

MCP is Anthropic's standard for connecting AI assistants to external systems. Instead of pasting invoice JSON into chat, Claude calls tools: list invoices, get client details, summarize dashboard KPIs, create draft invoices. Settlematic maps your org's API key to that tool surface and enforces scopes server-side.

Claude interprets intent; Settlematic enforces permissions. The assistant never bypasses your org isolation or key scopes.

Example prompts that work today

• What is our total outstanding accounts receivable?
• Create a $3,500 invoice for Nexus Digital — Net 14, USDC and ETH allowed
• List overdue invoices for clients in Europe
• Has invoice INV-1042 been viewed or paid?
• Summarize payments received this month by chain
• Which clients have the highest concentration of open balance?

Setting up API access for Claude

Generate a scoped API key in Settlematic settings. Start read-only: invoices:read, clients:read, reports:read. Validate answers against your dashboard for a week. Add invoices:write when your team is comfortable with draft-and-confirm flows.

Add the Settlematic MCP server to Claude Desktop with that key. Document key ownership — 'Finance Automation — read only' is better than an unnamed key shared in a wiki.

Confirmation before writes

Read operations return immediately. Creates and sends show a preview and ask for explicit confirmation — Claude should not email a client a wrong amount because of ambiguous phrasing. Treat write scopes like production database credentials: few people, named owners, rotation schedule.

Governance for AI-assisted finance

• Separate keys for read vs write automation
• Org isolation — assistants only see your tenant data
• Same permissions model as REST, webhooks, and dashboard roles
• Works with Claude Desktop, custom GPTs, internal Slack bots, and CI jobs
• Revoke keys instantly from settings if a laptop is lost

Beyond Claude: one API, many connectors

MCP is one connector, not a walled garden. The same REST surface powers Zapier middleware, custom GPT actions, Monday morning AR bots in Slack, and data warehouse syncs. Start with conversational read access; graduate to webhook-driven automation when volume justifies it.

Teams that succeed treat AI assistants as junior operators with least privilege — not as magic that replaces approval policy.