ledgerflow-mcp

The problem

Freelancers and small creative teams often track incoming invoices, bills, and payment dates across email, spreadsheets, and payment apps. That makes it hard to answer simple questions: what is still owed, what is due next, and what might cash look like over the next few months? Expensive accounting suites can also be too much for a one-person business.

The solution

LedgerFlow gives those teams one small, understandable finance workspace. Users can import an invoice, review and correct its extracted fields, organize receivables and payables in an editable ledger, and view 30-, 60-, and 90-day cash-flow projections. Overdue or unusual invoices are highlighted for review. The app can prepare payment actions using configured sandbox credentials; it does not move real money in this demo.

How it is built

• React, TypeScript, Vite, and AG Grid power the dashboard and editable invoice table.
• FastAPI and Pydantic provide the API, document intake, forecasting, trace display, and payment adapters.
• SQLite is the local default. An optional MongoDB adapter supports shared document storage and vector search; its built-in vectors are lexical features, not neural embeddings.
• Invoice extraction uses a deterministic local parser by default, with optional Evolus and OpenAI-compatible model providers.
• Forecasting tries TabPFN when configured, then scikit-learn, then a deterministic baseline.
• Four MCP tools expose invoice extraction, runway forecasting, vendor anomaly review, and sandbox payment-link creation.

What is working in the demo

The recording uses a fresh local SQLite database with five sample invoices. It shows the dashboard totals, invoice ledger, forecast view, anomaly review, and local trace. Optional cloud providers and payment credentials are not enabled, so the demo does not claim to have used them.

Business model

The initial customer is a freelancer or tiny service business that wants more visibility than a spreadsheet provides without adopting a full accounting suite. The business-model hypothesis is a free local solo tier, with a paid hosted tier for teams that want shared workspaces, automated imports, integrations, and longer history. A higher-priced accountant or studio tier could add multiple client workspaces and reporting. Prices and willingness to pay have not yet been validated; the next step is to interview freelancers and test whether they will pay for less manual reconciliation and clearer cash planning.

Scalability

Today, the default build is a single local service using SQLite. That is appropriate for an individual demo, but it is not yet a production multi-user deployment. The storage adapter already has an optional MongoDB path, and the service separates the API, forecasting, extraction, and payment providers. A production version would use a shared durable database, object storage for documents, a job queue for long-running extraction and forecasts, and horizontally scaled API workers. MCP makes the core actions reusable from compatible clients without duplicating the finance logic.

What is next

Validate the workflow with freelancers, improve extraction review and duplicate handling, add secure team access and durable hosted storage, and verify provider integrations against test accounts before enabling any live financial action.

Built With

• ag
• airwallex-sandbox-(optional)
• fastapi
• model-context-protocol-(mcp)
• mongodb-(optional)
• paypal-sandbox-(optional)
• python
• react
• scikit-learn
• sqlite
• tabpfn-(optional)
• typescript
• vite