Private AI assistant over company data (MCP server)
Lets Claude Desktop or any MCP client work with a company's CRM, tasks and inventory through ten audited tools, with read only guardrails and a dry run before anything is written.
AI assistantsPythonMCP SDKSQLite backend (one class to swap for HubSpot, Procore or SAP B1)Claude Desktop, Claude Code
The problem it solves
Giving an AI assistant access to company data is easy and dangerous. The useful version is a small set of tools with annotations, a read only mode, validation, a dry run for anything that writes, and an audit line for every call.
What it does
Ten tools over contacts, deals, tasks, notes, inventory and sales orders, each annotated read only or write.
A read only mode removes the write tools entirely. Input validation and stock checks refuse bad requests with an error the model can act on.
Orders are priced and checked with dry_run true by default; nothing is written until dry_run is false.
Resources expose the schema, low stock and the pipeline summary; a prompt template produces a daily operations briefing.
Six end to end tests through a real MCP client; every call appends an audit line.
How it was tested
Search, get contact, dry run order, committed order, refused order on an out of stock SKU, pipeline resource. 6 of 6.
Screens
MCP Inspector connected to the server: ten tools listed, read tools tagged read only.Dry run preview: the order is priced and checked against stock but nothing is written.Guardrail: an order on an out of stock SKU is refused with a clear error the model can act on.Pipeline resource: deals by stage, open value and deals closing in 30 days, readable by the model.
Who it is for
Firms that want an assistant over their own data without handing it the keys.
Built and tested on sample data before any live credential is connected. Every system Finamatik ships is adapted to the client's own tools and owned by the client outright.