14
ERP Tools Over MCP
The agent's six ERP tools served over the Model Context Protocol, read-only, with results redacted before they leave.
What it measured
The full test split, replayed with every ERP tool served over MCP, hit all 158 recordings and scored the same 155 of 158, with no row different from the in-process run.
Measured 2026-10-12 · the source is private, so this cannot be checked yet
Done when
Any MCP client can connect to the server from a fresh clone, and the source is public.
The tools started as Python functions called in-process. Moving them behind an MCP server raised one question: does the agent still see exactly the same thing? A test answers it for every development invoice by running each tool call both ways and requiring identical text.
The server holds no session state. It only reads the ERP, and its results pass through the same redaction and tripwire as any other request. A server failure marks the scenario as not run, the same as a provider outage, so it can never be scored as a decision.