MCP server
MachineMetrics runs a Model Context Protocol server. Connected to your coding assistant, it lets the assistant read the GraphQL schema, build and run production and GraphQL queries against your live data, and search your Knowledge Hub. The Carbide plugin's skills rely on it to check the data model and component rules rather than guess.
Connect
In Claude Code:
claude mcp add --transport http machinemetrics https://agent.machinemetrics.com/mcp
The first use opens a MachineMetrics sign-in. Approve it once and the assistant works as you.
Tools
Fifteen tools are exposed. Six are the general MachineMetrics set — production and GraphQL queries, Knowledge Hub search — and are described on the support site. The other nine are Carbide Data, and they are the ones the plugin's skills lean on:
| Tools | What they are for |
|---|---|
listCarbideSchemas, getCarbideSchema | Find and read existing schemas |
validateCarbideSchema, diffCarbideSchema | Check a schema, and see what a change would alter |
publishCarbideSchema | Publish it |
seedCarbideRecords, validateCarbideRecord | Put test data in, and check a record against the schema |
generateCarbideQuery, executeCarbideQuery | Build and run a query against the data |
You rarely call these by name — /carbide:carbide-data and the other skills invoke them. They
are listed here because the support guide documents the six general tools only, so a developer
reading it alone would conclude the Carbide tools do not exist.
These tools appear only for an admin role. custom-data:schema, the scope that gates them,
is implied by the itadmin role in Dashboard and by nothing else. A developer without that role
never receives the scope and never sees the Carbide tools, however they authenticate: signing in
again, reconnecting, or using a different client will not add them. If the assistant can read
production data but the Carbide tools are absent, the role is what to check, not the connection.
Two sign-ins, one account
Carbide setup asks you to authenticate twice, for two different things:
| What it authenticates | Fix it with | |
|---|---|---|
mmdev login | The CLI, for scaffolding, OAuth clients, deploy and registration | mmdev login again |
| MCP connection | Your assistant, so it can read the data model and run queries | /mcp → re-authenticate |
Both use your MachineMetrics account and neither replaces the other. Scaffolding or deploy failing points at the first; an assistant that cannot see your schema points at the second.
Everything else
Setup for Claude Desktop and other clients, the six general tools in detail, what the assistant can and cannot see, where your data goes, and troubleshooting are on the support site: Connect an AI Assistant with MCP.