Start with what the controller will give you, because that decides everything downstream. A Haas exposes MTConnect natively on recent control generations. A Fanuc needs FOCAS — either a licensed library talking to the control, or an MTConnect adapter in front of it. A Siemens 840D is reachable over S7comm or the OPC UA server on newer firmware. Okuma, Mazak, DMG Mori and Brother each have their own combination, and most modern controls offer OPC UA as well.
The values worth collecting are few and consistent: spindle load (rising at constant feed and speed means a dulling tool; a step change means a collision), spindle speed and feed rate for context, execution state for utilisation, part count for throughput, program name so anything you measure can be attributed to a part number, and alarm code.
Getting it out of the plant is a network conversation more than a technical one. Keep the collector on the machine network and push outward — never expose a control to the internet, and never let anything that can write sit in the path. Read-only protocols make that argument easier to win.
The part that surprises people is the fleet problem. Monitoring one machine is a weekend. Monitoring seven machines from four vendors means seven tag dictionaries, and a threshold rule that has to be rewritten per controller. Once the values are canonical — spindle_load_pct means the same thing on all seven — one rule covers the floor, and adding an eighth machine does not reopen the work.
A Haas with a recent control, because MTConnect is built in and read-only — no library licence, no engineering project, and a security posture the plant will accept.
For direct access, yes — FOCAS is the library that talks to the control. The common alternative is an MTConnect adapter that speaks FOCAS on one side and MTConnect on the other.
Spindle load. At constant feed and speed its trend is tool condition, its step changes are collisions, and its absence is downtime.