A scout that only reads what is in the demo.
prefire turns CS2 demo files into a call sheet on the radar: where they smoke, where they stand, where they die, counted from the ticks. AI is only allowed to phrase a fact that was already counted.
Why it exists
Stat trackers are good at telling you what the numbers were. They are not built to tell you the single habit that produced them. prefire starts from the other end: a deterministic engine walks the match tick by tick, counts a specific pattern, and the AI is only allowed to phrase that already-counted fact in plain language.
That constraint is the whole point. The model never sources a number, never rounds one, and every claim on a sheet carries a tier: Counted, Derived or Read.
How your demo is handled
The .dem file is uploaded over HTTPS and parsed once. The raw file is deleted the moment the parse finishes; what stays is the numbers and one replay per round. Nothing else is collected: no chat logs, no voice, no account data beyond what a FACEIT or Steam sign-in returns.
Delete a demo and its rows and clips go with it. Your demos are never resold and never used to train a model. The parser is @laihoe/demoparser2, the same native library used by CS Demo Manager and broadcast tooling.
Who is building it
prefire is made by a solo developer who plays CS2. It is early: the Major hub is free and open, Team Scout is the paid product, and Ask your own demos is still in build.
Questions, bugs or feature requests go to contact@prefire.team.