Let your coding agent feel your game.
Kite runs your game, records every frame, and measures how it feels — input latency, jump curves, coyote windows — in a vocabulary your agent understands. Say “the jump feels floaty” and it has the numbers to act on. No guessing.
The closed feel loop
Agents can't hold a controller. They can read numbers. Kite turns game feel into numbers and closes the loop.
Agent edits code
Claude Code or Cursor changes your movement code: gravity, curves, state machines.
Kite runs the game
Headless, with scripted inputs. Same script, same frames, every single time.
Telemetry becomes a report
Velocity arrays, frame timings, and state transitions turn into a structured feel report.
Agent converges
The agent reads the report, sees the gap, and re-tunes. Repeat until it feels right.
The numbers your agent reads
Real output from kite run. Named, objective measurements your agent maps your feel-words onto. Here, checked against an optional snappy target — a gradient, not a vibe.
Instrumentation over intuition.
Response curves
Jump arcs, acceleration ramps, decel easing. Measured curves plotted against the curve you wanted.
Frame-true latency
Frames between input, state change, and first movement on screen. Spikes get flagged per run.
Feel archetypes
Every run gets scored against named feel profiles. “Make it snappier” becomes a number. No golden baseline needed.
Roadmap
Godot, open source
Recorder addon, deterministic input replay, CLI + MCP server, platformer feel metrics. Free and MIT.
Unity
C# adapter on the same telemetry contract. The analysis core never touches engine code.
Kite CI for teams
Feel regressions caught on every pull request. Dashboards where teams set a numeric feel target before merging.