Full walkthrough
The full product and workflow walkthrough, self-hosted as MP4 for public playback.
Measured proof
Latest release: v0.2.0-beta.16
One controlled build task is enough to show how Kiwi Control changes agent workflow quality.
This proof page is built around one controlled greenfield A/B run of the same task on the same machine. Repo A used Claude Code directly. Repo B used Kiwi Control before implementation. Both runs captured direct Claude JSON usage data.
This is measured proof for one controlled run, not a universal benchmark.
37.3%
lower Claude cost
24.3%
fewer Claude turns
59.0%
lower Claude wall-clock time
Watch the demo
Video-first product proof
The main demo is self-hosted on this website as a browser-safe MP4. Raw `.mov` files remain linked as fallback evidence.
Terminal-heavy run
The terminal-heavy cut showing the command flow more directly.
Measured A/B results
Repo A vs Repo B on the same task
Absolute values are shown below. Lower is better for every metric in this table.
Claude cost (USD)
Repo A
1.386074
Repo B
0.869697
Claude turns
Repo A
37
Repo B
28
Output tokens
Repo A
27,974
Repo B
10,958
Cache read tokens
Repo A
1,301,005
Repo B
1,100,949
Cache creation tokens
Repo A
153,617
Repo B
99,989
Claude wall-clock time
Repo A
432s
Repo B
177s
Methodology
How we measured it
Same task
Both repos implemented the same Markdown Notes Organizer scope on the same stack.
Same machine
Both runs happened sequentially on the same laptop and local filesystem.
Repo A
Claude Code only, with no Kiwi workflow help before implementation.
Repo B
Kiwi Control status, guide, graph, pack, and review were captured before implementation.
Direct usage data
Claude metrics came from Claude’s own JSON output. `ccusage` was not installed.
Scope of claim
This is one measured controlled run, not a universal benchmark.
Inspect raw evidence
Everything linked here is source-controlled
Comparison bundle
Per-repo metrics
Raw Claude JSON
Repo B Kiwi outputs
Try Kiwi Control
Inspect the proof, then install the live beta.
The proof page is here to make the product claims falsifiable. The downloads page is where you verify the current release surface.