The onsystem tool checks AI code against your design system
In a test of 80 results, 9 were incorrect alarms, a rate of 11.2%, and an AI set the labels.
A developer, dgesteves, released onsystem, a tool that checks interface code that agents write. It compares the code with the design system of a team and finds incorrect colors and parts that the system does not have. A design system is the set of approved colors and parts of an interface. The tool gives a repair for each result, and it works on each edit.
How it works
The tool reads the parts and tokens (values for color) of a team from source code. It checks each edit that an agent makes and each pull request (a code change for inspection). It operates on your computer. It works with Claude Code, Codex, Cursor and other agents.
The numbers
In a demo, the tool found 11 problems in one draft. The TypeScript compiler found 3. On 12 open repositories it gave 14,539 results. In a sample of 80, 57 were correct, 14 were not clear and 9 were incorrect.
What is not known
An AI reviewer, Claude Opus 5.5, set the labels. A person did not. The sample is small. The range for the incorrect rate is 6.0% to 20.0%.
Insight: The check moves from the prompt to the pipeline
The source tells that text about a design system helps an agent, but only a check of the code holds. An AI also set the labels in the test of this tool. This reads as a move from text for agents to a gate that makes agents obey, and the team with the gate sets the limits. Leaders must find out who checks the checker.
Sources
Posted
