One Command Table — What the AI Can Do Is What You Can Do

Every feature of LensBench is a command in the command table. The same command is, at once:

• An action in the desktop app (in the command palette you fill in parameters while previewing the diff)
• A subcommand of the command line lb, such as lb paraxial my.lens
• A tool the AI client calls

What this means ​

• There is no "AI-only hidden channel". What the AI can do is exactly what you can do, and it goes through the same checks: a change that would introduce a new error is rejected outright, and every change goes onto the undo stack.
• Every step the AI took, you can reproduce. Which command it called and with what parameters, you can redo with the same command on the command line or in the desktop app.
• The documentation can't drift from the features. The command reference is generated automatically from the command table, and what the AI sees when it looks up help is the same document you see.

Any AI will do ​

LensBench connects through MCP (the common protocol by which AI clients connect to external tools): Claude Code, Claude Desktop, Cursor, Cherry Studio, Trae and others all work, and one command, lb mcp-config, generates or writes the configuration. When a client limits the number of tools (Trae, for instance), you can switch to the slimmed-down set of the 37 most commonly used tools. For each client's configuration see Native AI.

The AI can also "see" plots ​

Plots such as the layout, spot diagram and MTF are handed to the AI directly as images, together with the numbers page for the same data: the AI uses its eyes to look at the light path and the data to report numbers, never estimating numbers by reading a plot.

Security boundaries ​

• The AI can only read and write the project directory you specify.
• Anything written in the lens name or comments of a design file shared by someone else (even "ignore previous instructions…") is marked as data when handed to the AI, and is never executed as an instruction.