ICore Blocks / Features / iScript · agents

AI agents that edit the diagram

The diagram has a text form. That single fact is what lets a coding agent build and edit models for you, without a plug-in and without an API.

iScript — the diagram as text

Every diagram can be written as a short script. The interpreter evaluates these statements before the console's maths grammar, so their syntax is never mistaken for arithmetic:

blockA = block(Gain, Home)      // create a block, bind it to a handle
s      = subsystem(Home)        // create a subsystem and bind its block
connect(blockA<1>, blockB<0>)   // output 1 -> input 0
blockA.move(120, 40)            // position, +y is up
blockA.rename(Gain2)
blockA.resize(80, 60)
blockA.rotate(90)
blockA.commentOut               // exclude from the run

How an agent gets a script into the app

  • Script Runner (Code Engine → Script Runner) keeps scripts under <Project>/scripts/*.iscript with Run, Run All, New, Save and Delete. It stops at the first failing line, so a bad statement cannot half-build a diagram silently.
  • Import → Import as ICore Recipe… replays a recipe file into the level on screen.
  • The templates folder — drop a .iscript in and it appears in every template list on the next look. No restart; the catalog rescans each time it opens.
  • The command window — paste the lines in directly, or drive them headlessly with --console.
  • The terminal inside the app (Code Engine → Terminal) — a real shell, already in your ICore Blocks folder. Whatever coding agent you use runs here, and reaches the model the same ways you would — writing .iscript beside it, or driving the binary with --console. The canvas is a few inches from its output.
Code Engine → Terminal
A coding agent running in the app's Terminal panel, beside the attention stack it built on the canvas A coding agent running in the app's Terminal panel, beside the attention stack it built on the canvas
“Build me an Attention Stack!” — typed to Claude Code running in the app's own Terminal panel, which opens in your ICore Blocks folder. What came back is on the canvas beside it: Embed and position — a counter into an embedding lookup, positional encoding and layer normalisation — feeding an Attention subsystem of scaled dot-product attention, a softmax, and a display reading the weights. The agent is an ordinary command-line tool; the diagram is the model it wrote.

The round trip

An agent that can only write is working blind. generateRecipe — or Export → Export as ICore Recipe… — writes any level back out as iScript, so the agent can read what it just built, diff it, and correct itself. Script Runner's Import Diagram button does the same into a new script, which is the quickest way to start from something already drawn.

The copilot inside the app

Connect an Anthropic or OpenAI key and the copilot panel sees your live diagram in that same language, alongside the block catalog, and answers in it. Its access to the model is read-only — it describes and drafts; applying a script is your action, through the surfaces above. Your key is held in the system credential store.

See also: Python and MATLAB, both  ·  Block Wizard — build your own library

Get started

See it run on your own model.

Download the application from the customer portal, or read the documentation first — the manual, every block with its measured response, and the full command reference are public.