Shapes Workbench

The design-time surface where you author the shapes that govern your graph, exposed as an MCP server you connect to your AI client.

Penumbra has two surfaces, and they cascade. You design your domain in the Shapes Workbench, then you operate it with the SDK. This tab is the first half.

The Shapes Workbench is a design-time MCP server. You connect it to an AI client (Claude Code, Cursor, Codex), and author your domain model conversationally: define the types, properties, and relationships that structure your graph, compile them, and materialize them to a project. Everything the runtime reads is governed by what you design here.

Design time vs. run time

	Shapes Workbench (here)	SDK
When	Design time	Run time
Surface	MCP server, in your AI client	TypeScript client
You	Author the shapes that structure the graph	Operate a graph governed by those shapes
The question	What does my domain look like?	How do I put knowledge in and get it back out?

If you are deciding what your knowledge looks like, you are in the Workbench. If you are capturing, searching, and reasoning over it from an app or agent, you are in the SDK. The two are meant to be used together: design a shape here, then read and write against it with the SDK.

The loop

1. Connect

   Add the Workbench MCP server to your AI client. See Connect.

2. Model your domain

   Open a session and define types, properties, and relationships. See Authoring.

3. Compile and materialize

   Compile the model and materialize it to a project, so the runtime can use it. See Materialize.

4. Operate with the SDK

   Extract, capture, and search against the shape you just designed, with the SDK.

Next

Connect the MCP

Endpoint, key, and client config for Claude Code, Cursor, and Codex.

Author shapes

The workbench tools: open, model, compile, diff, version, preview.