Exposes Rhino control tools over the Model Context Protocol and forwards each call as newline-delimited JSON over TCP to the listener script running inside Rhino. Implements JSON-RPC over stdio for initialize, tools/list, and tools/call.
Pantograph defines 11 tools with explicit schemas and descriptions. Naming is verb-forward (graph_*, execute_*, get_*, capture_*) and generally clear. Descriptions are present and substantive (avg ~120 chars), explaining WHAT and WHEN. However, critical gaps exist: (1) parameter descriptions are sparse, many params lack type info or constraints (e.g., graph_connect's semantics param has no type declaration); (2) output schemas are entirely undocumented, no tool declares what it returns, forcing LLMs to guess; (3) error handling is absent, no guidance on recovery, retryability, or actionable error messages; (4) no tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (DESTRUCTIVE, IRREVERSIBLE tools exist). The graph_* tools are well-composed and domain-specific, but lack the polish needed for production agent use.
Capture a screenshot of the active Rhino viewport and return it as an image, so you can visually verify your modeling work.
ESCAPE HATCH — runs raw Python (rhinoscriptsyntax as `rs`) inside Rhino, bypassing the definition graph. Prefer the graph_* tools: raw code produces geometry nobody can edit. Use only for inspection or operations the op catalog cannot express.
Get the current Rhino document state: layers, object count, and a list of objects with their ids, types, layers and names. Call this before modifying existing geometry.
Add one node to the definition graph. Provenance is required: state which prompt clause this node answers and why it exists.
Clear the definition graph to start a new definition.
Wire an output port of one node into an input port of another. Rewires automatically if the input is already connected. Include `semantics` describing the dependency (e.g. 'twist grows with level').
No output schemas documented. LLMs cannot infer what fields to expect from any tool response, forcing them to guess downstream field names and risk broken tool chains.
Parameter descriptions missing or incomplete. graph_connect.semantics, graph_set_param.clause/reason, and many others lack type info or constraints. LLMs cannot validate inputs without explicit descriptions.
No error handling guidance. Destructive tools (graph_remove_node, graph_clear) and irreversible tools (execute_rhino_code) lack confirmation patterns or recovery hints. Agents cannot distinguish retryable from fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Compile the current definition graph to rhinoscriptsyntax, perform it in the live Rhino document (rebuilding the graph's layer), and capture the viewport so you can inspect the result. Compile or execution errors come back for you to repair.
List the op catalog: every operation a definition-graph node can be, with its params, input ports, and output ports. Call this before authoring a graph.
Read the current definition graph (nodes, params, edges, provenance) and its live validation issues.
Remove a node (and its edges) from the definition graph.
Change one param value on an existing node.
No tool annotations. Tools with clear risk levels (DESTRUCTIVE, IRREVERSIBLE, WRITE) lack readOnlyHint, destructiveHint, or idempotentHint annotations. Agents cannot reason about side effects.
Parameter type constraints missing. graph_add_node.params.value is described as 'number, boolean, or number[]' but lacks JSON Schema type union or enum. graph_set_param.value has no type at all.