Declarative, reversible Niagara authoring surface for Unreal Engine exposed through Epic's Toolset Registry. Provides tools for observing, planning, applying, validating, previewing, and rolling back Niagara particle system modifications.
Tools have basic descriptions and risk classifications, but critical gaps undermine production readiness. All 7 tools accept a single 'RequestJson' string parameter with minimal schema definition, no typed fields, no parameter descriptions beyond 'JSON request containing...'. Input validation logic exists in C++ (ValidateDesiredIr, IsStableId checks) but is not exposed in the tool schema. Output schemas are undocumented. Error handling returns JSON with code/message fields, but descriptions lack recovery guidance. Tool names follow verb_noun pattern (niagara_observe, niagara_apply) but are domain-specific and lack clarity for general LLM audiences. No tool descriptions explain WHEN to use each tool vs. alternatives or what downstream tools expect.
Apply an approved plan inside a NiagaraMCP-owned sandbox.
Report runtime features without mutating editor state.
Export a canonical observation of an allowed Niagara system.
Produce a deterministic plan without writing assets.
Render bounded preview evidence without persisting the candidate.
Revert an owned pending request or restore an approved backup.
Validate a sandboxed candidate without persisting it.
All tools accept a single 'RequestJson' string parameter with no typed schema. The parameter description is generic ('JSON request containing...') and does not specify the expected JSON structure, required fields, or format. LLMs cannot infer valid input without explicit schema.
Output schemas are not documented. Tool descriptions do not specify what fields are returned, their types, or how downstream tools should consume the response. This forces LLMs to guess at response structure.
Tool descriptions lack context for LLM selection. Descriptions are 30-50 characters and do not explain WHEN to use each tool, what prerequisites exist, or how tools relate to each other. E.g., niagara_plan says 'Produce a deterministic plan' but does not explain that it must be called before niagara_apply.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
Error responses include code and message fields but lack recovery guidance. E.g., 'invalid_ir' error does not tell the LLM what to do next or how to fix the input. Descriptions should include examples of valid input or suggest alternative tools.
Tool composition is unclear. The workflow (observe → plan → apply → validate/preview → rollback) is implicit in tool names but not documented. Descriptions do not specify tool dependencies, required inputs from prior calls, or expected sequencing.