Loading ToolBench
Two MCP servers: signal-processing and geometry analysis of 3D mesh dumps, renders and telemetry, plus local Blender/MPFB body-mesh generation
Server has 15 tools with significant quality gaps. First 6 tools (bodymesh suite) have detailed descriptions (100-250 chars) and reasonable schemas with typed parameters and enums. However, tools 7-15 (video/imaging/perceptual/geometry analysis) have minimal descriptions (7-40 chars) with NO visible input schemas in source. Bodymesh tools average ~65; analysis tools average ~25. Overall composition shows two distinct server implementations with vastly different quality levels. No tool annotations (readOnlyHint/destructiveHint) present. Error handling guidance absent across all tools.
Analyze photometric consistency in video
Compare wavelet signatures for spatial multiresolution analysis
Create one immutable MPFB candidate in a headless Blender subprocess. Macro and detailed target values are bounded to [0,1]; target names must come from list_bodymesh_parameters. The result includes neutral Blender/OBJ/DSP-PLY artifacts, silhouette renders, an exact 55-bone arms-down body_engine.glb, and validated character_bake_cli JSON with skin weights/tangents/clips. Repeated calls with revised values create new candidates for an LLM-driven render/compare/adjust loop. This is parametric fitting, not one-shot photo-to-3D reconstruction.
Evaluate occlusion boundaries in video
Evaluate perceptual similarity using FR-VQA
Evaluate spatiotemporal frequency characteristics of video content
9 analysis tools (tools 7-15) have minimal descriptions (<40 chars) with no visible input schemas in source code.
No tool annotations present (readOnlyHint, destructiveHint, idempotentHint). Bodymesh tools 3-4 and 6 are marked WRITE but lack destructiveHint annotation in schema. LLM cannot infer irreversibility without explicit annotation.
Error handling guidance absent. No tool descriptions explain what to do on failure, which errors are retryable, or how to recover. Bodymesh tools reference 'BodyMeshError' in CLI but MCP responses lack recovery hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
Return compact reference and candidate status for one MCP-owned body-mesh job.
Inspect the local Blender 4.2 executable and installed body add-ons without creating a mesh. Returns the resolved direct blender.exe (never the launcher), MPFB/MB-LAB versions, engine skeleton/bake availability, local data directory, allowed reference roots, and blocking warnings. MPFB 2.0.x is the supported generator; MB-LAB is reported but not invoked.
List MPFB's bounded macro controls and optionally search installed detailed target files. Call without query for the macro parameter semantics. With a two-character-or-longer query (for example 'waist', 'shoulder', or 'leg'), returns at most limit safe relative target names accepted by create_body_mesh. This does not infer values from photographs.
Copy consented body reference images into an isolated local job. Inputs must be beneath BODYMESH_INPUT_ROOTS. Front and explicit rig_sex (male/female) are required; calibrated side and known height materially improve fitting. Optional binary masks should normally come from the DSP MCP's segment_image tool and must match their source image dimensions. The tool normalizes files to PNG, records hashes, and returns a job_id; it does not upload images or run Blender.
Render the fixed identity-v1 portrait/body set for one completed candidate. Produces eight near-frontal closeups with deterministic expression/light variation and front/three-quarter/side body views. The source character.blend is resolved internally, remains unchanged, and outputs are returned through a separately versioned manifest.
Score bake quality for geometry
Verify camera projection consistency
Verify identity coherence
Verify motion consistency across video frames
Analysis tools (7-15) lack parameter descriptions entirely. Tool names alone (e.g. 'evaluate_spatiotemporal_frequencies') do not convey what inputs are required or what the tool returns. LLM cannot determine when to invoke these tools.
Output schemas not documented for any tool. Bodymesh tools return complex nested structures (job_id, candidate_id, manifests) but LLM has no schema to understand response shape. Analysis tools return unknown structures.