Cross-DCC Verification

Prove an asset produced by one DCC through its MCP server can be consumed, inspected, and validated by a second DCC. This page explains the contract dcc-mcp-core ships and how downstream DCC repositories plug into it.

Why a contract, not an implementation

Every DCC (Blender, Maya, Unreal, Photoshop, Houdini, ...) has its own native import API, its own scene traversal primitives, and its own idea of what counts as a "mesh". dcc-mcp-core stays out of that jungle on purpose: it only defines the shape of a verifier's result and the skill template every downstream implementation is expected to clone. Keeping the shape tiny (three fields) is what makes the contract portable.

The actual import_and_inspect logic lives in the downstream repos:

DCC	Repository	Skill name (convention)
Blender	dcc-mcp-blender	blender-fbx-verifier
Maya	dcc-mcp-maya	maya-fbx-verifier
Unreal	dcc-mcp-unreal	unreal-fbx-verifier
Photoshop	dcc-mcp-photoshop	photoshop-psd-verifier

The SceneStats contract

from dcc_mcp_core import SceneStats

observed = SceneStats(
    object_count=1,
    vertex_count=482,
    has_mesh=True,
    extra={"dcc": "blender-3.6"},  # optional DCC-specific enrichments
)

Field	Type	Meaning
object_count	int	Top-level objects seen after import.
vertex_count	int	Total vertex count across all mesh geometry.
has_mesh	bool	True when any imported object is polygon geometry.
extra	dict	Free-form enrichments. Survives serialisation, ignored by the core comparison.

Comparing producer vs verifier

The helper SceneStats.matches() implements the only comparison semantics core endorses:

• Strict on object_count (structural invariant).
• Strict on has_mesh (detects silent empty imports).
• Fuzzy on vertex_count (±5% by default) because FBX normals, UV seams, and tangent basis changes can split vertices on re-import.

produced = SceneStats(object_count=1, vertex_count=482, has_mesh=True)
observed = verifier_skill__import_and_inspect("/tmp/sphere.fbx")

assert produced.matches(observed, vertex_tolerance=0.05), (
    f"round-trip drift: expected {produced}, got {observed}"
)

Extra fields beyond the three core ones belong in extra. The comparator does not read extra, which keeps DCC-specific telemetry from ever breaking a round-trip assertion.

Deterministic capture responses

Capture implementations can use dcc_mcp_core.runtime.capture_contract to publish one bounded response shape:

from dcc_mcp_core import CaptureTargetSpec, build_capture_response

response = build_capture_response(
    frame.data,
    width=frame.width,
    height=frame.height,
    format="png",
    backend="mock",
    target=CaptureTargetSpec.active_viewport(),
    return_mode="artifact_only",  # image or both are explicit opt-ins
    unchanged_if_hash=previous_sha256,
)

artifact_only is the compatibility-safe default and returns an artefact://sha256/... reference plus bounded metadata (dimensions, bytes, sha256, format, backend, and target) without inline pixels. image adds one base64 image payload; both adds that payload alongside the artifact reference. A matching unchanged_if_hash suppresses image data and marks the response unchanged: true. Targets are explicit (dcc_window, active_viewport, scene_viewer, or a positive bounded crop); hidden or unknown targets fail closed.

Writing a verifier skill

Scaffold a skill with dcc-mcp-skills-creator, then wire it to dcc_mcp_core.SceneStats. To create a verifier for a new DCC:

1. Scaffold a new skill directory in your downstream repo (e.g. dcc-mcp-blender/skills/blender-fbx-verifier/).
2. Edit SKILL.md: set dcc: blender (or similar) and rename the skill to match your repo's convention.
3. Replace the stub body of scripts/import_and_inspect.py with your DCC's native import + inspection calls. Return a SceneStats.to_dict() payload wrapped by skill_success(...).
4. Add a CI job in your downstream repo that:
   • Boots two DCC processes (or one producer + one verifier).
   • Runs a producer skill to create and export an asset.
   • Runs your new verifier skill on the exported file.
   • Asserts produced.matches(observed) using dcc_mcp_core.SceneStats.

Example stub (Blender)

import bpy

from dcc_mcp_core import SceneStats
from dcc_mcp_core.skill import skill_entry, skill_success


def main(params):
    bpy.ops.wm.read_factory_settings(use_empty=True)
    bpy.ops.import_scene.fbx(filepath=params["file_path"])
    objects = list(bpy.context.scene.objects)
    meshes = [o for o in objects if o.type == "MESH"]
    vertex_count = sum(len(m.data.vertices) for m in meshes)

    stats = SceneStats(
        object_count=len(objects),
        vertex_count=vertex_count,
        has_mesh=bool(meshes),
        extra={"blender_version": bpy.app.version_string},
    )
    return skill_success("Imported and inspected", **stats.to_dict())


if __name__ == "__main__":
    skill_entry(main)

Where the round-trip CI lives

dcc-mcp-core does not ship a Blender-with-FBX round-trip job. That belongs in whichever downstream repo owns the producer or verifier binary, because only those repos pin the DCC version matrix. The invariant core asserts is shape-only, exercised by tests/test_verifier_contract.py.

Downstream repos following this contract should add a CI job along these lines:

1. Start producer DCC in headless mode, load producer skill.
2. mcpcall call: producer__create_sphere, producer__export_fbx.
3. Start verifier DCC in headless mode, load verifier skill.
4. mcpcall call: verifier__import_and_inspect(/tmp/sphere.fbx).
5. Python assertion: SceneStats.from_dict(...).matches(produced).

FAQ

Why only three fields? What about bounding boxes, materials, animations?

Every field that enters the core contract has to mean the same thing across every DCC, forever. Three fields (objects / vertices / has-mesh) is the set where we could rigorously define and test the semantics. Anything DCC-specific (materials, cameras, animation frame counts, bounding boxes) can travel through extra without committing core to a cross-DCC definition.

Can I extend SceneStats in a downstream repo?

Do not subclass. Put downstream-specific data in extra. If a field becomes genuinely universal across all supported DCCs we'll promote it to the core contract in a minor release.

What if my DCC doesn't expose vertex counts?

Return vertex_count=0 and set an explanatory note in extra. Round- trip tests that target that DCC should then pass has_mesh as the only real assertion.

Related

• dcc-thread-safety.md — main-thread dispatcher primitives that verifier skills rely on when running inside a live DCC session.
• host-adapter.md — the HostAdapter base class downstream repos extend to plug a verifier into their DCC's idle loop.
• skills.md — SKILL.md format the verifier template is built on.