dcc-mcp-shogun

A typed, local-first DCC-MCP adapter for Vicon Shogun Post motion-capture inspection, timeline control, bounded cleanup, processing, and file workflows.

The adapter uses Vicon's official local ViconShogunPost control-stream SDK and its Scene, Timeline, Offline, Channel, FIRFilter, and WeightedAverageFilter contracts. It does not expose arbitrary Python or HSL execution. Application controls not exposed by the SDK remain an explicit, exact-window DCC UI Control fallback.

Shogun Post ships an external Python SDK. The adapter runs in its own Python process and connects to the application's local control stream; it does not rely on a general-purpose Python interpreter embedded in the Shogun Post UI.

Capabilities

• inspect path-redacted scene metadata and bounded subject lists;
• inspect subject markers, skeleton hierarchy, constraints, and static subject parameters;
• query one marker trajectory value or an inclusive window of at most 2,000 frames;
• inspect the capability-gated Scene object graph, hierarchy, transforms, attribute/channel names, bounded channel samples and gaps, expression-safe labeling/solving setup parameters, rigid-body transforms and attached markers, optical/video-camera calibration and presentation summaries, selection, visibility, selectability, and opacity;
• inspect and explicitly change current frame, selected time ranges, range selection derived from keys, play/animation ranges, and playback through the capability-gated Timeline interface;
• inspect bounded Clip timing, offsets, time scale, lock and SMPTE alignment; read or explicitly change the active Clip; and update an allowlisted Clip timing subset with vendor read-back verification and rollback;
• inspect Character frame bounds and shot-QA flags, and update only six allowlisted Boolean workflow fields with vendor read-back verification and rollback, while excluding artist identities and free-form notes;
• repair one explicit marker sample with read-back verification; select channel keys from the current ranges, delete one explicit key or selected keys, and apply bounded FIR or weighted-average filtering that defaults to selected keys;
• inspect a stable allowlist of processing settings and invoke reconstruct, ROM labeling, subject calibration, auto-label, occlusion fixing, solve, QuickPost, or retarget through the official Offline interface; reconstruction, occlusion, and solving setting updates are allowlisted and rollback-aware;
• expose capability-gated import, save, and export calls that map directly to the official SDK;
• register one Shogun Post GUI instance with the DCC-MCP local gateway.

The mutating surface is deliberately narrow. Clip mutation cannot create, remove, rename, or reparent NLE objects, and Character mutation cannot read or write artist identities or free-form notes. Both mutation paths validate every field before connecting, verify the vendor read-back, and attempt rollback on partial failure. Cleanup never exposes the SDK's unbounded DeleteAllKeys operation. Single-sample writes are finite and bounded, return the previous value, and fail if Shogun does not return the requested sample. Processing requires an explicit current_frame or selected_ranges scope; the complete play range is not an available implicit default. The adapter does not expose scene replacement, arbitrary HSL/Python execution, or bulk raw trajectory writes. Existing outputs fail closed unless overwrite=true, and public results omit full file-system paths. The Scene surface does not expose object creation, removal, reparenting, setup-parameter creation/expression edits, or raw attribute writes. Attribute values, setup expression source, camera device identifiers, firmware, and capture/video file paths are intentionally omitted from inspection results.

The subject, marker, skeleton, constraint, parameter, and trajectory queries in shogun-scene are live-validated against Shogun Post 1.19. The newer official Scene object model is live-validated for scene-object, rigid-body, and video-camera inventories in an initialized 100-frame blank session. A fresh zero-frame placeholder session rejects those same official object-list commands with ControlError; the adapter reports that host-state boundary instead of claiming an empty inventory. Individual object-detail operations remain scene-state gated. A live 1.19 host also rejected the SDK's ImportFile call with ControlError, without partially changing the scene. The separate shogun-files Skill therefore remains explicitly host-build and license-capability gated: its tools are typed and fail closed, but import, save, and export are not claimed as live-supported on 1.19.

The setup-parameter, rigid-body, and video-camera extensions use the official 1.19 object types and are schema- and contract-tested. Wheel-installed and public-PyPI adapters both registered all 27 Scene tools in live 1.19 hosts. The three inventory calls completed in the initialized blank session; in a separate zero-frame placeholder they returned only bounded ControlError results, and follow-up typed inspection confirmed that the exact host remained available. Successful parameter and detail reads from a disposable non-empty scene are not yet claimed.

The same 1.19 host exposes the official Timeline and Offline Python classes but rejects their commands as invalid for that host application. The shogun-timeline and shogun-processing Skills therefore remain explicitly capability-gated: their schemas and SDK mappings are tested, while 1.19 support is not claimed. A rejection is returned as a bounded typed error and never triggers UI automation.

shogun-editing is separately capability-gated. Its schemas and SDK mappings are tested against the 1.19 SDK contract. All five tools were also loaded and dispatched through a live 1.19 host: against an empty scene, each mutation was safely rejected as a bounded ControlError, with the scene remaining empty. Trajectory write and filter effects on a non-empty disposable take are not yet claimed as live-validated. A host-build or scene-state rejection never falls back to arbitrary script execution.

shogun-production-context exposes eight Clip and Character tools. Its four read-only tools were loaded and dispatched through a live 1.19 host; the blank scene returned bounded ControlError responses and the host remained available. All eight current tools were then loaded from the packaged Skill in a live 1.19 adapter; get_active_clip reached the official Scene interface and returned the same bounded ControlError on the blank scene without affecting host availability. The active-Clip, Clip-timing, and Character-QA mutation contracts are SDK-mapped, schema-tested, read-back verified, rollback-aware, and remain capability-gated until a disposable non-empty scene is available for live mutation evidence. The official Database interface is intentionally not public: an isolated read probe coincided with host termination, so it remains deferred until that stability signal can be reproduced and resolved.

The implemented contracts follow Vicon's official Shogun Post documentation, Python scripting guide, and HSL command reference. See the maintained official SDK coverage matrix for implemented and intentionally deferred interface families.

Showcase

The repository includes an original, deterministic 240-frame BVH motion source and generator under examples/showcase. It is intended for reproducible import experiments without redistributing production capture data.

See docs/showcase.md for the full launch, discovery, inspection, and evidence workflow.

Local development

Start Vicon Shogun Post with your normal application launcher. Then install this adapter in a Python environment that can import dcc-mcp-core:

python -m pip install -e ".[dev]"
dcc-mcp-shogun --host-pid <SHOGUN_POST_PID>

The adapter discovers the official SDK beside the selected host process. An operator can instead set DCC_MCP_SHOGUN_SDK_PATH to the SDK's Win64 directory. The adapter resolves the selected host process's control-stream listener; DCC_MCP_SHOGUN_CONTROL_PORT may override it only when that port is owned by the same host process.

The DCC-MCP listener uses an OS-assigned loopback port unless --mcp-port or DCC_MCP_SHOGUN_PORT is explicitly set. Use dcc-mcp-cli list, search, describe, and call rather than storing the resolved endpoint.

Validation

python -m ruff check src tests tools
python -m ruff format --check src tests tools
python -m pytest
python tools/lint_skills.py
python -m build

Privacy and safety

• Connections are limited to the local Shogun control stream.
• Tool results reduce file paths to base names.
• Errors report exception classes, not SDK install paths or machine details.
• Attribute inspection returns names only, and camera summaries omit device IDs.
• The scene Skill limits mutation to selection and display state; the file Skill exposes only bounded import/save/export operations; editing requires an explicit object/channel or subject/marker/frame; processing never defaults to the full play range.
• Authentication, licensing, UAC, and security dialogs are never automated.

Vicon and Shogun are trademarks of Vicon Motion Systems Ltd. This independent adapter is not affiliated with or endorsed by Vicon.

License

MIT