Capabilities

GET /capabilities is how a client discovers what the server hosts. No request body. Response is a Capabilities object.

Response shape

{
  "protocol_version": "1.2",
  "rotation_format": "quaternion_xyzw",
  "coordinate_system": "right_handed_y_up",
  "units": "meters",
  "response_formats": ["gltf_2.0_json", "gltf_2.0_binary"],
  "models": [
    {
      "id": "kimodo-soma-rp",
      "fps": 30.0,
      "supports_retargeting": false,
      "supports_async": false,
      "supported_constraints": ["root_path", "effector_target", "pose_keyframe"],
      "supported_segments": ["text", "unconditioned"],
      "supported_guidance_types": ["nocfg", "regular", "separated"],
      "predicted_contact_joints": ["LeftHeel", "LeftToe", "RightHeel", "RightToe"],
      "native_clip_seconds": 10.0,
      "chunking": "stitched",
      "recommended_max_duration_seconds": 12.0,
      "canonical_skeleton": { "joints": [ ] },
      "limits": {
        "max_duration_seconds": 30.0,
        "max_num_samples": 16,
        "max_constraints_per_request": 64,
        "max_prompt_length": 1000,
        "max_request_bytes": 1048576
      }
    }
  ]
}

Model fields

Field	Type	Notes
id	string	Used as request.model
fps	number	Native frame rate of the model
supports_retargeting	bool	If false, requests MUST send canonical_skeleton verbatim
supports_async	bool	If true, server may return 202 Accepted for long generations (see Async)
supports_segment_seed	bool	If true, the model honours a per-segment seed
supports_loop	bool	If true, the model samples Options.loop as a seamless cycle (one segment)
supports_motion_reference_mixed	bool	1.2. true exactly when the model lists "motion_reference" in supported_segments (a reference composes like a text segment), false otherwise
supports_trajectory	bool	If true, every response sample carries its travel trajectory (MMCP_motion.samples[].trajectory, see Response)
supports_batch	bool	If true, POST /generate takes a batch body of up to limits.max_batch_size requests (see Endpoints)
supported_constraints	string[]	Constraint type names the model accepts
supported_segments	string[]	Segment types the model accepts. Defaults to ["text", "unconditioned"]; backbones with a specialized text-to-pose model add "pose", ones that take a motion as a prompt add "motion_reference" (1.2, see Segments), and ones that take a video as a prompt add "video_reference" (1.2, see Segments). A client SHOULD check this before sending a reference; a type not listed is refused with unsupported_segment
supported_guidance_types	string[]	Subset of ["nocfg", "regular", "separated"]
predicted_contact_joints	string[]	Joint names (from canonical_skeleton) for which the model emits foot_contacts. Empty means no contact prediction
native_clip_seconds	number	Duration the model was trained on. Requests beyond this are stitched per chunking
chunking	"none" | "stitched"	If "stitched", server may chunk longer requests internally
recommended_max_duration_seconds	number	Comfort-zone duration; clients SHOULD warn the user past this point
canonical_skeleton	Skeleton	The skeleton the model was trained on. See Skeletons →
limits	object	Per-request limits

Limits

Field	Type	Notes
max_duration_seconds	number	Largest accepted total duration
max_num_samples	int	Largest accepted Options.num_samples
max_constraints_per_request	int	Sum across all constraints in the request
max_prompt_length	int	Codepoints, per TextSegment.prompt
max_request_bytes	int	Body size limit. Exceeding returns 413 payload_too_large. The SDK server enforces it before parsing: from Content-Length, and while reading a body sent without one
max_batch_size	int	Most requests in one batch body
max_reference_frames	int, optional	1.2. Most frames a motion_reference segment's clip may carry, counted as sent: a server MUST accept a clip of at most this many frames at any fps, resampling or trimming it to its own window as needed (its duration_frames is capped like any segment's, by max_duration_seconds). Set by models that list "motion_reference"; absent otherwise
max_video_bytes	int, optional	1.2. Most bytes a video_reference's video may be. The SDK enforces it on inline video.data (decoded size, 413 payload_too_large); a server enforces it on a video.url when it fetches it. Set by models that list "video_reference"; absent otherwise
max_video_seconds	number, optional	1.2. Most seconds of video a video_reference may ask to read. The SDK enforces it on a trimmed video (end_s − (start_s or 0), 400 invalid_options); a server enforces it on an untrimmed one once it has it. It caps the video, not the output (that is max_duration_seconds). Set by models that list "video_reference"; absent otherwise

Using the canonical skeleton

A client that wants to use a model without retargeting reads canonical_skeleton from this endpoint and submits it verbatim in subsequent /generate requests:

curl -s https://server/capabilities \
  | jq '.models[] | select(.id == "kimodo-soma-rp") | .canonical_skeleton' \
  > canonical.json

Then in your request:

{
  "protocol_version": "1.0",
  "model": "kimodo-soma-rp",
  "skeleton": { "joints": [/* paste from canonical.json */] },
  "segments": [/* … */]
}

When supports_retargeting: true (the common case for production servers), you can ignore the canonical and send the user's actual skeleton.