Skip to main content

Roadmap

The v1 protocol surface is intentionally small. The items below split across three buckets:

  1. Protocol additions — features that need a wire-format change. Land as minor versions (1.1, 1.2, …) when there's a concrete use case behind them.
  2. SDK additions — features that any conforming server could ship. No protocol change required. Will land in the open-source motionmcp package over time.
  3. Cloud-only — features that only make sense as a hosted service. Won't ship in the open-source SDK. Available through Animatica Cloud, which any MMCP-compliant client can hit — the wire format is the same.

If something you need isn't here, open an issue on GitHub.

1 — Protocol additions (future minor versions)​

Each lands additively — clients on v1.0 keep working.

  • Streaming progress — SSE on /generate/stream for diffusion-step progress and partial frames.
  • Per-finger and facial constraint primitives — finger_pose, face_blend. v1 deliberately omits these to keep the constraint set reviewable.
  • Sparse-keyframe responses — bezier handles instead of dense linear channels, for clip sizes that scale with motion complexity rather than duration.
  • Round-trip motion editing — sending an existing motion as input alongside constraints for refinement.
  • Vendor extensions — formal extensions.<vendor>_* blocks in the response, for backbones that want to surface metadata not yet in core.

How additions land: server advertises a new capability (e.g. a new constraint type), client opts in by sending it. Old clients ignore unknown capabilities; new servers ignore unknown additive request fields. See Versioning →.

2 — SDK additions (open-source motionmcp)​

All of these will ship in the open-source SDK in future versions — they're just not there yet.

  • Async jobs — POST /generate returning 202 Accepted with a Location for GET /generate/jobs/{id} polling. Server-side queue + result retention.
  • Idempotency cache — honoring Idempotency-Key per the protocol's SHOULD. In-memory by default, pluggable backend for Redis etc.
  • Binary glTF (model/gltf-binary) — .glb responses for multi-sample or long clips, no base64 overhead.
  • fps resampling — server-side slerp/lerp between request fps and the model's native fps, so clients can request whatever fps suits their pipeline.
  • OpenTelemetry hooks — span-based tracing for /capabilities and /generate, request id propagation.

3 — Cloud-only​

These are deliberately not in the open-source SDK. They need ML expertise, hosted infrastructure, or run-cost economics that don't make sense for a single-machine setup. Animatica Cloud surfaces them behind the same MMCP wire contract — any client written against the protocol works.

  • Skeleton retargeting — accept any client skeleton (Mixamo, Maya HumanIK, Unreal mannequin, custom rigs) and retarget under the hood. The OSS path requires the canonical skeleton from /capabilities; Animatica Cloud lifts that constraint.
  • Hosted GPU compute — pay-per-generate without the user managing weights, drivers, or autoscaling.
  • Multi-tenant auth + billing — token-based access, usage metering, rate limit tiers.
  • Batch + scheduled generation — long-running jobs measured in hundreds of samples or many-minute clips.
  • Trained-model hosting — quality and fast variants, A/B'd model versions, opt-in to newer models without re-deploying anything.

If you need any of these, Animatica Cloud is the easier path. See Servers & clients → for the full catalog of supported pieces.

Contributing to the protocol​

The protocol is defined by these docs. Source lives in the motionmcp repo. Open an issue or PR.

For protocol-shaping discussions (new constraint types, new option fields), file an issue first describing the use case before opening a PR — the goal is to keep the surface small.