Quickstart — client

Building a server in Python? See Quickstart (server) →.

This walks through hitting an MMCP server with raw HTTP — no SDK required. Useful for plugin authors writing in non-Python languages, or for sanity-checking a server you're standing up.

1 — Discover

curl -s https://your-mmcp-server.example.com/capabilities | jq

The response lists every model the server hosts, plus its canonical skeleton, supported constraints, fps, and limits. Pick a model id — say, kimodo-soma-rp — and read its canonical_skeleton.

{
  "protocol_version": "1.0",
  "models": [
    {
      "id": "kimodo-soma-rp",
      "fps": 30.0,
      "supports_retargeting": true,
      "supported_constraints": ["root_path", "effector_target", "pose_keyframe"],
      "canonical_skeleton": { "joints": [ /* … */ ] }
    }
  ]
}

See Capabilities reference → for the full shape.

2 — Generate

The smallest useful request is a model id, a skeleton, and one text segment.

curl -s https://your-mmcp-server.example.com/generate \
  -H "Content-Type: application/json" \
  -d @request.json \
  > motion.gltf

Where request.json is:

{
  "protocol_version": "1.0",
  "model": "kimodo-soma-rp",
  "skeleton": { "joints": [ /* model's canonical_skeleton */ ] },
  "segments": [
    {
      "type": "text",
      "prompt": "a person walks forward, then waves hello",
      "duration_frames": 120
    }
  ]
}

The response is a standard glTF 2.0 document. Load it with any glTF parser — pygltflib, cgltf, the Khronos reference loader, three.js, or your DCC's built-in glTF importer.

3 — Apply

Each generated sample is one entry in animations[], named sample_0, sample_1, etc. Channels target nodes by joint name — the same names you sent in the request. Bake them onto your rig as you would any other animation clip.

The extensions.MMCP_motion block carries non-glTF metadata: fps, foot-contact masks, and chunk boundaries. Clients that don't need it can ignore the extension entirely and still get a valid glTF.

See Response → for the extension schema.

What's next

• Hand the user's actual skeleton. Most servers can retarget; you don't have to use the canonical. See Skeleton →.
• Add constraints. Pin the root to a path, an effector to a position, or fix a pose at one frame. See Constraints →.
• Bridge segments. Compose a motion as text → unconditioned → text for natural transitions. See Segments →.
• Tune generation. Seeds, sample counts, guidance, transition blending. See GenerateRequest →.