How rigging works
Rigging turns a static mesh into a character an engine can animate. Three things have to happen, and the API does all of them in one call.
A skeleton is predicted. A model infers where the joints belong from the shape of the mesh: hips, spine, shoulders, elbows, knees, and so on for a biped, or a spine plus four limb chains for a quadruped. No template is fitted and no manual marker placement is needed, which is what separates this from tools where you place joints by hand.
Skin weights are bound. Every vertex is assigned to the bones that move it, and by how much. This is what makes an elbow bend smoothly instead of tearing. Weighting is where most of the perceived quality of a rig lives.
The result is packaged. The skeleton, the weights, and your original materials and UVs are written back out as a GLB or FBX. Materials survive: the mesh you get back looks like the mesh you sent, not a grey blob.
What comes back
A rigged file with Mixamo-compatible bone naming. That naming is the de-facto standard for humanoid characters, so clip libraries and engine retargeters recognise it without a manual bone-mapping step. It is the difference between dropping a character into Unity's Humanoid pipeline and spending an afternoon in a bone-mapping dialog.
For non-humanoid rigs, bone names follow a structural convention (leg_front_left_* and similar) because no cross-industry standard exists for creature skeletons.
How long it takes
The Fast engine typically finishes in 15 to 30 seconds, most of which is queueing and mesh I/O rather than inference. The Pro engine averages around 150 seconds. Dense meshes take longer on both, because past a certain size the work is dominated by loading and writing geometry rather than by the model.
What it does not do
Facial rigs and blendshapes. The skeleton is body-only. Existing blendshapes on your mesh are preserved, but none are created.
Fingers, reliably. Hand bones are placed for humanoid rigs, but finger-level fidelity depends heavily on whether the source mesh has separated, well-formed fingers. Treat finger quality as best-effort.
Cleanup of bad geometry. Non-manifold meshes, interpenetrating limbs, and A-pose vs T-pose ambiguity all degrade the result. See model requirements.
Creature leg fold. This is worth being explicit about, because everyone in this market has the same problem. Quadruped rigs from every automatic rigger we have measured, ours included, place leg joints closer to a straight line than an artist would. A real animal's leg is folded, so the bone chain is roughly a quarter longer than the straight-line distance from hip to foot; automatic rigs cluster far below that. The practical consequence is that a quadruped rig walks acceptably but cannot crouch or leap convincingly without manual joint adjustment. We published our measurements rather than quietly shipping around it.
Choosing an engine
Start with Fast. It handles the large majority of models, costs a fraction of Pro, and returns in a fraction of the time. Fall back to Pro on the specific meshes where Fast fails or produces a poor skeleton. The engines guide has the details.