Meshy 3D generation API — text-to-3D, image-to-3D, multi-image-to-3D, remesh, retexture, auto-rigging, animation, 3D printing, plus nano-banana-backed 2D gen. Activate when the user asks for 3D models, rigging, animations, 3D printing, or configures any node whose id starts with `meshy-`. Sourced directly from docs.meshy.ai on 2026-04-17 — every parameter, endpoint, and price quoted here came from the canonical API reference.
Meshy 3D generation API — text-to-3D, image-to-3D, multi-image-to-3D, remesh, retexture, auto-rigging, animation, 3D printing, plus nano-banana-backed 2D gen. Activate when the user asks for 3D models, rigging, animations, 3D printing, or configures any node whose id starts with `meshy-`. Sourced directly from docs.meshy.ai on 2026-04-17 — every parameter, endpoint, and price quoted here came from the canonical API reference.
Meshy Skill
When to use
User asks for a 3D model (text description or image reference)
User wants to rig a character, animate a rigged character, or 3D-print a model
User wants to retexture or remesh an existing 3D model
User configures any meshy-* node
User asks about credit costs, rate limits, or which Meshy model to pick
Universal rules (all Meshy endpoints)
Every task is async. Submit returns {"result": "<task_id>"}. Poll GET /{endpoint}/{task_id} or subscribe to /stream for SSE. Task is done when status is SUCCEEDED, FAILED, or CANCELED.
Assets expire. Task responses include expires_at (ms since epoch). Download outputs promptly.
Image texturing limitation:"may not work optimally if there are substantial geometry differences" — keep the texture reference image close in form to the actual 3D model.
Multi-image-to-3D:"All images should depict the same object from different angles for best results" — 1 to 4 images, same subject, varied angles.
Workflow patterns
Full text-to-3D with textured output
POST /v2/text-to-3d with mode=preview, prompt, ai_model, optional topology/polycount — returns preview task_id
Poll preview task until SUCCEEDED
POST /v2/text-to-3d with mode=refine, preview_task_id=<id from step 1>, optional enable_pbr, texture_prompt or texture_image_url — returns refine task_id
Poll refine task until SUCCEEDED
Download from model_urls.glb (or whichever format)
Rig + animate a character
Generate or upload a textured humanoid GLB under 300k faces (use remesh first if larger)
POST /v1/rigging with model_url or input_task_id, height_meters (default 1.7)
Rigging result includes free walk + run animations AND a rig_task_id
POST /v1/animations with that rig_task_id and an action_id from reference/animation-library.md
Download result.animation_glb_url
Image → 3D → multi-color print
POST /v1/image-to-3d (textured, preferably meshy-6)
Poll until SUCCEEDED, capture its task_id
POST /v1/print/multi-color with input_task_id=<that id>, optional max_colors (1–16, default 4), max_depth (3–6, default 4)
Backend handlers in backend/handlers/meshy.py. All use the same _poll_meshy_task helper.
When chaining Meshy tasks (e.g. rigging → animate), the rig_task_id must come from a rigging node's output task_id port.
When a 3D node's output goes to the canvas, the GLB is downloaded locally and served via /api/outputs/… — frontend MeshPreview component renders via model-viewer.
Meshy's texture_image_url / image_style_url params accept public URLs or base64 data URIs. Our backend converts local file paths to data URIs via image_to_data_uri.
Known handler limitations are documented in reference/handler-gaps.md. Mention them to the user when they ask for something a Meshy feature supports but our node doesn't currently expose.
