Meshy 3d Generation

Meshy 3D Generation

Directly communicate with the Meshy AI API to generate 3D assets. This skill handles the complete lifecycle: environment setup, API key detection, task creation, polling, downloading, and chaining multi-step pipelines.

All paths below are relative to this skill's own directory (the directory containing this SKILL.md). Resolve them before running.

Security & Data Handling

• API key (MESHY_API_KEY) — sent only in the HTTP Authorization: Bearer header to https://api.meshy.ai. Never logged in full (only a key[:8]... prefix is ever printed). The bundled script never persists it; it is written to .env in the current working directory only when the user explicitly asks, and never to shell profiles, Windows user variables, or any path outside the working directory (see references/setup.md).
• Key sources read — the current session environment, then .env / .env.local in the current working directory. Home directories and shell profiles are never scanned.
• Network — the only external endpoint is https://api.meshy.ai. System proxies are bypassed (trust_env = False) so the key is never handed to an environment-configured proxy.
• Filesystem writes — .env in the working directory (on explicit request only) and ./meshy_output/ for downloaded models, thumbnails, and metadata. Input files (e.g. local images for image-to-3D) are read only at the exact path the user provides.
• Data leaving the machine — the API key, user-provided text prompts, and image URLs/data go to api.meshy.ai only. No other local data is transmitted; downloaded assets are saved locally.

IMPORTANT: 3D Printing → Use meshy-3d-printing Skill

If the user's request involves 3D printing (keywords: print, 3d print, slicer, slice, bambu, orca, prusa, cura, multicolor, 3mf, figurine, miniature, statue, physical model), use the meshy-3d-printing skill instead of this one for the entire workflow. The printing skill handles generation with correct print-optimized parameters (e.g. target_formats with "3mf" for multicolor), slicer detection, coordinate conversion, and slicer launch — all in one pipeline.

This skill's scripts/meshy_task.py is reused by the printing skill, but the workflow orchestration (what to generate, which formats, what to do after) must come from the printing skill when printing is involved.

Do NOT generate a model with this skill and then hand off to the printing skill — the printing skill needs to control parameters from the start (e.g. target_formats, should_texture).

IMPORTANT: First-Use Session Notice

When this skill is first activated in a session, inform the user:

All generated files will be saved to meshy_output/ in the current working directory. Each project gets its own folder ({YYYYMMDD_HHmmss}_{prompt}_{id}/) with model files, textures, thumbnails, and metadata. History is tracked in meshy_output/history.json.

This only needs to be said once per session, at the beginning.

IMPORTANT: File Organization

All downloaded files MUST go into a structured meshy_output/ directory in the current working directory. Do NOT scatter files randomly.

• Each project gets its own folder: meshy_output/{YYYYMMDD_HHmmss}_{prompt_slug}_{task_id_prefix}/
• For chained tasks (preview → refine → rig), reuse the same project_dir
• Track tasks in metadata.json per project, and global history.json
• Auto-download thumbnails alongside models

The bundled CLI implements this: project-dir, record, and thumbnail subcommands.

IMPORTANT: Shell Command Rules

Use only standard POSIX tools in shell commands. Do NOT use rg (ripgrep), fd, or other non-standard CLI tools — they may not be installed. Use these standard alternatives instead:

IMPORTANT: Run Long Tasks Properly

Meshy generation tasks take 1–5 minutes. When polling for completion:

• The bundled CLI prints unbuffered progress in real time — run each poll as a single Bash call and let it finish.
• Be patient with long-running polls — do NOT interrupt or kill them prematurely. Tasks at 99% for 30–120s is normal finalization, not a failure.
• Pass a larger --timeout (e.g. --timeout 600) for heavy tasks instead of retrying a timed-out poll.

IMPORTANT: Never Rebuild Bundled Scripts

scripts/meshy_task.py is the single source of truth for create_task / poll_task / download / get_project_dir / record_task / save_thumbnail. Never retype, paraphrase, or "reconstruct" these helpers from memory — not even partially. Compose CLI calls in bash, or write a small Python script that does sys.path.insert(0, "<this skill's scripts dir>") and from meshy_task import .... Reimplementing them inline causes silent behavior drift and doubles the token cost of every run.

Step 0: Environment Detection (ALWAYS RUN FIRST)

Before any API call, run the bundled environment check:

Only check the current session environment and .env files in the current working directory. Do NOT scan home directories or shell profile files.

python3 scripts/meshy_task.py check-env

It reports ENV_VAR (current environment), DOTENV (.env / .env.local in the working directory), PYTHON_REQUESTS, and a final READY: line. The bundled CLI loads the key itself (env var → .env → .env.local), so no manual export is needed to use it.

Decision After Detection

• READY: key=... → Proceed to Step 1.
• READY: NO_KEY_FOUND → Go to Step 0a.
• PYTHON_REQUESTS: MISSING → Run pip install requests.

Step 0a: API Key Setup (Only If No Key Found)

Follow references/setup.md. It walks the user through creating a key at https://www.meshy.ai/settings/api (Pro plan required), setting it for the current session only, and verifying it against GET /openapi/v1/balance.

Never persist the key yourself — no shell profiles, no Windows user environment variables, no file outside the current working directory. The only exception is .env in the working directory, and only when the user explicitly asks. Otherwise print the persistence instructions and let the user apply them.

Step 1: Confirm Plan With User Before Spending Credits

CRITICAL: Before creating any task, present the user with a summary and get confirmation:

I'll generate a 3D model of "<prompt>" using the following plan:

  1. Preview (mesh generation) — 5-20 credits (meshy-6/lowpoly: 20, others: 5)
  2. Refine (texturing with PBR) — 10 credits
  3. Download as .glb

  Total cost: 30 credits
  Current balance: <N> credits

  Shall I proceed?

For multi-step pipelines (e.g., text-to-3d → rig → animate), present the FULL pipeline cost upfront:

Note: Rigging automatically includes basic walking + running animations for free (in result.basic_animations). Only add Animate (3 credits) if the user needs a custom animation beyond walking/running.

Wait for user confirmation before executing.

Intent → API Mapping

Step 2: Execute the Workflow

CRITICAL: Async Task Model

All generation endpoints return {"result": "<task_id>"}, NOT the model. You MUST poll.

NEVER read model_urls from the POST response.

The Bundled CLI: scripts/meshy_task.py

Every workflow is a sequence of calls to the bundled CLI — do not write your own API code:

Pick the Workflow

Follow the matching recipe in references/pipelines.md — each lists the exact payload options and the full create → poll → download → record call sequence:

• Text to 3D (preview → refine) — the default for "make a 3D model of X"
• Image to 3D / Multi-Image to 3D
• Retexture / Remesh
• Convert / Resize / UV Unwrap (lightweight mesh utilities)
• Auto-Rigging + Animation — requires a textured humanoid model (rig the refine task, never the preview), t-pose, and a ≤300k face-count gate; rigging includes walking/running for free. A custom animation needs a real action_id from the public catalog
• Text to Image / Image to Image — see the 2D pre-step below

(Optional but strongly recommended) 2D Optimization Pre-Step

Prefer the image-to-3d route over direct text-to-3d — it's higher quality and more controllable, so for a text-only request make a design image first, then 3D-ify.

Image quality directly determines 3D model quality. Before calling /openapi/v1/image-to-3d or /openapi/v1/multi-image-to-3d, evaluate the user's input and proactively suggest a 2D pass:

The optimized image URL feeds directly into /openapi/v1/image-to-3d's image_url. 3-9 extra credits typically buy a noticeable quality bump, and downstream refine / texture-on-mesh stages benefit too.

Skip when: the user already provided a clean front-facing studio shot — go straight to image-to-3d. Also skip for Creative Lab products (figure / lamp / keychain / fridge-magnet): they apply their own built-in stylization, so feed the raw photo (or text, for lamp) straight to Creative Lab — do not pre-generate a design image.

Step 3: Report Results

After task succeeds, report:

1. Downloaded file paths and sizes
2. Task IDs (for follow-up operations like refine, rig, retexture)
3. Available formats (list model_urls keys — may include glb, fbx, obj, usdz, 3mf)
4. Thumbnail URL if present
5. Credits consumed and remaining balance (run balance; each task JSON also has consumed_credits)
6. Suggested next steps:
   • Preview done → "Want to refine (add textures)?"
   • Model done → "Want to rig this character for animation?"
   • Rigged → "Want to apply an animation?"
   • Any model → "Want to remesh / export to another format?"
   • Any textured model → "Want to 3D print this? Multicolor printing is available!" (requires meshy-3d-printing skill)
   • Any model → "Want to 3D print this model?" (requires meshy-3d-printing skill)

Error Recovery

On any failure, follow references/troubleshooting.md: HTTP status handling (401/402/422/429/5xx), retry policy, and known task FAILED messages. The bundled CLI already auto-reports the current balance on 402 and exits non-zero with the server's error message on failure.

Known Behaviors & Constraints

• 99% progress stall: Tasks commonly sit at 99% for 30–120s during finalization. This is normal. Do NOT kill or restart.
• CORS: API blocks browser requests. Always server-side.
• Asset retention: Files deleted after 3 days (non-Enterprise). Download immediately.
• PBR maps: Must set enable_pbr: true explicitly.
• Format availability: Check keys in model_urls before downloading — not all formats are always present (the poll summary lists them). 3MF is available from the Multi-Color Print API.
• Download format: ALWAYS ask the user which format they need before downloading. Recommend: GLB (viewing), OBJ (white model printing), 3MF (multicolor printing), FBX (game engines), USDZ (AR). Do NOT download all formats.
• 3MF format: 3MF is NOT included in default output of generation endpoints. To get 3MF, pass "3mf" in target_formats on generate/refine/remesh/retexture, or use the Convert API (POST /openapi/v1/convert, 1 credit). For multicolor 3D printing, the Multi-Color Print API outputs 3MF directly — no need to request it from generate/refine.
• Deprecated params: symmetry_mode no longer affects output; art_style is ignored by Meshy-6; use pose_mode instead of the old is_a_t_pose flag; use texture_resolution ("2k"/"4k"/"8k") instead of hd_texture; on image-to-3d use model_type: "smart-topology" (with ai_model: "meshy-t2") instead of the deprecated "lowpoly". meshy-4 is retired (returns 400).
• Smart Topology is image-to-3d only: Text to 3D and Multi-Image to 3D have no smart-topology — for clean low-poly from text, go text-to-image → image-to-3d, or remesh down afterwards.
• Rigging needs textures: rig the textured task (text-to-3d refine, or image-to-3d with should_texture: true). Rigging a mesh-only preview fails — untextured meshes are unsupported.
• Inspect before downloading: pass multi_view_thumbnails: true on image-to-3d / multi-image-to-3d and read thumbnail_urls (front/right/back/left, 512×512 PNG) instead of pulling a 50–200 MB GLB just to check the result. ~3s extra latency.
• Never hardcode action_id: fetch GET https://api.meshy.ai/web/public/animations/resources (public, no key, ?category= to narrow) and match the user's intent against name / category. IDs are not 1..N — the catalog includes -2, -1, 0.
• Failed tasks are free: a FAILED task reports consumed_credits: 0 (credits are refunded), so a transient failure can be retried without re-asking the user to approve the spend.
• consumed_credits: Every task GET response includes consumed_credits — read it to report the real credits spent rather than estimating.
• Timestamps: All API timestamps are Unix epoch milliseconds.
• Large files: Refined models can be 50–200 MB. The CLI streams downloads with timeouts; just be patient.

Execution Checklist

• Ran environment detection (check-env, Step 0)
• API key present and verified
• Presented cost summary and got user confirmation
• Composed the workflow from bundled scripts/meshy_task.py calls (never retyped the helpers)
• Followed the matching recipe in references/pipelines.md
• Reported file paths, formats, task IDs, and balance
• Suggested next steps