codex-meshy-multiview
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@codex-meshy-multiview Generate a consistent four-view concept sheet and a textured 3D model with an adaptive low remesh."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Meshy Multi-View for Codex
Turn concept sheets into four consistent reference views, generate a textured 3D model with Meshy, and create a separate Adaptive / Low remesh. Includes a Codex skill, an MCP server, and a resumable command-line workflow. Existing single images and finished view sets are also supported.
flowchart LR
A[Concept sheet or existing references] --> B[Four separate orthographic views]
B --> C[Multi-Image to 3D + PBR textures]
C --> D[Adaptive Low remesh]
C --> E[Original GLB]
D --> F[Low-poly GLB + inspection report]The image preparation skill works with characters, creatures, props, vehicles, furniture, and other objects. For new designs, ImageGen creates a concept sheet by default, establishing one design across complementary views and useful detail studies. The skill checks the sheet for inconsistencies before preparing the four clean inputs. It also accepts existing references or an explicitly requested single concept image, preserving the selected variant, pose, asymmetry, materials, and camera scale. ImageGen runs through Codex's image-generation tool; this package does not implement a separate image-generation API client.
What it does
The MCP/CLI generation step accepts exactly four PNG/JPEG files or public HTTPS image URLs: front, back, left, right.
Uses
meshy-7.1, textured generation, PBR maps, and all four views as texture references by default.Runs remesh as a separate task with
topology: triangleanddecimation_mode: 4.Saves task IDs, supports resume, and never automatically retries a paid POST.
Downloads both GLBs, returned texture maps and thumbnails, and inspects triangle counts and embedded base-color texture bindings.
Provides eight MCP tools and an English skill for the entire concept-to-model workflow.
Adaptive Low is a relative level, not a fixed polygon budget. Meshy ignores target_polycount when adaptive decimation is enabled. GLB inspection verifies metadata and embedded texture bindings, not artistic quality or reference fidelity.
Related MCP server: ai-image-gpt-mcp
Verified examples
1. Concept sheet: Survey Pod
ImageGen concept sheet | Meshy Adaptive Low result |
|
|
The default workflow starts with a concept sheet, checks agreement between its views and details, and prepares four separate inputs. A live MCP run produced 352,496 → 6,249 triangles, with embedded base-color textures verified in both GLBs, for 35 reported API credits. See the complete image sequence and validation.
2. Single concept image: Utility robot
ImageGen concept | Meshy Adaptive Low result |
|
|
A live MCP run using a single concept as the starting reference produced 221,954 → 5,865 triangles, with embedded base-color textures verified in both GLBs, for 35 reported API credits. See the complete image sequence and validation.
Both examples include the initial reference, all four input views, the resulting model preview, and a reproduction command. Counts and generation quality vary between runs; live verification establishes API completion and texture retention, not exact design fidelity.
Requirements
uv on PATH; Python 3.11+ (uv can provision it).
A Meshy API key with sufficient API credits.
Codex with image generation for the optional concept/view preparation step. Existing four-view images also work without ImageGen.
Install
git clone https://github.com/KeeVeeG/codex-meshy-multiview.git
cd codex-meshy-multiview
uv sync --locked
uv run codex-meshy-multiview doctorConfigure credentials outside the repository. Never paste a real key into a Codex prompt, an issue, a committed file, or a command saved in shell history.
PowerShell, hidden input for the current shell:
$meshySecret = Read-Host 'Meshy API key' -AsSecureString
$env:MESHY_API_KEY = [System.Net.NetworkCredential]::new('', $meshySecret).PasswordBash:
read -rsp 'Meshy API key: ' MESHY_API_KEY; echo
export MESHY_API_KEYAlternatively, set MESHY_API_KEY_FILE to a private UTF-8 file containing only the key. The environment key takes precedence. .env.example documents the names; .env files are not automatically loaded. The MCP process must inherit one of these variables. See Codex setup for desktop configuration.
Use from Codex
Install the runtime and bundled plugin:
uv tool install .
codex plugin marketplace add .
codex plugin add codex-meshy-multiview@meshy-toolsStart a new Codex task after installation. The plugin launches the installed runtime with uv tool run --offline, so install the runtime before enabling the plugin. A typical request is:
Use the Meshy Multi-View skill. Start from my attached concept, preserve its design, create four separate consistent orthographic views, generate a textured model, and remesh it with Adaptive Low. Save both GLBs under my project's outputs directory.
For a new design:
Create a concept sheet for a stylized stone lantern with ImageGen. Check that its views and details describe the same design, then prepare matching front, left, back and right images and use Meshy to produce a textured model and an Adaptive Low remesh.
For a concept sheet only:
Use the Meshy Multi-View skill to create a concept/reference sheet for a stylized stone lantern. Show the same design from useful angles, with close-ups of its materials and assembly. Prepare the sheet only; do not start Meshy generation.
A concept sheet can contain multiple panels, detail studies, and requested variants. The skill selects one design and variant before preparing the four separate, clean Meshy inputs. A request for concept images or a sheet alone does not start a paid Meshy task.
The skill includes reusable, subject-independent concept-sheet guidance and a view-generation prompt. See image preparation for source choices and consistency checks.
Use from the CLI
Validate four inputs without an API request:
uv run codex-meshy-multiview plan --front front.png --back back.png --left left.png --right right.png --output ./runsRun generation, remesh, and downloads:
uv run codex-meshy-multiview run --front front.png --back back.png --left left.png --right right.png --output ./runsEach run has a unique directory. JSON results go to stdout; progress goes to stderr. Use absolute image/output paths when working from Codex. The first view sent to Meshy is always front, followed by right, back, left.
uv run codex-meshy-multiview resume ./runs/meshy-RUN_ID
uv run codex-meshy-multiview status ./runs/meshy-RUN_ID
uv run codex-meshy-multiview download ./runs/meshy-RUN_IDFor short calls suitable for agents, use start, then repeated advance, then download. advance polls once and submits remesh when generation succeeds. status only reads local state. Waiting has a timeout and remains resumable.
--texture-prompt "..." is optional. It replaces four-view texture guidance because Meshy disallows combining texture_prompt and texture_image_urls. Omit it to preserve the supplied views as the texturing reference.
Recovery and outputs
The run directory contains workflow.json, generation/model.glb, remesh/model.glb, and any texture/thumbnail files returned by Meshy. State records include signed asset URLs; treat run directories as private. Generated assets, run state, and secrets are ignored by Git.
If a POST times out after it may have reached Meshy, the workflow stops with submission_uncertain. Repeating advance cannot submit that task again. Identify the correct existing task in your Meshy account (or with list), then attach its ID:
uv run codex-meshy-multiview list --kind multi-image-to-3d --limit 10
uv run codex-meshy-multiview recover ./runs/meshy-RUN_ID --stage generation --task-id EXISTING_TASK_ID
uv run codex-meshy-multiview resume ./runs/meshy-RUN_IDFor uncertain remesh, use --kind remesh and --stage remesh. Recovery checks the ID and endpoint, but cannot prove it corresponds to the same reference images: select the task deliberately. If no task exists, start a new run only after resolving the uncertain submission. A rejected remesh can be diagnosed from saved state; there is no automatic paid retry.
Download promptly: Meshy's API asset retention is limited. Identical downloads are reusable; conflicting local files are not silently overwritten.
API contract
The implementation uses the Multi-Image API followed by the Remesh API. Remesh receives the generated GLB URL because the documented input_task_id types do not explicitly include Multi-Image tasks. The four-view workflow sends two paid submissions; check current Meshy pricing before use.
There is no key bundled with this project. Automated tests use mocked responses and synthetic GLBs; they do not spend credits or establish real-world Meshy output quality.
Development
uv sync --locked
uv run ruff check .
uv run pytest
uv buildCI checks Python 3.11 and 3.12 on Linux and Windows. See CONTRIBUTING.md, SECURITY.md, and the MCP tool reference.
MIT licensed. Unofficial integration; not affiliated with Meshy or OpenAI.
Available Tools
8 toolsmeshy_advance_workflowA
Check progress once and submit PAID adaptive low-poly remesh when generation succeeds.
Returns promptly without polling. Saved IDs prevent resubmitting completed stages. An uncertain submission requires recovery, never a repeat submission or a new run.
Inputs: run_dir (required). Returns: object
meshy_doctorARead-onlyIdempotent
Check whether local Meshy credentials are configured, without exposing them or calling the API.
Inputs: no named parameters. Returns: object
meshy_download_workflowAIdempotent
Download completed model and texture artifacts into the saved run directory; submits no paid tasks.
Inputs: run_dir (required). Returns: object
meshy_list_tasksARead-onlyIdempotent
List up to 100 existing Meshy tasks for inspection or uncertain-submission recovery; read only.
Inputs: kind (required), limit. Returns: object
meshy_plan_multiviewARead-onlyIdempotent
Validate four image paths/HTTPS URLs and preview redacted requests. No paid task is submitted.
Inputs: back (required), left (required), front (required), right (required), +2 more. Returns: object
meshy_recover_submissionAIdempotent
Attach a verified existing task ID after an uncertain submission. Never creates a task.
Inspect Meshy task history first and select the task belonging to this exact run and stage.
Inputs: stage (required), run_dir (required), task_id (required). Returns: object
meshy_start_multiviewA
Submit one PAID textured Multi-Image to 3D task using four views; return the saved run directory.
Local images are sent to Meshy. Continue this run with meshy_advance_workflow. Do not repeat start after an uncertain submission; inspect tasks and recover its ID.
Inputs: back (required), left (required), front (required), right (required), +2 more. Returns: object
meshy_workflow_statusARead-onlyIdempotent
Read a saved workflow and its next step. Local only; does not refresh remote task status.
Inputs: run_dir (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 8 tools
Each tool has a distinct role in the workflow, and the paid-task boundaries are clearly marked. There is mild potential confusion between meshy_advance_workflow and meshy_workflow_status, but the local-vs-remote distinction is clearly described.
All tools share a consistent meshy_ prefix and use snake_case, which is easy to follow. Minor deviations like meshy_doctor (noun) and meshy_workflow_status (noun_noun) break the otherwise verb_object pattern.
Eight tools is well-scoped for a 3D generation workflow: credential check, planning, paid submission, workflow advancement, status, download, recovery, and task listing. No tool feels redundant or unnecessary.
The core lifecycle from planning through submission, remesh, status, and download is covered, including a recovery path for uncertain submissions. Minor gaps exist around cancellation, cleanup, or credential configuration, but agents can complete the main workflow without dead ends.
Maintenance
Related MCP Connectors
Generate game-ready 3D models, textures, and audio from natural language, over MCP.
Focused MCP server for OpenAI image/audio generation (v2.0.0). Wraps endpoints via HAPI CLI.
MCP server for Qwen Image 3 AI image generation
Chat, image, video, and audio models from many vendors behind one remote MCP server and one key.
Related MCP Servers
- AlicenseAqualityBmaintenanceAI 3D model generation and post-processing MCP server — text/image/multiview-to-3D via Tripo, retopology, format conversion (GLB/FBX/OBJ/STL/USDZ), and stylization. Single Go binary, 10 tools.285Apache 2.0
- FlicenseAqualityDmaintenanceStandalone MCP server and CLI for generating images via ChatGPT backend, with reliable exact file paths and multi-account login support.25-
- AlicenseNot gradedqualityBmaintenanceAn MCP server that enables an image-to-3D-to-rigged character pipeline, offering tools to generate images with FLUX, create textured 3D meshes with TRELLIS, and rig characters with SkinTokens/TokenRig, all driven through MCP.GPL 3.0
- AlicenseNot gradedqualityDmaintenanceMCP server for Meshwright, enabling agents to create 3D models from text prompts via a LangGraph pipeline. Supports configurable image providers, background removal, and TRELLIS reconstruction backends.MIT



