Nous Research's Hermes Agent Now Builds 3D Models From Plain Text

Nous Research has connected Meshy’s 3D generation stack to Hermes Agent documentation through an official plugin. Developers can ask Hermes to create a low-poly treasure chest for Godot, convert a photograph into a printable model, or prepare a character for animation, then inspect the result inside the conversation.

The integration combines Meshy’s first-party Model Context Protocol server with a Hermes workflow skill. Hermes handles tool selection, credit approval, task tracking, and output formats; Meshy performs the hosted image and geometry processing.

A pinned MCP server with a workflow skill

The plugin starts Meshy’s MCP server as a local process over standard input and output. MCP gives an agent a standard interface for calling external tools, while the local process communicates with Meshy’s API using the configured account key.

Each new session launches meshy-mcp-server@0.6.4 through npx. Hermes omits an MCP server’s embedded instructions from model context, so the plugin also bundles a meshy-3d skill containing routing rules, scenario playbooks, cost checks, format guidance, and a current tool table.

Hermes exposes 28 Meshy tools whose schemas total about 19,000 tokens. Tool search brings individual schemas into model context only when relevant, keeping the full catalog out of ordinary conversations.

• Generation: text-to-3D, image-to-3D, multi-image-to-3D, text-to-image, and image-to-image
• Post-processing: remeshing, retexturing, conversion, resizing, and UV unwrapping
• Characters: rigging, animation, text-to-motion, and animation-library browsing
• 3D printing: printability analysis and repair, multicolor 3MF, automatic part splitting, slicer handoff, and Creative Lab products
• Account operations: task status, model listings, downloads, balance checks, and usage data

3D previews stay in the conversation

Hermes Desktop renders generated models inside chat, displays the Meshy credit balance in its status bar, and maintains a gallery of previous creations. After downloading a model, the agent ends its response with a meshy-model directive that points Desktop to the local file.

The resulting card opens an interactive viewer with orbit, zoom, wireframe, and grid controls, plus an option to move the model into its own tab. When Hermes starts an asynchronous job, it can return the task ID immediately while the card displays Meshy’s progress. Once processing finishes, the card shows a thumbnail with Next and Download actions.

Messaging apps and other non-Desktop clients receive a short preview video, preserving a viewable result on devices that cannot render the interactive card.

Each card lists the formats produced for its task and saves the selected artifact beside the local .glb file. Meshy prepares those formats during generation, so choosing another download does not consume additional credits.

Credit controls fail closed

Most Meshy generation and processing calls consume between 5 and 62 credits. The plugin classifies the server as untrusted, which places approval prompts in front of write-capable operations such as creation, download, and cancellation.

The 200-credit setting acts as a chat-level reapproval threshold. Crossing it requires another decision from the user, limiting the effect of an overly broad session approval.

Unattended cron jobs and gateways cannot authorize paid operations, so Hermes blocks those requests before they can drain the associated Meshy account.

On Meshy 7, image-to-3D costs 20 credits without a texture, 30 credits with a standard texture, and 35 credits with an 8K texture. Actual consumption depends on the selected operation and settings.

Three commands to install

Installing and configuring the plugin requires three commands after Hermes is available:

hermes plugins install meshy
hermes plugins enable meshy
hermes config set MESHY_API_KEY msy_...

The runtime requires Node.js 18 or later, with npx available on PATH. Meshy API keys require a Pro plan or higher and can be created from the account settings page. The first launch downloads the pinned server package into npm’s cache.

The exact package pin keeps sessions reproducible across machines. Server upgrades proceed through a reviewed catalog pull request before the plugin manifest changes.

Every derivative keeps its lineage

Hermes Desktop’s gallery includes a Made from panel for each model. It can show the original prompt, texture or style prompt, source image, generation settings, seed, credits consumed, parent and child steps, and an Open chat link to the originating conversation.

Meshy retains text prompts for its own tasks. For image-based generation and later operations such as remeshing, retexturing, rigging, animation, and resizing, the plugin also records the creating tool call in provenance.json. Each record can include the prompt, settings, parent task, local image path, or source image URL with its query string removed.

Those records preserve the path from an initial prompt or photograph through each derived asset. Teams can audit settings, locate the source conversation, and retrace a multi-step pipeline without building a separate lineage system.

From prompt to game, print, or rig

Tool-using language models fit this workflow because the model handles orchestration while Meshy handles geometry. Hermes selects an operation, confirms its cost, monitors the task, and routes the resulting files into the requested workflow.

• Game development: generate placeholder or production assets in formats accepted by Godot, Unity, and Unreal Engine
• 3D printing: convert phone photographs into STL or 3MF files, then run printability analysis and geometry repair
• Character work: prototype rigs and animations without opening Blender for every iteration
• Product design: generate multi-view concepts, select a direction, and convert it into a 3D model

Nous’s Meshy release joins its official plugin catalog alongside integrations for Blender, Home Assistant, Spotify, and other services.

Plugin authors can reuse the same division of responsibilities: expose concrete capabilities through MCP, then place agent-specific routing, approval policy, and output handling in a workflow skill. The Meshy integration shows how that structure supports expensive asynchronous tools without loading every schema into the prompt or surrendering control over account spend.