trident-mcp
# trident-mcp
[](https://go.dev)
[](LICENSE)
An unofficial MCP server for the [Tripo 3D](https://www.tripo3d.ai) API that covers every documented, deployed v3 workflow. A few endpoints are deliberately left out; they are listed in [docs/MCP_TOOLS.md](docs/MCP_TOOLS.md#not-exposed). Agents can:
- generate meshes from text, one image or several views;
- make reference images and multiview sets;
- texture, retopologize, UV-unwrap, segment, rig, animate and convert meshes;
- produce Gaussian splats.
The outputs land as files on disk. It ships as one Go binary over stdio, plus a Claude Code plugin with a workflow skill.
## Quick start
**Claude Code.** Put the binary on your `PATH`, then install the plugin:
```bash
go install github.com/mordor-forge/trident-mcp/cmd/trident-mcp@latest # or grab a release archive
```
```text
/plugin marketplace add mordor-forge/claude-plugins
/plugin install trident@mordor-forge
```
The plugin asks for your Tripo key and output folder, and adds the `tripo-3d` skill. Then ask for what you want, e.g. *"make a game-ready low-poly treasure chest and export it as FBX"*.
**Any MCP client:**
```json
{
"mcpServers": {
"trident": {
"command": "trident-mcp",
"env": { "TRIPO_API_KEY": "tsk_...", "MODEL_OUTPUT_DIR": "/abs/path/generated_models" }
}
}
}
```
## How it works
1. A task-creating tool (`text_to_model`, `rig_model`, `convert_model`, …) validates its arguments locally, uploads any local files, and returns a `task_id`.
2. `wait_for_task` waits with progress notifications, then saves the outputs to `<output dir>/<task_id>/`. For example: `model.glb`, `rendered_image.webp`, `generated_image.png`, `front_view.png`.
3. Pass a `task_id` as the next tool's `input` to chain steps: image → mesh → rig → animation → FBX.
Arguments use Tripo's own field names, so [Tripo's docs](https://developers.tripo3d.ai) apply to them directly. A few convenience fields differ from Tripo's wire format, such as multiview `inputs` given as a `{front, left, back, right}` map. See [docs/MCP_TOOLS.md](docs/MCP_TOOLS.md) for all 28 tools.
## Configuration
| Variable | Default | |
| --- | --- | --- |
| `TRIPO_API_KEY` | required | From <https://platform.tripo3d.ai> |
| `MODEL_OUTPUT_DIR` | `~/generated_models` | Created if missing |
| `TRIPO_BASE_URL` | `https://openapi.tripo3d.ai/v3` | China-region keys: `https://openapi.tripo3d.com/v3` |
## Documentation
| Need | Read |
| --- | --- |
| Every tool, argument and behaviour | [docs/MCP_TOOLS.md](docs/MCP_TOOLS.md) |
| Architecture and the `internal/tripo` module | [ARCHITECTURE.md](ARCHITECTURE.md), [docs/design/tripo-module.md](docs/design/tripo-module.md) |
| Domain vocabulary | [GLOSSARY.md](GLOSSARY.md) |
| Contributing, change recipes | [AGENTS.md](AGENTS.md), [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md), [docs/patterns/common-changes.md](docs/patterns/common-changes.md) |
| Running, live tests, failure handling | [docs/OPERATIONS.md](docs/OPERATIONS.md) |
| Decisions | [docs/adr/](docs/adr/) |
| Security | [SECURITY.md](SECURITY.md), [THREAT_MODEL.md](THREAT_MODEL.md) |
## Tripo terms and trademarks
trident-mcp is an independent project. It is not affiliated with or endorsed by Tripo, VAST or Holymolly Ltd, and "Tripo" is a trademark of its owner. Your use of the Tripo API is governed by [Tripo's terms](https://www.tripo3d.ai/terms). Under them:
- Use your own API key, and keep it private.
- Tripo's written authorization is required before offering a shared or hosted deployment to other people.
- Ownership of generated models depends on your Tripo plan.
Credit prices quoted in the docs and skill are dated summaries; [Tripo's pricing page](https://developers.tripo3d.ai/en/pricing) is authoritative.
## License
Apache-2.0. See [LICENSE](LICENSE). The license covers this repository's code, not Tripo's service or the models you generate.
TDQS
Scored across 28 tools
Each generation tool is keyed to a distinct input/output modality (text_to_model, image_to_model, multiview_to_model, image_to_splat, text_to_image, etc.), so selection is usually unambiguous. The only mild overlaps are segment_mesh vs smart_segment, and get_task vs get_tasks vs get_usage, which all retrieve task/credit info but are differentiated in their descriptions.
All names use snake_case and are consistently readable, with a recurring verb_noun or source_to_target structure. Minor deviation in that many tools use a 'source_to_target' convention (text_to_model) rather than a leading verb (generate_model), but this is coherent across the set.
At 28 tools the set is on the heavy side, sitting in the borderline-heavy band. Most tools correspond to genuine distinct pipeline operations (generation, texturing, rigging, UV, segmentation, conversion, credits), so each earns its place, but the volume is high enough to feel sprawling.
The surface covers the full 3D asset lifecycle: image/mesh generation, multiview, texturing, retopology, UV unwrap, segmentation, rigging, animation retargeting, conversion, download and credit checks. The notable gap is task management — there is no cancel/delete/cleanup operation for failed or unwanted tasks.