splatmaker-mcp
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., "@splatmaker-mcp generate a navigable 3D splat world from ./capture.mp4"
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.
splatmaker-mcp
Self-hosted, zero-marginal-cost Gaussian-splat generation. The FOSS alternative to worldlabs-mcp's Marble API — point it at a video or a set of images, get a navigable 3D Gaussian-splat world back, run entirely on local hardware (RTX 4090).
Status: v0.4.0, engine implemented AND verified. Server shell, tool schemas, job tracking, fleet health/registration contract, 11-page webapp, AND the real Nerfstudio/Splatfacto subprocess pipeline are all real and working — ns-train --help exits 0 (nerfstudio 1.1.4, verified 2026-11-01; earlier that day the CLI was broken by missing deps and the status check lied — now the server proves it with a cached --help probe). uv sync --extra engine installs torch+CUDA/nerfstudio/gsplat (verified on Goliath 2026-07-13: torch 2.6.0+cu124, CUDA true, RTX 4090 detected, gsplat 1.0.0). What's genuinely still pending: a full end-to-end training run against real capture data — nerfstudio's own Google Drive-hosted demo datasets are currently blocked by an external permission/quota issue on Google's end (not our code, tried two different capture names, same failure both times), so mechanical CLI/pipeline verification (flags checked against real --help output, executables confirmed present, control flow unit-tested, 8/8 passing) is done, but a real photorealistic splat hasn't been produced yet. That was always the plan anyway — needs Sandra's own real photos/video, same "get real reference material before writing the wrapper" discipline as vcv-rack-mcp's reference patches and Boomy's Leash's ARKit survey.
Why this exists
worldlabs-mcp (Marble API) isn't cheap. This fleet has an RTX 4090 sitting idle for exactly this kind of workload. A FOSS Gaussian-splat pipeline (Nerfstudio, decided 2026-07-13 — see "Engine status") wrapped as an MCP server gets the same "photo/video → navigable 3D world" capability at zero marginal cost, feeding the same downstream chain (blender-mcp → resonite-mcp / Resonite, VRChat) that worldlabs-mcp already feeds.
There's also a plausible case this is useful beyond this fleet — the Resonite and VRChat world-building communities currently either pay for splat SaaS or fight research-grade CLI tooling directly. If this ships clean, it may be worth public release per FLEET_PROMOTION.md.
Related MCP server: comfyui-mcp-server
Engine status
DECIDED 2026-07-13: Nerfstudio. Postshot scratched entirely (Sandra: "'download free' then 'account with useful CLI needs £40pcm' is a bit unverschämte Frechheit" — fair, and the free-download/paid-CLI split is a bait-and-switch shape regardless of the underlying business reason). gsplat ruled out as bare-library/no-CLI/too-much-glue. This section keeps the full comparison research below for the record, but the decision is made — next step is implementation, not further evaluation.
Researched 2026-07-13 (stars/age/CLI quality/speed/collider-mesh support/known bugs — sources: PyPI, GitHub, radiancefields.com, thefuture3d.com, polyvia3d.com benchmark, docs.nerf.studio). None chosen for implementation yet — this is a comparison to inform the choice, not a decision.
gsplat | Nerfstudio | Postshot | |
Type | CUDA rasterization library (PyTorch) | Full training/export framework (uses gsplat backend) | Proprietary desktop app (Jawset) |
Age | Oct 2023 | Oct 2022 | Beta Dec 2023 |
Stars / forks | ~5.3k / 897 | ~11.8k / 1.6k | N/A — closed source |
License | Apache-2.0 | Apache-2.0 | Proprietary. CORRECTED 2026-07-13: earlier "verified live: all tiers €0.00" claim in this doc was WRONG — jawset.com's pricing page uses a JS-driven calculator (monthly/yearly toggle, currency selector, license-count stepper); the automated fetch used to "verify" it doesn't execute JS and was reading a pre-hydration placeholder state (identical €0.00 across three separate fetches was the tell, in hindsight). Sandra's actual browser shows Studio at roughly £40/month — trust that over anything this doc claimed about live pricing. What IS reliable: the static feature-bullet list shows "Command-line interface" listed ONLY under Studio, not Free or Indie — CLI access is Studio-exclusive, paid, not a free-tier-with-GUI-restriction-lifted situation. Single-maintainer sole proprietorship (Jascha Wetzel, Munich, Germany) regardless of price |
Language | Python + CUDA | Python + CUDA (via gsplat) | Not published, native Windows app |
CLI | None — write your own training script | Real CLI: | GUI-first; CLI historically gated to paid Studio tier |
Collider mesh sidecar | No — pure rasterizer | Partial — | No, confirmed via direct |
Speed (RTX 4090) | ~8 min (fastest) | ~10 min | ~30 min (different benchmark scale) |
VRAM | ~4GB | ~6GB | Not specified |
Quality (PSNR, same benchmark) | 28.1 dB | 27.9 dB | Not benchmarked head-to-head |
Known gaps | JIT CUDA compile breaks on some PyTorch/CUDA combos; zero pipeline, all DIY glue | No fisheye/equirectangular/orthographic camera support (perspective only) — check against actual capture method | PLY exports have been reported missing attributes (normals, opacity, scale, rotation, SH coefficients) needed by some downstream tools (SOGS) |
Working recommendation (not yet implemented): Nerfstudio DECIDED: Nerfstudio. See the decision note at the top of this section.
Risk framing, revised again 2026-07-13 (pricing correction supersedes the free-tier framing below): the "install Studio, hope for the best" discussion below was conducted under the mistaken belief that Studio (and its CLI) was genuinely free. It is very likely NOT — Studio appears to be a real paid tier (~£40/month per direct browser observation), gating CLI access specifically. The "frozen vs. actively churning" business-stability discussion below still stands as reasoning, but the practical upshot changes: this isn't a zero-cost hedge anymore, it's a real ~£480/year recurring cost question against a ~€100/month total AI-tools budget, competing directly with everything else in that budget. Re-evaluate before subscribing to anything — don't let the sunk research cost here push toward a subscription that wasn't actually free after all.
Risk framing, revised 2026-07-13 (Sandra pushback, worth keeping both sides): the "agentic-transition instability" read above assumed Postshot v1 is being actively (if distractedly) touched. The more parsimonious read: v1 is frozen in place — attention (and any v2/rebrand) has moved elsewhere, and the free-tier reset reflects "not worth maintaining a paywall we're not really selling anymore," not ongoing churn. Under that reading the risk profile inverts: a genuinely abandoned tool can't drift under you (nobody's touching it), so it's actually MORE stable day-to-day than a tool under active development. The real remaining risks are narrower: (1) today's known bugs (e.g. the PLY/SOGS attribute gap above) are permanent, never getting fixed — test against them up front, not an ongoing hazard; (2) the download can vanish with zero warning since there's no commercial relationship obligating Jawset to keep it up. Mitigation for (2), and the actual action item: download and archive a local copy of the Studio installer now, while it's free — a frozen binary already on disk can't disappear. Given that, Postshot Studio is back on the table as a real SplatBackend candidate worth testing, not ruled out — evaluate it alongside Nerfstudio rather than defaulting past it. We genuinely don't know which scenario (active churn vs. frozen abandonware) is true; the installer-archive move is cheap and correct either way.
Correction to the above (2026-07-13, later same session): the installer archival was and remains a good, genuinely free move — downloading Postshot costs nothing regardless of tier. What's NOT free, per the pricing-page correction, is unlocking Studio-tier features (the CLI) for actual use. The --login/--password flags the other verification pass found on the CLI are almost certainly the Studio license/activation check, not an optional extra — meaning unattended SplatBackend automation via Postshot would require an active paid subscription tied to those credentials, not just a downloaded binary. Given Nerfstudio's Apache-2.0 status has zero equivalent cost, this materially strengthens the Nerfstudio recommendation over Postshot for anything meant to run unattended — Postshot's remaining honest use case is Sandra manually running the GUI (which the Free tier does support) for one-off high-quality captures, not the CLI-driven automation this server needs.
Done, 2026-07-13: installer archived at D:\Dev\archives\postshot\ on Goliath (Authenticode-verified: signed by "Jascha Wetzel, Munich, DE", matching the site's own imprint exactly — legitimate, not tampered). Turned out Postshot 1.1.0 was already installed on Goliath (C:\Program Files\Jawset Postshot\bin\, build dated 2026-05-13, predates this investigation) — verified working, not just present:
> postshot-cli.exe --version
Postshot v1.1.0 (2026-05-13)
> postshot-cli.exe --help
Subcommands: train, exportGround-truth confirmation of the collider-mesh finding (previously sourced from write-ups, now from the actual --help output): export's only output option is --export-splat (PLY or SPZ). No mesh, no collider, nothing geometry-adjacent in the CLI at all — the earlier secondary-source research was accurate, now verified directly.
New from the real CLI, not previously known: train takes --import (images/video/poses/point-clouds), a real depth of tuning flags (--profile {Splat ADC, Splat MCMC, Splat3}, --splat-density, --max-num-splats, crop/ROI boxes, sky-model generation, anti-aliasing), and outputs a .psht project file that export then reads from separately — a real two-stage train → export pipeline, scriptable. Also present: --login/--password flags on the CLI, meaning even the free tier may want an activation check before training actually runs — untested, next thing to verify before assuming this works fully unattended.
Load-bearing finding, not a nice-to-have: none of the three produce a collider mesh as a clean sidecar the way worldlabs-mcp's collider_mesh_url does. This means the ⚠️ warning above ("route through blender-mcp's blender_splatting before Resonite") is mandatory for any of these three, not a fallback for edge cases — budget it as a required pipeline stage in whatever implements SplatBackend.
splat_engine_status tool reports this live. Set SPLATMAKER_ENGINE env var once one is chosen — but the actual subprocess/API wrapper in server.py's SplatBackend class still needs to be written; the env var alone does not make generation work.
⚠️ Important: collision mesh
Unlike worldlabs-mcp (which returns assets.mesh.collider_mesh_url alongside every splat), FOSS splat pipelines generally do NOT produce a collision mesh. Any consumer of this server's output (Blender, Resonite) needs to route the raw splat through blender-mcp's blender_splatting tool to generate a usable collision mesh before it's usable in a VR world — skipping this step produces a splat that looks correct in a screenshot and has no floor to stand on. See mcp-central-docs/architecture/FLEET_DEEP_ANALYSIS_2026-07-13.md §2.2.1 for the full reasoning (this exact mistake was made once already for worldlabs, in the opposite direction — don't repeat it here).
Tools
splat_generate(portmanteau) — ops:from_video,from_images,status,list,get_assetsplat_engine_status— reports which engine is configured, plus the cachedns-train --helptruth (verifiedok/broken/missing)server_shutdown— graceful shutdown (matchesPOST /api/shutdown)show_status_card(app=True) — engine + job-queue status as a Prefab card
REST (backend :11091): GET /api/health, GET /api/status, GET /api/capabilities,
GET /api/jobs, POST /api/jobs/status, GET /api/skills, GET /api/logs,
GET /api/llm/models, POST /api/llm/chat, POST /api/shutdown.
Full reference: llms-full.txt. All tool returns carry success + message.
Webapp pages (frontend :11092): Dashboard, Generate, Jobs, Inbox, Gallery, Tools, Skills, Chat (local Ollama), Logs, Help (incl. what-Nerfstudio-is), Settings.
Ports
Backend (FastMCP HTTP
/mcp+ Starlette REST/api/*): 11091Frontend (Vite React dashboard): 11092
Operator backend (installed Tauri app, side-by-side rule): 11259
Registered in mcp-central-docs/operations/WEBAPP_PORTS.md.
Stack
Backend: Python 3.12, FastMCP 3.x, Starlette, uvicorn, SQLite (job store)
Frontend: React 18, Vite 5, TailwindCSS 3, Zustand, react-router-dom
Native shell: Tauri 2 (NSIS target)
Sub-documents: INSTALL.md, CHANGELOG.md, llms-full.txt,
docs/CONFIGURATION.md, docs/DEVELOPMENT.md, docs/TOOLS.md,
docs/TROUBLESHOOTING.md, docs/ONBOARDING.md.
Quick start
See INSTALL.md. Short version: uv sync, then uv run splatmaker-mcp (stdio) or uv run splatmaker-mcp --http (HTTP on :11091).
Claude Desktop one-liner (installs the latest .mcpb release bundle):
irm https://github.com/sandraschi/splatmaker-mcp/releases/latest/download/install.ps1 | iexRoadmap
Archive a local copy of the Postshot Studio installerDONE 2026-07-13, then Postshot scratched entirely 2026-07-13 (paid CLI, see decision note above — archived installer kept for reference/manual GUI use only, never load-bearing). Engine decided AND implemented 2026-07-13: Nerfstudio.SplatBackendis real —ns-process-data→ns-train splatfacto→ns-export gaussian-splat, background async execution, flags verified against real--helpoutput.Real end-to-end run against real capture dataDONE 2026-11-01 on synthetic data (job0fb0758f: 24 generated views → COLMAP 10 registered → 15k iters → 41 MB / 174k-Gaussian.plyin the depot). The one thing genuinely still pending: the same run against real capture data - needs Sandra's own photos/video of a real space.SQLite job persistence (v1 is in-memory only — jobs don't survive a restart).
Webapp: Gallery, Generate, Jobs pages per
WEBAPP_SOTA_STANDARDS.md.MCPB packaging +
glama.json(stub present, needs real health/tier data now that the engine is real).If pursuing public release: naked-PC install testing, README screenshots, community-facing docs pass.
operation="from_prompt"— an Ollama-driven prompt-to-spec stage, per DeepSeek V4 Flash's 2026-07-13 cross-examination review (deepseek comments.md) and Sandra's follow-up worked example ("Falling Water Villa living room" → disambiguation between real Fallingwater and North by Northwest's fictional Vandamm House). Full pipeline sketch indocs/FROM_PROMPT_DESIGN.md: disambiguation → grounding search (plex-mcp + web) → scene spec → comfyops-mcp hero image → comfyops-mcp i2v camera-path video → feeds straight into the already-builtfrom_videopipeline. ~90% reuse of existing fleet tools; the technical risk (does i2v maintain enough multi-view consistency for COLMAP?) is flagged unverified, not assumed.SplatBackendas a plugin registry rather than a single enum+env-var engine, per the same review — not urgent (current design isn't wrong), but worth doing before a second backend (e.g. a GUI-assisted manual-capture path) ever needs to coexist with Nerfstudio.Post-3DGS quality improvements (Scaffold-GS, 3DGS-MCMC, 2DGS, AbsGS) — full prioritized list with verified arXiv citations in
docs/SPLAT_QUALITY_ROADMAP.md. Literature review only, nothing implemented.
Available Tools
4 toolsserver_shutdownA
Gracefully shut down this server. Matches the REST /api/shutdown endpoint - both call the same underlying implementation.
Return Format
{"success": True, "status": "shutting_down", "message": str}.
Examples
server_shutdown() -> {"success": True, "status": "shutting_down", "message": "Server will exit in ~0.5s"}
Inputs: no named parameters. Returns: object
show_status_cardA
Show splatmaker engine + job-queue status as a rich Prefab card.
Return Format
A ToolResult with plain-text content (same numbers as the card) and
structured_content carrying the PrefabApp view for Prefab-capable hosts.
Examples
show_status_card() -> card with engine state, job counts by status, backend port.
Inputs: no named parameters. Returns: object
splat_engine_statusA
Report which splat engine is configured and whether the real CLI
tools are actually present (requires uv sync --extra engine).
Return Format
{"success": True, "engine": str, "configured": bool, "max_iterations": int, "note": str, "message": str, "verified": "pending"|"ok"|"broken"|"missing", "engine_version": str, "engine_error": str} — message mirrors
note so fleet tooling can render the result without special-casing it.
configured is the fast exe check; verified is the cached
ns-train --help truth (pending until the background probe lands).
Examples
splat_engine_status() -> {"success": True, "engine": "nerfstudio", "configured": True, "max_iterations": 15000, "note": ..., "message": ...}
Inputs: no named parameters. Returns: object
splat_generateA
Generate or inspect a Gaussian-splat world from local media.
Operations: from_video - start a job from a video file (video_path required) from_images - start a job from a set of images (image_paths required) status - check a job's status (job_id required) list - list all known jobs get_asset - return asset paths for a completed job (job_id required)
from_video/from_images return IMMEDIATELY with status="running" - the real ns-process-data -> ns-train -> ns-export pipeline runs as a background task (real training takes real minutes, not seconds). Poll with operation="status" until status is "done" or "failed".
Return Format
Always a dict with a success bool and a human-readable message.
Success paths also carry the documented payload keys (job_id, status,
jobs, asset_paths). Error paths carry success: False plus an error
key with the machine-readable reason; the error key is kept for backward
compatibility with existing clients.
Examples
splat_generate(operation="from_video", video_path="D:/cap/walk.mp4")
-> {"success": True, "job_id": "...", "status": "queued", "message": ...}
splat_generate(operation="status", job_id="<id>")
-> {"success": True, "job_id": "<id>", "status": "running", "message": ...}
splat_generate(operation="list") -> {"success": True, "jobs": [...], ...}
Inputs: operation (required), job_id, video_path, image_paths. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
splat_generate multiplexes five operations (create, status, list, get_asset) while show_status_card and splat_engine_status both report engine/job state, so an agent can plausibly reach for splat_generate(operation='status') or show_status_card for the same information. The descriptions distinguish them somewhat (card vs. raw status), but boundaries are soft.
All names are snake_case and readable; two share the splat_ prefix while server_shutdown and show_status_card do not, and verbs vary (generate, status, shutdown, show). Minor deviation from a strict verb_noun pattern but nothing confusing.
Four tools is a tight surface, but splat_generate absorbs five operations internally, so the effective capability set is reasonable for a focused splat-generation server. Nothing feels padded, though the count is on the low side.
The pipeline lifecycle is covered end to end: start job, poll status, list jobs, fetch assets, plus engine diagnostics. However there is no cancel/delete job operation and no way to inspect a job's logs or failures, which are meaningful gaps for long-running background jobs.
Maintenance
Related MCP Connectors
Remote MCP for RunComfy: ComfyUI deployments, hosted models, LoRA training. 31 tools.
Generate, edit, and deploy immersive 3D/WebGL web projects from any MCP assistant.
Cloud Blender for AI agents: build, inspect, render and animate 3D scenes over remote MCP. Keep editable .blend projects and export GLB or STL. Make your first 3D asset free—30 compute minutes per UTC month, no credit card. First 100 eligible Free account owners to use all 30 minutes in one UTC month by December 31, 2026 UTC receive one calendar month of beta free: 4 compute hours per UTC month, 2 concurrent workers per deployment, 2 deployments, 2048×2048 renders at up to 256 samples, and 10 GiB storage. Existing usage counts toward the 4-hour allowance. Limited to 100 rewards; expires January 1, 2027 00:00 UTC. Free resumes afterward without an automatic charge. Examples, eligibility and terms: https://sceneplane.online/launch?utm_source=glama&utm_medium=directory&utm_campaign=first100
Generate 3D models, worlds, PBR materials, and coordinated asset packs from AI agents.
Related MCP Servers
- AlicenseAqualityBmaintenanceMCP gateway to World Labs Marble + Spark 2.0 for generating navigable 3D worlds from text, images, panoramas, multi-view sets, or video, and grounding a voice agent in scene coordinates.2125MIT
- AlicenseNot gradedqualityCmaintenanceTurns curated ComfyUI workflows into typed MCP tools and REST endpoints, enabling image generation and job management through natural language or HTTP calls.MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI coding assistants and agents to natively generate images, videos, 3D meshes, Gaussian splats, and speech, and to manage media assets, through natural language commands or automated scripts over the Model Context Protocol. It exposes 55 tools alongside a Python CLI automation suite for batch generation, camera-orbit passes, credit/cost estimation, and companion API bridge integration.51Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables agents to turn a single image into a textured 3D mesh (GLB, OBJ with vertex colours, plus a turntable PNG) entirely on local hardware. Exposes capability reporting, dry-run planning, job queuing, and status/result retrieval so quality presets, hardware tiers, and consented model downloads can be driven through natural language.3Apache 2.0