Generates Foundry D&D table art via a local headless ComfyUI instance, with Foundry-shaped tools for drafting, curating, and finishing images.
generate-image – Create art by purpose preset (handout, scene-background, portrait, token) with fixed generation/output resolutions.
draft mode – Fast FLUX.2-klein batches (1–8 images) for quick curation; can include reference images for consistent character faces.
final mode – High-quality FLUX.1-dev render finished at the preset output resolution.
refine mode – Dev img2img over a picked draft (denoise adjustable) to re-render in dev style while keeping the scene skeleton.
Configure prompt, slug, seed, batch size.
upscale-image – Finish an existing draft or image through the model upscale pipeline (4x-UltraSharp + downsample) to the target kind's output resolution.
artificer-status – Health check: ComfyUI connectivity/version, VRAM, queue depth, workflow integrity, and required model presence.
Returns absolute PNG file paths so Claude can read and curate actual images.
Uses pinned workflow JSONs, not free-form graphs, and keeps resolution pipeline locked (generate small, upscale, downscale).
Click on "Install 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., "@fvtt-mcp-artificer Generate a portrait of a grizzled dwarven ranger for my NPC"
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.
A Foundry-specific image-generation Model Context Protocol
server for D&D table art, driven by Claude Code. It wraps a headless
ComfyUI instance on the local machine and exposes a
small set of Foundry-shaped tools so Claude can author prompts, generate illustration batches,
curate the results by actually looking at them, and hand the winners to the Foundry pipeline in
its sister server, fvtt-mcp-molten5e
(upload-asset → set-actor-art / add-journal-image / scene backgrounds).
The whole loop is fast enough to be conversational on the target hardware (RTX 5090): a 6-image
draft batch lands in ~10 s, a finished 2560×1600 render in ~19 s, and the full
prompt → draft → curate → final → in-world-journal cycle was proven end to end in
notes/m3-loop-proof.md.
This is not a generic ComfyUI bridge, by decree. Tools speak Foundry vocabulary — actor
portraits, tokens, handouts, scene backgrounds — and get tweaked freely for Foundry work. And it
stays separate from fvtt-mcp-molten5e: that server is scoped to Foundry content authoring and
must not couple to image generation. This server never talks to the Foundry bridge; the handoff
between them is files on disk plus the molten5e upload tools.
Same house philosophy as the rest of the family: tools do, skills decide. Correctness (workflow
execution, dimensions, the upscale pipeline, file conventions) lives in tested tools here; judgment
(prompt craft, curation taste, which actor or journal gets the art, house style) lives in a later
illustration-builder skill.
Claude ──MCP──> fvtt-mcp-artificer ──HTTP──> ComfyUI (headless, local)
│
└── pinned workflow JSONs (draft / final / final-refine / upscale)ComfyUI runs headless in API mode; this server submits the pinned workflow JSONs under
workflows/ — never free-form graphs — with prompt/seed/batch/preset
substituted in (by node id, guarded by class-type drift assertions). Tools return absolute file
paths; Claude reads the PNGs directly to curate.
Related MCP server: FoundryVTT MCP Server
generate-image takes a kind, not width/height:
kind | generates at | finished output |
| 1536×960 | 2560×1600 |
| 1024×1280 | 2048×2560 |
| 1024×1024 | 2048×2048 |
Resolution pipeline (locked): never generate at output size — composition degrades past ~1.5 MP. Generate at the preset's native resolution, model-upscale ×4 with 4x-UltraSharp, lanczos downsample to the finished size, all in one pinned graph.
FLUX.1-dev fp8 (Comfy-Org all-in-one) — final-quality renders, ~15–19 s finished.
FLUX.2-klein 4B (Apache 2.0) — the draft model: 4-step, ~1–2 s/image, batch 6–8, pick, re-render with dev.
4x-UltraSharp — the model upscaler in the pipeline tail.
tool | what it does |
|
|
| Finish an existing image (usually a draft that won curation outright) through the upscale tail to its kind's output resolution. |
| Health check: ComfyUI reachability/version, VRAM, queue depth, pinned-workflow integrity, required-model presence. |
The substitution contract between tools and the pinned graphs is documented in
workflows/README.md.
Windows + NVIDIA GPU. Developed and proven on an RTX 5090 (Blackwell needs CUDA 12.8+ PyTorch; the current ComfyUI portable build ships it).
ComfyUI (standalone portable build, v0.34+) running headless — see
scripts/launch-comfyui.ps1 for the launch convention
(API mode on 127.0.0.1:8188, pinned --output-directory).
The model files in ComfyUI's models/ tree (~33 GB, all ungated):
checkpoints/flux1-dev-fp8.safetensors, diffusion_models/flux-2-klein-4b.safetensors,
text_encoders/qwen_3_4b.safetensors, vae/flux2-vae.safetensors,
upscale_models/4x-UltraSharp.safetensors. artificer-status reports anything missing.
Node.js 22+ for the MCP server itself.
npm install
npm run buildTests: npm test (offline unit suite; the pinned workflow JSONs are the fixtures). The live suite
— real ComfyUI, real renders — is gated behind npm run test:integration. Quality gates:
npm run typecheck, npm run check (biome), npm run knip.
House convention registers the server at user scope (new MCP tools ⇒ restart Claude Code). Either use the CLI:
claude mcp add -s user artificer -- node D:/path/to/fvtt-mcp-artificer/dist/index.jsor copy .mcp.json.example into a .mcp.json Claude Code reads (or merge
into ~/.claude.json mcpServers) with absolute paths. On Windows point command at the full
node.exe path if Node isn't on PATH.
Copy .env.example to .env (gitignored):
COMFY_URL — the headless instance (default http://127.0.0.1:8188).
COMFY_OUTPUT_DIR — must match the --output-directory ComfyUI was launched with; the server
reads generated PNGs straight off this path.
ARTIFICER_TIMEOUT_MS — per-job wait cap (default 300000).
MIT License — see LICENSE for details.
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Manage TTRPG campaigns: NPCs, locations, factions, quests, sessions, lore, and knowledge graphs.
Generate logos, social posts, app screenshots, comic panels & visual-novel assets from prompts.
Design, save, and run outcome-aligned AI workflows and verifiers, with reliable image output.
Generate AI images, video, music, and sound effects, and upscale them, from any MCP client.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/Txpple/fvtt-mcp-artificer'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
