nanoodle MCP Server

Description → canonical character → separate cutout parts → local rig and animation baking. One text call + two image calls. The companion rig.mjs extracts transparent parts and bakes 32 frames: id… text:Character -> llm:Character designer -> image:Canonical character -> resize:Reference -> edit:Rig parts; returns image + image (muse-image/edit, 1:1) saved to disk (file paths in result). Runs on NanoGPT — $0.06 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.0221, ~86s. Example: Iron Verdict (https://nanoodle.com/examples/iron-verdict/). A playable fighting game built with character-sprites. This graph generates a character reference and parts sheet; the local skill bakes animated sprites, and a coding agent adds combat, gravity and game rules. Agent skill: https://github.com/nanoodlecom/noodle-skills/tree/main/skills/character-sprites

Draft + Context + Guidance → 5 writers (Qwen3.8 Novelist/friend, DeepSeek V4.1 Flash/forum, Cydonia-24B v4.3/peer (Starry Ocean fallback on mcp), Qwen3.8 antislop/plain, Mercury 3/quick), each with… text:Draft -> text:Context -> text:Guidance -> join:Context+Guidance -> join:Writer input -> join:Judge: context+original -> llm:Writer A (Qwen3.8-27B Novelist, voice: friend) -> llm:Writer B (DeepSeek V4.1 Flash 1017, voice: forum) -> llm:Writer C (Cydonia-24B v4.3, voice: peer) -> llm:Writer D (Qwen3.8-27B antislop, voice: plain) -> llm:Writer E (Mercury 3, voice: quick) -> join:Judge: +A -> join:Judge: +B -> join:Judge: +C -> join:Judge: +D -> join:Judge: +E -> llm:Decider (GLM-5.3); returns text. Runs on NanoGPT — a deposit is quoted for each request from its length and maxChars and caps what the customer pays; it settles at the lesser of the metered cost plus 20% and that deposit, and any difference is returned; the operator absorbs any overrun. When Cydonia fails and the Starry Ocean fallback runs, both calls are charged. This refund and cap apply to hosted payments. Direct-wallet spending has no cap; last run $0.0023+, ~45s.

Idea → concept art → 3D model. Write one object in plain words. An LLM turns it into a concept-art description (shape, parts, materials, colors). Join adds fixed render rules (one whole object, thr… text:Object idea -> text:3D-ready render rules -> llm:Concept description -> join:Concept prompt -> image:Concept art -> model3d:TRELLIS.2 → GLB; returns model3d (trellis-2/image-to-3d) saved to disk (file path in result). Runs on NanoGPT — $0.51 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed. Example: 3D sample (https://nanoodle.com/examples/gallery/#idea-to-3d-model). One line becomes a textured 3D model: a teal-and-cream wind-up toy submarine with a brass periscope, portholes, a wind-up key and a striped tail. GLM writes the concept, Muse paints it, and TRELLIS.2 turns that one picture into a GLB that holds up from every side. The viewer shades it flatter than the concept, and the thin key and propeller blades go soft at some angles.

Compare four image models on the same brief. Run makes four paid images. Judge the exact heading and footer, object count, tool shapes, composition and color. Change the shared brief to test your o… text:Shared test brief -> image:Contender 1 -> image:Contender 2 -> image:Contender 3 -> image:Contender 4; returns image (muse-image/text-to-image, 1:1) + image (gpt-image-2.5/flare/text-to-image, 1k) + image (grok-imagine-image/v2.0/text-to-image, 1:1) + image (recraft-v4, 1024×1024) saved to disk (file paths in result). Runs on NanoGPT — $0.40 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.264, ~73s. Example: Reviewed sample (https://nanoodle.com/examples/gallery/#image-model-arena). The same bicycle-repair poster brief sent to four image models. All four retained the requested text in this run; only Grok rendered recognizable tire levers. Compare all outputs before choosing.

Mix four references into one picture. Each reference is its own node: who (the character), what they wear, what they hold, and where they stand. FLUX.3 Klein Edit gets them in port order and the in… text:Who · character -> text:What they wear · outfit -> text:What they hold · prop -> text:Where · place -> text:Mix instruction (names each image) -> image:1 · Character -> image:2 · Outfit -> image:3 · Prop -> image:4 · Place -> edit:Mix · FLUX.3 Klein Edit; returns image (flux-3/klein/edit-image, 1024×1024) saved to disk (file path in result). Runs on NanoGPT — $0.08 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed. Example: Composite sample (https://nanoodle.com/examples/gallery/#mix-the-references). Four separate references become one picture: a paper-craft fox (image 1) wears a yellow knitted raincoat (image 2), holds one red paper lantern (image 3) and stands on a cardboard pier at dusk (image 4). FLUX.3 Klein Edit gets them in port order and the instruction names each image by number. Flaws in the saved run: the paper boats from the place reference mostly drop out and the coat comes out longer than its flat-lay reference.

seed → GLM writes Anima still prompt → anima/text-to-image → GLM writes H3 Max motion prompt → minimax/h3-max i2v at 480p/5s. Change the seed; both LLMs rewrite prompts. ~$0.27 sampled. Hold both f… text:Fight seed -> llm:Still prompt -> llm:Motion prompt -> image:Fight still -> ivideo:Clash clip; returns video (h3-max) saved to disk (file path in result). Runs on NanoGPT — $0.63 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.2601, ~47s.

Create a first frame with Muse, then animate one small movement with MiniMax H3 Spicy at 480p for five seconds. Change the still and motion briefs together. Inspect object stability and motion; a s… text:Still brief -> text:Motion brief -> image:First frame -> ivideo:Animated clip; returns video (image-to-video) saved to disk (file path in result). Runs on NanoGPT — $0.51 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.21, ~1m 30s. Example: Reviewed sample (https://nanoodle.com/examples/gallery/#photo-to-video). A short tea-on-a-desk animation. Sampled frames retained the mug while steam moved. This thumbnail is one frame; the example does not establish a seamless loop.

Say whether an image, or evenly sampled frames of a video, is safe for work. One yes/no decision per picture (perplexity/pplx-decider-v1.2). The check passes only when every picture is yes. Video sampling needs ffmpeg on the server; a still image does not. There is no switch to skip this check. Runs on NanoGPT — a deposit is quoted for each request from how many pictures it checks and is a cap on the charge; it settles at the metered cost plus 20% or the quote, whichever is less, and a picture that would exceed the remaining budget is not checked; the difference is returned.

Start with a song idea and the bands you like. Three text calls write lyrics, arrange them, and identify what would clash with that arrangement. The arrangement and avoid-list become one musical di… text:Song idea -> text:Preferred bands -> join:Idea + musical references -> llm:Write lyrics -> join:Brief + finished lyrics -> llm:Arrange the song -> llm:What would clash? -> join:Arrangement + avoid-list -> music:Sing; returns audio (music-3) saved to disk (file path in result). Runs on NanoGPT — $0.37 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.1502, ~2m 35s. Example: Audio sample (https://nanoodle.com/examples/gallery/#sing). A 170-second generated closing-credits song. Gemini 3.8 Flash reported clear gentle acoustic vocals and a chorus matching the supplied lyrics. This is model-assisted audio review, not human listening sign-off.

Make a two-frame story through a handoff between models. Muse establishes the opening. Nano Banana stages the next beat from that actual image. Gemini compares both frames; Muse Edit uses the draft… text:First scene -> text:Next beat -> text:Continuity rules -> image:Muse opens the story -> join:Opening brief + next beat -> resize:Frame 1 -> join:Next beat + continuity rules -> edit:Nano Banana stages the next beat -> llm:Draft continuity check -> resize:Draft frame 2 -> join:Story + observed repair notes -> edit:Muse repairs the continuity -> resize:Frame 2 -> llm:Continuity notes; returns image + image + image + text; media saved to disk (file paths in result). Runs on NanoGPT — $0.19 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.0652, ~2m 4s.

Rewrite a draft, run humanize-comment on that rewrite, then ask whether the result is safe to post and how funny it is. Order is refine (liquid/lfm-2.5-2.6b), then humanize-comment, then a liquid/d1 yes/no gate, then a liquid/d1 funny score. Returns the comment, the gate verdict, the funny score, and the cost of each step. A "no" gate does not discard the comment. Runs on NanoGPT — a deposit is quoted for each request from its context and guidance and is a cap on the charge; it settles at the metered cost plus 20% or the quote, whichever is less, and the run stops before a stage the remaining budget cannot cover; the difference is returned. Needs the humanize-comment graph mounted.

Create a short spoken introduction with a fictional presenter. Muse makes the portrait, MiniMax Speech reads the script, and LongCat combines the image and audio into a speaking video at 480p. Keep… text:Presenter look -> text:Spoken script -> text:Delivery -> image:face -> tts -> lipsync; returns video (longcat-avatar-1.5) saved to disk (file path in result). Runs on NanoGPT — $0.66 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed. Example: Video sample (https://nanoodle.com/examples/gallery/#talking-avatar). A fictional presenter reads a short workshop introduction. Sampled frames preserve the presenter; check precise lip-sync timing in playback. Generation can take several minutes.

Muse builds a miniature. Wan animates that exact image. Mirelo watches the resulting silent clip and adds foley. Three models pass actual media forward. Compare the silent and sounded takes. Defaul… text:World -> text:Motion -> image:Build the miniature -> ivideo:Silent take -> resize:First frame -> vedit:Film with foley; returns image + video (sfx1.6/video-to-video) saved to disk (file paths in result). Runs on NanoGPT — $0.65 deposit per call, paid in Nano (XNO) — settles at actual model cost + 20%, change returned; no account needed; last run $0.27, ~2m 11s.