# Comfy Model Downloader > Download missing built-in comfy-agent-tools model files on demand by capability. Use when a generation/edit/upscale/music/video request needs local models that are missing, when comfy-models reports missing_model_file, or when the user asks to download supported base models. Do not use to download all models unless explicitly requested. - **Type:** Skill - **Install:** `agentstack add skill-quinteroac-comfy-agent-tools-comfy-model-downloader` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [quinteroac](https://agentstack.voostack.com/s/quinteroac) - **Installs:** 0 - **Category:** [Content & Media](https://agentstack.voostack.com/c/content-and-media) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [quinteroac](https://github.com/quinteroac) - **Source:** https://github.com/quinteroac/comfy-agent-tools/tree/main/skills/comfy-model-downloader ## Install ```sh agentstack add skill-quinteroac-comfy-agent-tools-comfy-model-downloader ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # comfy-model-downloader Use this skill to download missing model files for supported built-in profiles. Downloads are capability-scoped: download only what the user's current request needs. If `comfy-models` is not available, use `comfy-tools-setup` first. In this repo, prefer `uv run comfy-models`; outside the repo, use the installed CLI. ## Capability Mapping - Image generation: `imagegen.generate` - Image editing: `imagegen.edit` - Ideogram 4 local generation: `imagegen.ideogram4-generate` - Krea2 Turbo local generation: `imagegen.krea2-generate` - Image upscale: `imagegen.upscale` - Text to video: `videogen.t2v` - Image to video: `videogen.i2v` - First/last frame video: `videogen.flf2v` - Image plus audio to video: `videogen.ia2av` - WAN 2.2 image to video: `videogen.wan22-i2v` - WAN 2.2 first/last frame video: `videogen.wan22-flf2v` - WAN 2.2 sound to video: `videogen.wan22-s2v` - WAN 2.2 video plus audio: `videogen.wan22-video-audio` - WAN 2.2 Bernini video edit: `videogen.wan22-bernini` - Music generation: `musicgen.generate` - Image description: `imagedescribe.describe` (local HuggingFace model directory, not a single-file download) Do not map Seedance 2.0 API requests to downloads. `videogen.seedance2-t2v`, `videogen.seedance2-r2v`, and `videogen.seedance2-flf2v` use the remote `seedance2-api` profile and require `COMFY_ORG_API_KEY`, not local model files. Do not map SeedVR2 upscale requests to downloads either. `videogen.seedvr2-upscale` uses the `seedvr2` profile and the pinned SeedVR2 upstream CLI auto-downloads its own model files. Do not map Grok Imagine API requests to downloads either. `imagegen.grok-generate` and `imagegen.grok-edit` use the remote `grok-imagine-api` profile. Ideogram 4 is local and should be downloaded with `imagegen.ideogram4-generate` when requested. ## Flow 1. Resolve the capability from the user's request. 2. Preview the download: ```bash uv run comfy-models download imagegen.generate --dry-run ``` 3. Tell the user which profile, files, sources, and approximate size will be used, especially for large video/music models. 4. Download only after the request implies the capability is needed: ```bash uv run comfy-models download imagegen.generate --yes ``` 5. Validate the profile, then return to the original generation skill: ```bash uv run comfy-models validate-profile anima-base ``` ## Supported Built-In Profiles - `anima-base`: Anima Base v1.0 image generation with the turbo LoRA. - `qwen-edit2511`: Qwen Image Edit 2511 generation/editing. - `flux-klein-9b-snofs`: FLUX.2 Klein 9B FP8 generation/editing with SNOFS. - `ideogram4-fp8`: local Ideogram 4 FP8 text-to-image with structured JSON/bbox prompting. - `krea2-turbo`: Krea2 Turbo FP8 text-to-image with Qwen3-VL text encoder and conditioning rebalance. - `clear-reality`: ClearReality image upscaling. - `ltx23-10eros`: LTX 2.3 video and IA2AV. - `ltx23-dasiwa-golden-lace-v3`: Dasiwa LTX 2.3 Golden Lace v3 video and IA2AV. - `wan22-t2v`: WAN 2.2 text-to-video. - `wan22-i2v`: WAN 2.2 image and first/last-frame video. - `wan22-dasiwa-tastysin-t2v`: Dasiwa TastySin WAN 2.2 T2V local UNets plus shared UMT5/VAE. - `wan22-dasiwa-boundbite-t2v`: Dasiwa BoundBite WAN 2.2 T2V local UNets plus shared UMT5/VAE. - `wan22-s2v`: WAN 2.2 sound-to-video with wav2vec2 audio encoder. - `wan22-dasiwa-littledemon-v2-s2v`: Dasiwa LittleDemon v2 S2V, 4-step profile. - `wan22-dasiwa-littledemon-v2-video-audio`: Dasiwa LittleDemon v2 S2V for 16 fps video+audio, audio-driven, and external-mask lipsync. - `ace15-base`: ACE-Step 1.5 music generation. - `qwen3vl-2b-instruct`: Qwen3-VL 2B Instruct image description (local HuggingFace model directory; not distributed by the downloader). `seedance2-api` is intentionally excluded: it is a remote Comfy API profile, not a downloadable model profile. `grok-imagine-api` is excluded for the same reason. `ideogram4-fp8` is not remote; it uses Hugging Face model files under `models_dir`. Do not download models for custom local profiles unless the CLI provides a source. For unknown custom checkpoints, use `comfy-model-onboarding` and ask the user for the file. ## Auth And Sources Supported sources are Hugging Face and direct HTTP model links. Set `HF_TOKEN` for gated Hugging Face repositories; FLUX.2 Klein 9B requires accepting the Black Forest Labs license terms before download. Ideogram 4 may require accepting the Ideogram model terms before download. Set `CIVITAI_API_TOKEN` if Civitai requires authenticated downloads. The tool writes into the active `models_dir`, usually `/mnt/models/comfyui`, and creates missing subfolders. Partial downloads use `*.part` files and are renamed only after validation. ## Limits - Do not download every model unless the user explicitly asks for all supported models. - Do not download ad hoc LoRAs; `--extra-lora` files are managed by `comfy-lora-onboarding`. - Do not change model defaults as part of downloading. - Do not install ComfyUI custom nodes or start a server. - Do not download or validate local files for Seedance 2.0 API generation. - Do not download or validate local files for Grok Imagine API generation. - Do not redistribute FLUX.2 Klein or SNOFS weights, and do not use SNOFS for a public/commercial generation service without separate licensing. ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [quinteroac](https://github.com/quinteroac) - **Source:** [quinteroac/comfy-agent-tools](https://github.com/quinteroac/comfy-agent-tools) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-quinteroac-comfy-agent-tools-comfy-model-downloader - Seller: https://agentstack.voostack.com/s/quinteroac - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.