Civitai + comfyui-mcp

Civitai ships an official remote MCP server — discovery and the whole community surface — and comfyui-mcp owns the local machine: download, wire, queue, tag. They compose. comfyui-mcp deliberately does not proxy or re-expose Civitai's tools; MCP clients connect to both servers at once, so the right pattern is pairing, not wrapping.

Is the Civitai MCP connected?

Look for mcp__civitai__* tools in your session (e.g. mcp__civitai__search_models).

The comfy plugin bundles it — its .mcp.json declares the Civitai remote server alongside comfyui, so plugin users get mcp__civitai__* automatically with no manual setup and no key required (the Authorization header defaults to an empty Bearer, which Civitai accepts for its read/browse tools). It just works headless.

• Present → use it for discovery (this skill's main path).

• Absent (standalone MCP install, not the plugin) → either install the plugin, or add the server once:

  claude mcp add --transport http civitai https://mcp.civitai.com/mcp \
    --header "Authorization: Bearer YOUR_CIVITAI_API_KEY"
  

  Until then, fall back to search_models (HuggingFace) + download_civitai_model by id/URL.

API key (optional). Browsing/search needs none. To unlock the user's account context (favorites, posting, gated/early-access downloads), set CIVITAI_API_TOKEN (from civitai.com/user/account) in the environment — the same variable comfyui-mcp uses for its own downloads, so one secret powers both the Civitai MCP header and download_civitai_model.

The handoff that makes this work

Civitai's discovery tools return a model id and a model-version id — exactly what comfyui-mcp's download_civitai_model takes. No URL scraping:

mcp__civitai__search_models  ──▶  pick a result, read modelVersions[].id
        │
        ▼
download_civitai_model({ model_version_id, target_subfolder })   # local fetch
        │
        ▼
list_local_models  →  panel_add_node loader / generate_image      # use it

target_subfolder must match the model type: checkpoints, loras, vae, controlnet, embeddings, upscale_models, etc. (see the model-registry and model-compatibility skills). Prefer model_version_id — a Civitai page can list several versions and the user usually means a specific one.

Recipes

"Find me a good anime LoRA for Flux and install it"

1. mcp__civitai__search_models with the query + a type/base-model filter (use mcp__civitai__list_enums to get valid filter values).
2. Present the top 3–5 with name, creator, base model, downloads, and the version id. Let the user pick (base-model match matters — a Flux LoRA will not load on an SDXL checkpoint; see model-compatibility).
3. download_civitai_model({ model_version_id, target_subfolder: "loras" }).
4. In the panel: panel_add_node a LoraLoader, panel_set_widget the lora_name, wire it between checkpoint and sampler. Headless: generate_image / build the workflow.

"What's trending on Civitai for product photography?"

• mcp__civitai__search_models sorted by trending + search_images for example outputs and their embedded generation params. Summarize; offer to install any the user likes via the handoff above.

"Download this Civitai page for me" (user pastes a URL)

• Parse the modelVersionId from the URL if present; otherwise mcp__civitai__get_model to resolve the latest version id, then download_civitai_model. (You can also pass a raw civitai.com/api/download/... URL straight to download_model with CIVITAI_API_TOKEN set.)

In the sidebar panel

The panel's agent is the user's Claude Code session, so if both servers are connected the panel already has every mcp__civitai__* tool next to the panel_* tools — discover on Civitai and wire onto the live graph in one turn, no extra setup. If only comfyui-mcp is connected, suggest adding the Civitai MCP, then proceed with the HF/id fallback.

Boundaries — what to confirm before doing

Civitai's MCP also exposes write/social tools (post, comment, review, direct message, follow, bounties). Those publish on the user's behalf — treat them like any outward-facing action: surface what you're about to post and get an explicit yes first. This skill is about discovery → local install → generation; don't post or message without being asked.

See also

• model-registry — curated direct download URLs (HF + Civitai notes)
• model-compatibility — base-model / VAE / CLIP pairing (why a LoRA won't load)
• prompt-engineering — Civitai image params are a prompt goldmine