Comfy Skills

Home of Comfy skills and plugins for AI coding agents. comfy-cloud is the Agent Skills and MCP connection that generate images, video, audio, and 3D, search models and nodes, and run ComfyUI workflows through Comfy Cloud. comfy-sdk is guidance for building apps on the Comfy SDK and Comfy Router.

Contributions welcome. We'd love your help strengthening these skills — submissions and pull requests are open, and we look forward to building this out with the community.

Install (Claude Code) — comfy-cloud

The comfy-cloud plugin bundles the Comfy Cloud MCP connection and Agent Skills in one step. The same plugin is what chat, Cowork, and Claude Code install. Skills load in all three. Slash commands in claude-code/commands/ still run by name in Cowork and Claude Code; in chat, Anthropic loads those command files as skills too.

/plugin marketplace add Comfy-Org/comfy-skills
/plugin install comfy-cloud@comfy-skills

Then run /mcp to sign in (OAuth). That's it — the plugin registers the hosted MCP server (https://cloud.comfy.org/mcp) and adds the skills.

In chat and Cowork, open Customize, then the Plugins tab (in Cowork, open the Cowork tab first). Select Add, then Add marketplace, choose Add from a repository, and enter Comfy-Org/comfy-skills. Open Discover, select comfy-cloud, and click Add. Then sign in from the plugin's Connectors tab. Plugins need a paid Claude plan, and a Team or Enterprise admin may limit which marketplaces can be added.

The connector URL is the same hosted server. Skills are namespaced under the plugin, e.g. /comfy-cloud:generate-image, /comfy-cloud:search-models, /comfy-cloud:help. Asking in plain language loads the matching skill.

Full per-client setup, including adding the server without the plugin, lives in the docs: https://docs.comfy.org/cloud/mcp

Install (Claude Code) — comfy-sdk

Building an app on the Comfy SDK (Python comfy-sdk, TypeScript @comfyorg/sdk) and Comfy Router? The comfy-sdk plugin is guidance only (no MCP server): one skill that steers the agent on client choice, server-side credentials, discover-then-call, run vs submit, timeouts, idempotency and error buckets.

/plugin marketplace add Comfy-Org/comfy-skills
/plugin install comfy-sdk@comfy-skills

Other agents: the skill is one file, claude-code-sdk/skills/comfy-sdk/SKILL.md. Copy it into .claude/skills/comfy-sdk/, your Cursor rules, or AGENTS.md; or, with comfy-cli, run comfy skills install --skill ./claude-code-sdk/skills/comfy-sdk.

Skills

Agent Skills for Comfy Cloud live in claude-code/skills/<name>/SKILL.md. Claude loads a skill when the request matches its description. The same names are invocable as /comfy-cloud:<name>.

Skill	What it does
generate-image	Generate, edit, or restyle a still image
generate-video	Generate, edit, or extend a video
generate-audio	Generate speech, music, or sound effects
clone-voice	Clone a voice from a reference recording and speak with it
generate-3d	Generate a 3D model
remove-background	Remove or replace an image background
upscale-image	Upscale an image
search-models	Search available models
search-nodes	Search available nodes
search-templates	Search workflow templates
consistent-characters	Keep a character matching across shots, using the server's live technique
help	Show what Comfy Cloud can do

combine-people stays a slash command only (claude-code/commands/combine-people.md). The live composite recipe is the MCP server's technique tool, fed by skills/technique-combine-people.md. The plugin does not ship that recipe a second time as a skill. The same is true of the other technique recipes under skills/technique-*.md (consistent characters is the one short skill that points at the live technique; the recipes themselves stay on the server).

Slash commands for the other workflows remain next to the skills for explicit invocation in Claude Code and Cowork. Their bodies are the longer procedures. The skills add the cross-surface rules (OAuth, no API key in chat, discover the live catalog) and do not replace the MCP server's own prompts.

Repo layout

• claude-code/ — the comfy-cloud plugin: skills/ (Agent Skills), commands/ (slash commands for Cowork and Claude Code), .mcp.json (the hosted connector), and .claude-plugin/plugin.json (metadata). Edit the Comfy Cloud plugin here. There is no techniques/ directory on main.
• claude-code-sdk/ — the comfy-sdk plugin: a single skills/comfy-sdk/SKILL.md plus .claude-plugin/plugin.json (metadata only, no MCP server). Keep it one file: comfy-cli and the copy-it-raw targets read only SKILL.md. CI (.github/workflows/sdk-skill-smoke.yml) checks its code blocks against the latest published SDKs.
• .claude-plugin/marketplace.json — the marketplace manifest so /plugin marketplace add Comfy-Org/comfy-skills resolves both plugins.
• skills/ — legacy flat files for the deprecated comfy-cloud-mcp curl installer. Frozen. The Comfy Cloud MCP server's refresh-skills job mirrors skills/*.md into production prompts. Do not add Agent Skills here and do not rewrite these files to change the plugin.

comfy-cli's own agent skills (comfy, comfy-debug, comfy-relay, comfy-director, comfy-build, ...) no longer live here — they ship bundled inside comfy-cli itself (comfy_cli/skills/), versioned with each CLI release, and are installed the same way as ever via comfy skills install.

Authoring rule: steer the approach, defer the specifics

A skill or command installed on a user's machine is frozen at install time. So it must never hard-code facts that change -- which templates exist, which model names are live, which nodes are available. Those rot, and a user who installed in spring is then following guidance written in winter.

Write skills and commands in two layers:

1. Stable steering (safe to freeze): the approach. For example, "look for a template first, clone it, swap the input, then submit; prefer structured discovery over guessing."
2. Live specifics (never freeze): tell the agent to discover current options at run time via the tools, rather than listing them inline. For example, use search_templates with the tag filter, or search_nodes with its typed output_type / category params, instead of pasting today's template names or node lists into the command.

If you find yourself writing a literal model name, template name, or node id into a command, stop and point the agent at the tool that returns the current set instead. That keeps the command correct as the catalog moves, and it keeps the agent from going off on a tangent before it engages the right tool.

• Build steps must leave the user a canvas-readable artifact: titled nodes + a saved workflow link. Any step that has the agent build a workflow JSON tells it to give every node a _meta.title naming its role, and, when the user asked for a workflow rather than just an output, to save_workflow it with overwrite: false (a new name on collision, never replacing the user's existing workflow) and return the get_workflow_canvas_url link for the workflow_id the save returned (looked up with list_saved_workflows if the result omits it) — never by filename, which publishes a public share snapshot.

Contributing

Skills are claude-code/skills/<name>/SKILL.md with name and description frontmatter. The description is the trigger: the situations a user would be in. Commands are plain markdown with a short YAML frontmatter description under claude-code/commands/. The comfy-sdk skill stays the single file claude-code-sdk/skills/comfy-sdk/SKILL.md. Keep them thin per the rule above, and open a PR. Run claude plugin validate ./claude-code and claude plugin validate ./claude-code-sdk before pushing. Keep prose clear and free of emoji.

Status

comfy-cloud is the Comfy Cloud plugin (hosted MCP connector + Agent Skills, with slash commands kept for Cowork and Claude Code). comfy-sdk is the SDK and Router guidance plugin (no MCP server). The legacy skills/ flat layout remains the MCP prompt snapshot and the deprecated comfy-cloud-mcp installer source. Cursor / Codex support can be added later as sibling folders.