Build Plugins

Even plugin development is an agent job now: the repository ships a dedicated skill, and your coding agent analyzes, implements, and tests a debugged ComfyUI workflow end to end. Once registered, workspace users and MCP-connected agents automatically share one stable capability — zero MCP integration code.

from workflow to mcp
ComfyUI workflowdebugged
→
Plugin folderplugin.json registers it
→
Gateway scanrestart to register
→
Workspace + MCP toolsautomatic, zero glue code
discovered by list_capabilities · inspected by get_capability · submitted by submit_generation — all generic tools

The repository ships .agents/skills/videofree-plugin-development/. Install the whole folder into your coding agent’s skills directory (Claude Code, Codex, Cursor, …), then send:

Use $videofree-plugin-development to turn my ComfyUI workflow into a VideoFree plugin. First analyze inputs, outputs, model dependencies, and the right plugin shape; then implement, test, and run the release self-check. All model-specific code and assets must stay inside the plugin folder.

The agent analyzes the workflow, picks the right plugin shape, then writes the implementation and runs the tests. Sections 2–4 below are its ground rules — and your acceptance checklist.

2Choose the plugin shape

  • Generation plugin: wraps an image, video, speech, or music workflow;
  • Action plugin: adds second takes, enhancement, or business operations;
  • Project mini-plugin: manages scripts, progress, and other data in the current project or category;
  • Application plugin: owns a full standalone page, like the editor.

3Keep all model-specific content inside the plugin

A plugin’s plugin.json is the only registration entry. Workflows, compilers, collaboration skills, Vue pages, icons, presets, and fixed samples must ship with the plugin. Model-specific rules never go into the framework, and plugins never read the platform database or another plugin’s code.

Once registered, MCP access is automatic: the capability shows up in list_capabilities, agents inspect its input limits with get_capability, and submit jobs with submit_generation — never touching nodes. Your only job is to get the plugin.json fields right.

What models are needed and where they go belongs in the plugin’s own INSTALL.md — the same file the installation guide tells users and agents to install from.

4Develop and validate against the real protocol

Detailed fields, runtime interfaces, and UI rules live in the repository: plugins/PLUGIN_DEVELOPMENT.md, docs/plugin-protocol.md, and ARCHITECTURE.md. When done, at minimum verify plugin scanning, related tests, the frontend build (when there is UI), and videofree.bat --check (./videofree.sh --check on Linux).

To verify the MCP side: have any MCP-capable AI call list_capabilities once. If your new capability is in the list and get_capability returns its full input schema, the integration is done.

Plugins are trusted local code, not a security sandbox. Drain related jobs before installing or updating, restart the gateway afterwards, and remember plugins never install ComfyUI models or nodes by themselves.