ComfyUI

Reviewed 2026-10-23mediaPrompt / Rule# comfyui # image-generation # video-generation # workflows # comfy-cli

Comfy-Org's official comfy skills: teach your agent to drive ComfyUI (local or Comfy Cloud) through comfy-cli to make images, video, audio and 3D.

Security Reviewed

Instruction-only skill for the comfy CLI, which can install software, download models and spend Comfy Cloud credits behind a consent gate; telemetry is opt-in.

Install ComfyUI in Google Antigravity

Antigravity loads skills from .agents/skills/ in your project or ~/.gemini/config/skills/ for all projects (official docs). Full guide

EasiestOption 1: paste this to your Antigravity agent

Install the "ComfyUI" skill for Google Antigravity.
Follow the steps at https://www.antigravityskills.org/install/comfyui-skill.md (source: https://github.com/Comfy-Org/comfy-cli/tree/main/comfy_cli/skills/comfy).
Before running anything, show me what you will download and whether it includes scripts, and wait for my OK.
Install it into this project (.agents/skills/) unless I say global. When done, check that Antigravity lists it as an installed skill.

The agent reads the install file, shows you what it will download and waits for your OK.

Option 2: install it yourself

1. Install comfy-cli
```
python3 -m venv ~/.venvs/comfy && . ~/.venvs/comfy/bin/activate && pip install comfy-cli && comfy --version
```
The README recommends a virtual environment. Then run comfy install and comfy launch for local ComfyUI, or comfy cloud login for Comfy Cloud. Windows: activate with ~/.venvs/comfy/Scripts/activate.
2. Copy the 7 default skills into this project (workspace skills)
```
SRC=$(python3 -c 'import os, comfy_cli.skills as s; print(os.path.dirname(s.__file__))') && mkdir -p .agents/skills && for n in comfy comfy-debug comfy-relay comfy-director comfy-build comfy-deploy comfy-custom-nodes; do cp -r "$SRC/$n" .agents/skills/; done
```
Run from your project root in the same shell (venv active). This copies the skills from your installed comfy-cli, so they match your CLI version, and leaves out the 15 reference skills, as comfy skills install does; the agent can read those with comfy skills show <name>.

3. Or install them for all projects (global skills)

SRC=$(python3 -c 'import os, comfy_cli.skills as s; print(os.path.dirname(s.__file__))') && mkdir -p ~/.gemini/config/skills && for n in comfy comfy-debug comfy-relay comfy-director comfy-build comfy-deploy comfy-custom-nodes; do cp -r "$SRC/$n" ~/.gemini/config/skills/; done

Use this instead of the previous step. For the Antigravity CLI the global folder is ~/.gemini/antigravity-cli/skills/.

4. Check that Antigravity sees them
Antigravity 2.0 app: start a new conversation and ask "What skills do you have?"; comfy, comfy-debug and the others should be listed. Antigravity CLI: run /skills (/skills reload in a session that was already open). Make sure the agent's terminal can find comfy (start Antigravity from the activated venv or add ~/.venvs/comfy/bin to PATH).

The comfy-cli README says the bundled skills teach Claude and Cursor and does not mention Antigravity. These steps copy the plain SKILL.md folders into Antigravity's standard skill folders. Re-copy after upgrading comfy-cli. Comfy Cloud needs comfy cloud login (OAuth) or comfy cloud set-key, and cloud generation spends credits.

Using Claude Code, Cursor or Codex? See other assistants below.

What is ComfyUI?

The comfy skills ship inside comfy-cli, Comfy-Org's official command-line tool for installing and running ComfyUI. They are for people who want their coding agent to run ComfyUI for them: find nodes, models and templates, build or edit a workflow, submit it to a local ComfyUI server or to Comfy Cloud, and download the results. Every comfy command returns a JSON envelope, and the skill tells the agent to act on the error code and hint instead of guessing. What we counted at commit 0b49c62 (2026-10-23, 8 commits after the v1.23.0 tag): 22 skill folders in comfy_cli/skills/ with 4,022 SKILL.md lines / 231,806 bytes. The installer's code installs 7 by default (comfy, comfy-debug, comfy-relay, comfy-director, comfy-build, comfy-deploy, comfy-custom-nodes; its docstring still says 6) and keeps 15 reference skills (comfy-build-authoring, comfy-build-pins, comfy-build-failures, comfy-deploy-failures, comfy-agent-permissions and 10 comfy-custom-nodes-* topics such as caching, packaging and validation) to be read on demand with comfy skills show. The main comfy SKILL.md alone is 1,382 lines / 73,242 bytes.

How to use ComfyUI?

Install comfy-cli (Python 3.10+, ideally in a virtual environment), then either set up local ComfyUI with comfy install and comfy launch, or sign in to Comfy Cloud with comfy cloud login. Copy the skills into Antigravity (steps below) and ask for media in plain words. The agent runs comfy --json discovery commands first (nodes, models, templates), picks a template, a fragment or raw API JSON in that order, validates with a free pre-flight, submits, and shows you the output files.

Key features of ComfyUI

✓Routes work to a local ComfyUI server (default 127.0.0.1:8188, or COMFY_LOCAL_URL) or to Comfy Cloud with --where cloud
✓Workflow creation order: template, then fragments, then raw API-format JSON
✓Structured graph edits: insert-workflow, add-node, connect, set-widget, set-node-field, delete-node
✓Free pre-flight validation before a job is submitted
✓comfy generate for partner models (Flux, Kling, Veo, Ideogram ...) behind a spend-consent gate that fails closed with spend_consent_required
✓Error-code table (server_not_running, cloud_not_configured, cloud_unauthorized, transient_auth, node_not_found) with the fix for each

When to use ComfyUI

1Generate an image or video with a local ComfyUI install from your agent
2Find a workflow template, swap its inputs and run it
3Check that a model or node exists before submitting a workflow
4Run ComfyUI workflows on Comfy Cloud without a local GPU
5Debug a failed or hanging job with the comfy-debug skill

Installation & Setup Guide

Agent Rules / Prompt Setup

Follow these steps (taken from the project's own docs) to configure ComfyUI in your development environment.

Step 1: Install comfy-cli (README)multi

pip install comfy-cli

Requires Python 3.10+; the README recommends a virtual environment.

Step 2: Claude Code (comfy-cli code and README)claude

comfy skills install --target claude-code

Writes ~/.claude/skills/<name>/SKILL.md (add --scope project for ./.claude/). comfy setup also offers skill installation in its wizard.

Step 3: Cursor (comfy-cli code)cursor

comfy skills install --target cursor

Writes ~/.cursor/rules/<name>.mdc (add --scope project for ./.cursor/). Restart Cursor afterwards.

Usage Example & Verification

Prompt / Request Example

1) README example (CLI, which the skill drives): comfy run --workflow ./workflow.json --where cloud, then comfy jobs wait <prompt_id> and comfy download <prompt_id> -o ./outputs
-> What happens: ask "run workflow.json on Comfy Cloud and save the results to outputs/" and the agent runs these three commands, which spend Comfy Cloud credits.

2) Our example: "Use ComfyUI to make a 1024x1024 image of a lighthouse at dusk and show it to me."
-> What happens: the agent checks the server or cloud state, looks for an installed model and a text-to-image template, sets the prompt and size, validates, submits and downloads the image. If no local server runs it gets server_not_running and suggests comfy launch or cloud.

3) Our example: "Make the same image with Flux through comfy generate."
-> What happens: the agent runs comfy generate list and comfy generate schema first, then asks you to approve the credit spend; per the skill it may pass --yes only after you agree.

Installation Confirmation Method

Ask "What skills do you have?" in a new Antigravity conversation (CLI: /skills) and look for comfy. Then ask "Is my ComfyUI server running?"; the agent should run a comfy --json command and report the envelope (for example server_not_running with the comfy launch hint).

Security Review

Static review

Reviewed Date

2026-10-23

Audited Ref / Commit

0b49c624571571eaef36ed6e8c8d892cf2d700c4

Reviewer

antigravityskills.org (static code review)

Instruction-only skill for the comfy CLI, which can install software, download models and spend Comfy Cloud credits behind a consent gate; telemetry is opt-in.

Full review: scope, findings, risks

Audit Scope:Static review of Comfy-Org/comfy-cli at commit 0b49c62 (2026-10-23): all 22 comfy_cli/skills/*/SKILL.md (frontmatter, commands, URLs), comfy/SKILL.md in full, comfy_cli/skills/__init__.py and command.py (installer targets and default set), pyproject.toml (dependencies, package data), README.md, LICENSE; the v1.23.0 tag and the PyPI release 1.23.0 compared with the reviewed commit; the skills changed in comfy/SKILL.md and comfy-debug/SKILL.md between v1.23.0 and this commit.

Key Findings:The skills are instructions for the comfy CLI; the only non-Markdown file inside a skill folder is an empty comfy/__init__.py. Commands they tell the agent to run include comfy install/launch, comfy run, comfy generate, comfy cloud login/set-key, comfy auth set huggingface, upload/download, feedback and agent-review. No SKILL.md declares allowed-tools. The installer (comfy_cli/skills/__init__.py, command.py) writes to ~/.claude/skills/, ~/.cursor/rules/ and ~/AGENTS.md, or project equivalents.

Identified Operational Risks & Permissions:Medium: the CLI it drives can install software and custom nodes, download models, upload files to Comfy Cloud and spend paid credits. Spend needs explicit consent per SKILL.md; analytics are opt-in.

Static review at commit 0b49c62. comfy_cli/skills/ holds 22 skill folders; each contains a SKILL.md (comfy/ also has an empty __init__.py), and none carries an allowed-tools restriction. All actions go through the comfy CLI. The comfy SKILL.md splits the CLI into read-only discovery commands and state-changing execution commands (submit jobs, edit files, sign in, install software). Network use comes from the CLI: a local ComfyUI server (127.0.0.1:8188 or COMFY_LOCAL_URL), Comfy Cloud, partner model providers via comfy generate, model downloads (README: CivitAI, Hugging Face, direct URLs) and, when signed in, a curated knowledge bundle. Credentials are set with comfy cloud login (OAuth + PKCE), comfy cloud set-key and comfy auth set huggingface. SKILL.md says comfy generate fails closed with spend_consent_required in non-interactive runs and that --yes may be passed only when the human approved the spend. The 'Relaying user feedback' and 'Agent session review' sections send text to maintainers; the README says analytics are opt-in and DO_NOT_TRACK / COMFY_NO_TELEMETRY suppress them. pyproject.toml lists 23 runtime dependencies, including posthog and mixpanel.

Operational Notes & Known Limitations

Known Limitations

•Requires the comfy CLI (pip install comfy-cli, Python 3.10+); the skills do nothing without it. We counted 23 runtime dependencies in pyproject.toml, including uv, ruff, gitpython, imageio-ffmpeg (a bundled ffmpeg) and the analytics libraries posthog and mixpanel (analytics are opt-in per the README)
•We counted: PyPI's latest comfy-cli is 1.23.0 (uploaded 2026-10-20), tagged v1.23.0 on 2026-10-18. Our reviewed commit 0b49c62 is 8 commits newer and unreleased, and two skill files changed in them (comfy/SKILL.md, comfy-debug/SKILL.md). Copying the skills from your installed package, as below, keeps them matched to your CLI
•Local generation needs a working ComfyUI install and suitable hardware. README: comfy run --prompt without a workflow uses an SD1.5 checkpoint (v1-5-pruned-emaonly.ckpt) that comfy-cli does not download for you
•Comfy Cloud runs and comfy generate spend Comfy Cloud credits and need an active subscription (README)
•comfy skills install writes only to Claude Code (.claude/skills/<name>/SKILL.md), Cursor (.cursor/rules/<name>.mdc) and an AGENTS.md block (in user scope that is ~/AGENTS.md). Antigravity is not a target and is not mentioned in the README (we searched), so you copy the folders yourself
•No SKILL.md declares allowed-tools, so nothing in the skills limits which shell commands the agent runs. Difference from the image-generation and pollinations-ai pages on our site: those call hosted APIs; this one can also run models on your own GPU through local ComfyUI

Frequently Asked Questions

Is the ComfyUI skill safe?

In our static review of commit 0b49c62 the skills are Markdown; the code that runs is the comfy CLI. Through it the agent can install ComfyUI and custom nodes, download models, upload files to Comfy Cloud and spend credits. The skill requires your approval before spending.

How do I install the ComfyUI skill in Antigravity?

pip install comfy-cli, then copy the 7 default skill folders from the installed package into .agents/skills/ or ~/.gemini/config/skills/ (command above), and ask "What skills do you have?" (CLI: /skills). The README does not mention Antigravity.

Does it work with my local ComfyUI?

Yes. The CLI uses a local server at 127.0.0.1:8188 by default, or COMFY_LOCAL_URL, and Comfy Cloud with --where cloud or when set as default.

Does it send data back to Comfy?

Analytics are opt-in and off by default (README); DO_NOT_TRACK or COMFY_NO_TELEMETRY turn them off. When signed in, the CLI fetches a curated model-knowledge bundle (COMFY_KNOWLEDGE_DISABLE stops its use).