comfy-cli: A Command Line Tool for ComfyUI

comfy-cli is a command-line tool for installing, running, and extending ComfyUI — the open-source generative-media engine. Set up ComfyUI, install custom nodes and models, run workflows, and call hosted partner image models, all from your terminal.

Demo

Features

• 🚀 One-command ComfyUI install and launch
• 🎨 Direct calls to partner image and video nodes (Flux, Ideogram, DALL·E, Recraft, Stability, Gemini/nano-banana, Kling, Luma, Runway, Pika, Vidu, Hailuo, Seedance, …) via comfy generate, no workflow JSON required
• 🔧 Custom node management — install, update, snapshot, bisect
• 📦 Fast dependency resolution with uv (--fast-deps, --uv-compile)
• 🗄️ Model downloads from CivitAI, Hugging Face, and direct URLs
• 🎬 Run workflows against a local ComfyUI server, including auto-conversion of UI-format JSON
• 🧪 Test ComfyUI and frontend pull requests with one flag
• 💻 Cross-platform: Windows, macOS, Linux
• ☁️ Route any workflow to Comfy Cloud with --where cloud (no GPU required)
• 🤖 Agent-friendly: every command emits structured --json envelopes
• 📚 Bundled skills teach Claude / Cursor to drive comfy natively

Quick Start

pip install comfy-cli
comfy setup

comfy setup walks you through everything — local or cloud routing, authentication, and agent skill installation — in one interactive wizard. Pass -y for non-interactive (CI/scripted) installs.

Run a workflow in the cloud

No local GPU? Route any API-format workflow to Comfy Cloud with --where cloud:

comfy cloud login                                        # sign in via your browser (OAuth)
comfy run --workflow ./workflow.json --where cloud       # submits, prints a prompt_id, returns immediately
comfy jobs wait <prompt_id> --where cloud                # block until the job finishes (or: jobs status for a one-shot check)
comfy download <prompt_id> --where cloud -o ./outputs    # save the results locally

comfy run submits asynchronously and prints the prompt_id you feed to jobs/download; add --wait to block inline instead. Check your sign-in anytime with comfy cloud whoami. Export the workflow JSON from ComfyUI via File → Export (API) (UI-format JSON is auto-converted). Set cloud as your default target so you can drop the flag: comfy set-default --where cloud.

Credits: cloud generation (comfy run --where cloud, comfy generate) consumes Comfy Cloud credits and needs an active subscription. Discovery and inspection commands — comfy cloud whoami, comfy cloud status, comfy jobs status/ls, comfy templates ls, comfy generate list — don't. Check your balance and tier with comfy cloud status.

Installation

1. (Recommended) Activate a virtual environment (venv or conda).

2. Install with pip (requires Python 3.10+):

   pip install comfy-cli
   

Shell Autocomplete

Install shell completion so comfy <TAB> expands commands and options:

comfy --install-completion

Usage

Installing ComfyUI

To install ComfyUI using comfy, simply run:

comfy install

This command will download and set up the latest version of ComfyUI and ComfyUI-Manager on your system. If you run in a ComfyUI repo that has already been setup. The command will simply update the comfy.yaml file to reflect the local setup

• comfy install --skip-manager: Install ComfyUI without ComfyUI-Manager. To use a custom Manager fork or specific version, skip the default installation and install your own into the workspace venv:

  comfy install --skip-manager
  # Then install your custom Manager:
  pip install -e /path/to/your-manager-fork   # editable install
  # or
  pip install comfyui-manager==4.1b8          # specific version
  

• comfy --workspace=<path> install: Install ComfyUI into <path>/ComfyUI.
• comfy install --fast-deps: Use uv instead of pip for faster dependency resolution during initial ComfyUI installation. comfy-cli's built-in resolver compiles all requirements (core + custom nodes) into a single lockfile and installs from it. Also handles GPU-specific PyTorch wheel selection automatically.
• For comfy install, if no path specification like --workspace, --recent, or --here is provided, it will be implicitly installed in <HOME>/comfy.

Python environment handling

When you run comfy install, comfy-cli picks a Python environment for ComfyUI dependencies using the following precedence:

1. An active virtualenv or conda environment (VIRTUAL_ENV / CONDA_PREFIX) is used as-is.
2. An existing .venv or venv directory inside the workspace is reused.
3. Otherwise the choice depends on how comfy-cli was installed:
   • pip install comfy-cli (global / system Python): dependencies go directly into the same Python environment. This is the typical Docker setup.
   • pipx install comfy-cli or uv tool install comfy-cli (isolated tool environment): a .venv is created inside the ComfyUI workspace. Use comfy launch to start ComfyUI with the correct Python.

Updating ComfyUI

comfy update brings an existing workspace up to date:

• comfy update (or comfy update comfy): pull the branch the workspace is currently on and reinstall requirements.txt.
• comfy update all: also update every installed custom node.
• comfy update cli: upgrade comfy-cli itself.

By default comfy update all exits 0 even when the custom-node update step fails — the error is printed, but the exit code says success, so scripts and wrappers can't tell. Pass --exit-on-fail (the same flag name and default as comfy node install --exit-on-fail) to have that failure exit non-zero instead. The flag only changes the all target; comfy update comfy and comfy update cli already exit non-zero when they fail, and accept the flag as a no-op so it can be forwarded unconditionally.

One limit worth knowing: --exit-on-fail can only surface what ComfyUI-Manager's cm-cli update reports. cm-cli currently handles a single pack failing to update by printing ERROR: ... and carrying on, still exiting 0, so that particular case stays invisible until ComfyUI-Manager propagates it. Unlike comfy node install, the flag is not forwarded to cm-cli — only its install subcommand accepts --exit-on-fail; its update subcommand has no such option and would reject it.

When the flag does fire, the exit code is cm-cli's own, normalized so it is always usable: a process killed by a signal becomes 128+N rather than a truncated value, an exit code that would truncate to 0 becomes 1, and 2 becomes 1 so it can't be mistaken for a CLI usage error. In --json mode the failure is also reported as an update_custom_nodes_failed error envelope carrying cm-cli's raw status in details.cm_cli_returncode.

Switching to a specific version

comfy update comfy --version <X> moves an existing workspace to a specific ComfyUI version — a downgrade (rollback) or an upgrade — without prompting for anything, so it is safe to run headlessly or from a script. <X> is nightly (the repo's default branch), latest (the newest stable release), or a version number such as 0.3.0 (a leading v is optional).

comfy update comfy --version 0.3.0      # roll back to the v0.3.0 release
comfy update comfy --version latest     # newest stable release
comfy update comfy --version nightly    # roll forward to the default branch

Behavior worth knowing:

• The target is validated before anything is touched. An unknown version exits non-zero, lists the nearest available versions, and leaves the working tree exactly as it was.
• Uncommitted changes are stashed by default (git stash push -u) and are never popped or dropped automatically — the stash ref is printed so you can restore them with git stash pop. Pass --no-stash if you would rather the command refuse to run on a dirty tree.
• A version number checks out a tag, which leaves a detached HEAD. That is expected. Roll forward again with comfy update comfy --version nightly (or --version latest); a plain comfy update cannot advance a detached HEAD.
• Dependencies are reinstalled from the target version's requirements.txt. PyTorch is deliberately left alone: the ComfyUI version doesn't determine your torch build, your machine does. If the dependency install fails, the command exits non-zero and says so — the tree is already on the new version, and re-running the same command is safe.
• --version and --no-stash apply only to target comfy; combining --version with all or cli is an error.

Specifying execution path

• You can specify the path of ComfyUI where the command will be applied through path indicators as follows:

  • comfy --workspace=<path>: Run from the ComfyUI installed in the specified workspace.
  • comfy --recent: Run from the recently executed or installed ComfyUI.
  • comfy --here: Run from the ComfyUI located in the current directory.

• --workspace, --recent, and --here options cannot be used simultaneously.

• If there is no path indicator, the following priority applies:

  • Run from the default ComfyUI at the path specified by comfy set-default <path>.
  • Run from the recently executed or installed ComfyUI.
  • Run from the ComfyUI located in the current directory.

• Example 1: To run the recently executed ComfyUI:

  • comfy --recent launch

• Example 2: To install a package on the ComfyUI in the current directory:

  • comfy --here node install comfyui-impact-pack

• Example 3: To update the automatically selected path of ComfyUI and custom nodes based on priority:

  • comfy node update all

• You can use the comfy which command to check the path of the target workspace.

  • e.g comfy --recent which, comfy --here which, comfy which, ...

Default Setup

The default sets the option that will be executed by default when no specific workspace's ComfyUI has been set for the command.

comfy set-default <workspace path> ?[--launch-extras="<extra args>"]

• --launch-extras option specifies extra args that are applied only during launch by default. However, if extras are specified at the time of launch, this setting is ignored.

Launch ComfyUI

Comfy provides commands that allow you to easily run the installed ComfyUI.

comfy launch

• To run with default ComfyUI options:

  comfy launch -- <extra args...>

  comfy launch -- --cpu --listen 0.0.0.0

  • When you manually configure the extra options, the extras set by set-default will be overridden.

• To run background

  comfy launch --background

  comfy --workspace=~/comfy launch --background -- --listen 10.0.0.10 --port 8000

  • Instances launched with --background are displayed in the "Background ComfyUI" section of comfy env, providing management functionalities for a single background instance only.
  • Background-running ComfyUI can be stopped with comfy stop.

• To point every command (not just one) at a ComfyUI on a non-default address — e.g. a server you started outside comfy-cli on :8189 — export the COMFY_LOCAL_URL environment variable:

  export COMFY_LOCAL_URL=http://127.0.0.1:8189

  • Accepts http://host:port, host:port, or http://host (the port defaults to 8188; the scheme is optional and, if present, must be http). IPv6 literals are bracketed: COMFY_LOCAL_URL=http://[::1]:8189.
  • Honored by comfy env, comfy run, comfy jobs, comfy upload/download, comfy nodes, and every other local-targeting command, so comfy env's "Comfy Server Running" line now probes and reports the resolved address.
  • Precedence (per command): a per-command --host/--port flag wins, then COMFY_LOCAL_URL, then a comfy-cli-launched background server, then the 127.0.0.1:8188 default. A malformed value is ignored with a one-line stderr warning rather than breaking the command.

• to run ComfyUI with a specific pull request:

  comfy install --pr "#1234"

  comfy install --pr "jtydhr88:load-3d-nodes"

  comfy install --pr "https://github.com/comfyanonymous/ComfyUI/pull/1234"

  • If you want to run ComfyUI with a specific pull request, you can use the --pr option. This will automatically install the specified pull request and run ComfyUI with it.
  • Important: The --pr option cannot be combined with --version or --commit and will be rejected if used together.

• To test a frontend pull request:

  comfy launch --frontend-pr "#456"
  comfy launch --frontend-pr "username:branch-name"
  comfy launch --frontend-pr "https://github.com/Comfy-Org/ComfyUI_frontend/pull/456"
  

  • The --frontend-pr option allows you to test frontend PRs by automatically cloning, building, and using the frontend for that session.
  • Requirements: Node.js and npm must be installed to build the frontend.
  • Builds are cached for quick switching between PRs - subsequent uses of the same PR are instant.
  • Each PR is used only for that launch session. Normal launches use the default frontend.

  Managing PR cache:

  comfy pr-cache list              # List cached PR builds
  comfy pr-cache clean             # Clean all cached builds
  comfy pr-cache clean 456         # Clean specific PR cache
  

  • Cache automatically expires after 7 days
  • Maximum of 10 PR builds are kept (oldest are removed automatically)
  • Cache limits help manage disk space while keeping recent builds available

• To check VRAM/RAM usage: comfy system-stats (add --where cloud to target Comfy Cloud instead of local)

• To unload models / free the executor cache: comfy free (pass --free-memory to also reset the executor cache)

Managing Custom Nodes

comfy provides a convenient way to manage custom nodes for extending ComfyUI's functionality. Here are some examples:

• Show custom nodes' information:

comfy node [show|simple-show] [installed|enabled|not-installed|disabled|all|snapshot|snapshot-list]
                             ?[--channel <channel name>]
                             ?[--mode [remote|local|cache]]

• comfy node show all --channel recent

  comfy node simple-show installed

  comfy node update all

  comfy node install comfyui-impact-pack

  Note: the argument is the node's Comfy Registry ID, which is lowercase (e.g. comfyui-impact-pack), not the GitHub repository name (ComfyUI-Impact-Pack). Passing the repo-name casing fails to resolve with an error like Node 'ComfyUI-Impact-Pack@unknown' not found. Find a node's ID on its Comfy Registry page or via comfy node show all.

• Managing snapshot:

  comfy node save-snapshot

  comfy node restore-snapshot <snapshot name>

• Install dependencies:

  comfy node install-deps --deps=<deps .json file>

  comfy node install-deps --workflow=<workflow .json/.png file>

• Generate deps:

  comfy node deps-in-workflow --workflow=<workflow .json/.png file> --output=<output deps .json file>

install vs registry-install

comfy-cli offers two ways to install a custom node, backed by different mechanisms:

• comfy node install <id>... delegates to ComfyUI-Manager's cm-cli. It accepts one or more node IDs, resolves them through the Manager's channel database (--channel, --mode), and installs dependencies via the Manager (so it also supports --fast-deps, --no-deps, and --uv-compile). This is the recommended command for day-to-day node management, and it requires ComfyUI-Manager to be present in the workspace.

  comfy node install comfyui-impact-pack

• comfy node registry-install <id> [--version <v>] talks directly to the Comfy Registry API. It downloads the published archive for a single node (optionally pinned with --version), extracts it into custom_nodes/, and runs the node's own install script. It does not go through ComfyUI-Manager, so it does not require the Manager to be installed — this is the command the Registry's own install instructions use.

  comfy node registry-install comfyui-impact-pack

  comfy node registry-install comfyui-impact-pack --version 1.0.0 # install a specific version

  Because registry-install bypasses the Manager's dependency machinery and simply runs the node's bundled install script, it only accepts --force-download — it does not accept --fast-deps, --no-deps, or --uv-compile. If you want fast/unified dependency resolution, use comfy node install instead.

Unified Dependency Resolution (--uv-compile)

Requires ComfyUI-Manager v4.1+. Instead of installing dependencies per-node with pip install, --uv-compile delegates to ComfyUI-Manager's unified resolver which batch-resolves all custom node dependencies via uv pip compile with cross-node conflict detection — it can identify which node packs have incompatible dependencies and why.

• Install with unified resolution:

  comfy node install comfyui-impact-pack --uv-compile

• Available on: install, reinstall, update, fix, restore-snapshot, restore-dependencies, install-deps

• Run standalone (resolve all existing custom node dependencies):

  comfy node uv-sync

• --uv-compile is mutually exclusive with --fast-deps and --no-deps — except on restore-snapshot, where --fast-deps selects the --uv-compile path instead (see --fast-deps below).

• To make --uv-compile the default for all commands, see uv-compile default below.

• Use --no-uv-compile to override the default for a single command:

  comfy node install comfyui-impact-pack --no-uv-compile

--fast-deps

--fast-deps swaps comfy-cli's dependency installation from pip to comfy-cli's built-in uv-based resolver (DependencyCompiler), which is significantly faster and only requires uv (no ComfyUI-Manager). On a dependency version conflict it prompts you interactively to pick a version.

• Accepted by: comfy install, comfy node install, comfy node reinstall, and comfy node restore-snapshot.

  comfy install --fast-deps

  comfy node install comfyui-impact-pack --fast-deps

• Not accepted by comfy node registry-install: that command installs a node directly from the Comfy Registry by running the node's own bundled install script, so it never touches comfy-cli's dependency resolver (see install vs registry-install above).

• Mutually exclusive with --no-deps and --uv-compile, where those flags exist — the exact pairing is per-command:

  Command	Rejected combination
  comfy node install	--fast-deps --no-deps, --fast-deps --uv-compile
  comfy node reinstall	--fast-deps --uv-compile (no --no-deps on this command)
  comfy node restore-snapshot	--fast-deps --no-uv-compile (see below)
  comfy install	none (neither flag exists on this command)

• Exception — comfy node restore-snapshot: here --fast-deps selects the --uv-compile fast path rather than conflicting with it. ComfyUI-Manager's cm-cli has no --no-deps on restore-snapshot, and its --uv-compile already implies no-deps internally, so the two are synonyms on this command. Combining --fast-deps with --no-uv-compile is the contradiction, and it errors: Cannot use --fast-deps with --no-uv-compile.

--fast-deps vs --uv-compile

Both flags use uv for faster dependency resolution, but they work differently:

When to use which:

• For initial ComfyUI installation with uv: comfy install --fast-deps
• For custom node management with Manager v4.1+: --uv-compile (recommended)
• For custom node management with older Manager: --fast-deps

Bisect custom nodes

If you encounter bugs only with custom nodes enabled, and want to find out which custom node(s) causes the bug, the bisect tool can help you pinpoint the custom node that causes the issue.

• comfy node bisect start: Start a new bisect session with optional ComfyUI launch args. It automatically marks the starting state as bad, and takes all enabled nodes when the command executes as the test set.
• comfy node bisect good: Mark the current active set as good, indicating the problem is not within the test set.
• comfy node bisect bad: Mark the current active set as bad, indicating the problem is within the test set.
• comfy node bisect reset: Reset the current bisect session.

Managing Models

• Model downloading

  comfy model download --url <URL> ?[--relative-path <PATH>] ?[--set-civitai-api-token <TOKEN>] ?[--set-hf-api-token <TOKEN>]

  • URL: CivitAI page, Hugging Face file URL, etc...
  • You can also specify your API tokens via the CIVITAI_API_TOKEN and HF_API_TOKEN environment variables. The order of priority is --set-X-token (always highest priority), then the environment variables if they exist, and lastly your config's stored tokens from previous --set-X-token usage (which remembers your most recently set token values).
  • Tokens provided via the environment variables are never stored persistently in your config file. They are intended as a way to easily and safely provide transient secrets.

• Model remove

  comfy model remove ?[--relative-path <PATH>] --model-names <model names>

• Model list

  comfy model list ?[--relative-path <PATH>]

Running on Comfy Cloud (--where cloud)

Comfy Cloud runs your workflow on Comfy's GPUs — no local ComfyUI install and no GPU required. The same verbs you use locally take --where cloud; the only extra step is signing in once.

Prerequisites — a Comfy account with a credit balance (add credits; cloud runs are metered per GPU-second):

comfy cloud login                   # opens your browser (OAuth + PKCE), stores a session
comfy cloud whoami                  # confirm who you're signed in as
comfy cloud status                  # workspace, credit balance, tier, concurrency limit

Then submit, watch, and collect:

comfy run --workflow my_workflow_api.json --where cloud   # submits, prints a prompt_id, returns
comfy jobs ls --where cloud                               # queue + history for your account
comfy jobs status <prompt_id> --where cloud               # one job's state
comfy jobs watch <prompt_id> --where cloud                # live progress until it finishes
comfy download <prompt_id> --where cloud                  # fetch the outputs

Notes:

• --where cloud is per-invocation. Make it the default with comfy set-default --where cloud; every command then honors it without the flag, and --where local overrides it for one call. Clear it again with comfy set-default --clear-where.
• Add --wait to comfy run to block until the job completes instead of returning immediately.
• comfy run --prompt "<text>" (no --workflow) runs a bundled default text2img graph. That graph loads the SD1.5 checkpoint v1-5-pruned-emaonly.ckpt, which comfy-cli does not download for you — install it into models/checkpoints, or point the graph at a checkpoint you do have with comfy run --prompt "…" --set checkpoint=<name>. The same applies on cloud, where the checkpoint must exist in your cloud assets.
• comfy download also reads a prompt_id from piped stdin, so comfy run ... --where cloud | comfy download --where cloud works.
• Models and custom nodes must exist on the cloud side. comfy models search --where cloud lists the cloud asset catalog, and comfy nodes ls --where cloud lists the node classes cloud can run.
• Sign out with comfy cloud logout. If a run fails with cloud_unauthorized, your session expired — re-run comfy cloud login.

Calling partner nodes (comfy generate)

comfy generate calls Comfy's partner nodes directly from the terminal — no local ComfyUI or workflow JSON required. It hits the same hosted partner nodes you'd otherwise wire into a ComfyUI workflow, but as one-shot CLI calls. Image models (Flux, Ideogram, DALL·E, Recraft, Stability, Runway, Reve, xAI Grok, Google Gemini Flash Image aka nano-banana, …) and video models (Kling, Luma, Runway Gen-3, Pika, Vidu, Moonvalley, Hailuo, Grok video, ByteDance Seedance) are all covered; video jobs run async and the CLI polls until the result is ready.

Prerequisites — a Comfy API key and a credit balance:

• Create an API key
• Browse partner nodes and per-call credit costs · pricing table
• Add credits

Set the key once, then go:

export COMFY_API_KEY=comfyui-...   # or pass --api-key on each call

comfy generate list                                  # browse available models
comfy generate schema flux-pro                       # see params for one model
comfy generate flux-pro --prompt "a cat on the moon" \
    --width 1024 --height 1024 --download cat.png

Reference images can be passed as local paths — the CLI uploads them through the cloud's storage endpoint (or base64-encodes inline, as each partner requires):

comfy generate flux-kontext --prompt "add a top hat" \
    --input_image ./photo.jpg --download out.png

comfy generate upload ./photo.jpg                    # explicit upload

Async models (every video model plus the Flux family) block until ready by default. Pass --async to return immediately with a job id, then resume later with comfy generate resume <model> <job_id>. Examples:

comfy generate kling --prompt "a paper boat drifting on a river at dusk" \
    --duration 5 --download boat.mp4

comfy generate luma --prompt "..." --aspect_ratio 16:9 --async
# → prints job id; resume with:
comfy generate resume luma <job_id> --download out.mp4

Gemini Flash Image (nano-banana) — text-to-image and image edits in one alias. Pass --image (repeatable) for reference images. The response is inline base64, so --download is required to save:

comfy generate nano-banana --prompt "a watercolor of a sleeping fox" \
    --download fox.png

# Image edit — reference accepted as a local path, http(s) URL, or data URI:
comfy generate nano-banana --prompt "add a top hat" \
    --image ./cat.png --download out.png

# Switch model variants:
comfy generate nano-banana --prompt "..." --model gemini-3-pro-image-preview \
    --download out.png

Seedance — text-to-video and image-to-video, up to 1080p / 12s clips. Resolution, ratio, duration, fps, etc. get passed through as flags; the CLI inlines them into Seedance's prompt syntax for you:

comfy generate seedance --prompt "a hummingbird hovering over a flower" \
    --resolution 1080p --duration 5 --download bird.mp4

# Image-to-video: pick a lite/i2v variant and pass a first frame.
comfy generate seedance --model seedance-1-0-lite-i2v-250428 \
    --prompt "the wave crests and crashes" \
    --image ./still.jpg --download wave.mp4

Managing ComfyUI-Manager

• Disable ComfyUI-Manager completely (no manager flags passed to ComfyUI):

  comfy manager disable

• Enable ComfyUI-Manager with new GUI:

  comfy manager enable-gui

• Enable ComfyUI-Manager without GUI (manager runs but UI is hidden):

  comfy manager disable-gui

• Enable ComfyUI-Manager with legacy GUI:

  comfy manager enable-legacy-gui

• Clear reserved startup action:

  comfy manager clear

• Migrate legacy git-cloned ComfyUI-Manager to pip package:

  comfy manager migrate-legacy

uv-compile default

Set --uv-compile as the default behavior for all custom node operations:

comfy manager uv-compile-default true

When enabled, all node commands (install, reinstall, update, fix, restore-snapshot, restore-dependencies, install-deps) will automatically use --uv-compile. Use --no-uv-compile on any individual command to override.

To disable:

comfy manager uv-compile-default false

Beta Feature: format of comfy-lock.yaml (WIP)

basic:

models:
  - model: [name of the model]
    url: [url of the source, e.g. https://huggingface.co/...]
    paths: [list of paths to the model]
      - path: [path to the model]
      - path: [path to the model]
    hashes: [hashes for the model]
      - hash: [hash]
        type: [AutoV1, AutoV2, SHA256, CRC32, and Blake3]
    type: [type of the model, e.g. diffuser, lora, etc.]

  - model:
  ...

# compatible with ComfyUI-Manager's .yaml snapshot
custom_nodes:
  comfyui: [commit hash]
  file_custom_nodes:
  - disabled: [bool]
    filename: [.py filename]
    ...
  git_custom_nodes:
    [git-url]:
      disabled: [bool]
      hash: [commit hash]
    ...

Curated model knowledge

Discovery commands can attach a knowledge block to their --json output: which model a name refers to, whether it is deprecated, ranked picks per capability, and known pitfalls. A signed-in install (comfy cloud login) fetches the bundle from the cloud knowledge channel on its own. Signed out, the fetch fails quietly and no block is emitted unless a bundle is already cached.

Point the CLI at a different bundle with either of:

export COMFY_KNOWLEDGE_URL=https://.../knowledge.json   # fetched and cached
export COMFY_KNOWLEDGE_FILE=/path/to/knowledge.json     # read directly, never cached

Check what is loaded:

comfy knowledge status
comfy knowledge resolve "Kling 3.0"
comfy knowledge pick lipsync

Turn it off again:

export COMFY_KNOWLEDGE_DISABLE=1

Signing out is not an off switch. Once a bundle is cached it keeps being served, stale or not. COMFY_KNOWLEDGE_DISABLE suppresses envelope enrichment outright; the comfy knowledge verbs keep working under it, since those are you asking for the bundle directly.

Enrichment only ever reads the cache, so no command waits on a fetch. The cache refreshes during comfy skills install and in the background during comfy launch, or on demand with comfy knowledge status --refresh.

Analytics

Analytics are opt-in and off by default. The first time you run the CLI in an interactive terminal it asks whether to enable tracking, and that prompt defaults to no. Nothing is sent unless you answer yes. A non-interactive run (a pipe, CI, an agent) never enables it on its own. Setting DO_NOT_TRACK or COMFY_NO_TELEMETRY in the environment overrides the setting and suppresses everything.

Change your mind at any time:

comfy tracking enable
comfy tracking disable

When tracking is on, we use Mixpanel to understand usage patterns and know where to prioritize our efforts.

One event carries text you typed. If a curated knowledge bundle is loaded (a signed-in install fetches one on its own), then comfy nodes search, comfy templates ls, comfy generate schema, comfy generate list and comfy models search send the search terms you typed and which curated entries they matched. This tells us what people look for and fail to find. It needs both the bundle and your tracking consent, so with either one absent nothing is sent.

Check out the usage here: Mixpanel Board

Contributing

We welcome contributions to comfy-cli! For ideas, suggestions, or bug reports, open an issue at Comfy-Org/comfy-cli. For code changes, fork the repo and open a pull request.

See CONTRIBUTING.md for setup, the checks CI enforces, and PR conventions. Notable changes are recorded in CHANGELOG.md.

License

Released under the GNU General Public License v3.0.

Support

Questions or issues? Open an issue or reach us on Discord.

Happy diffusing with ComfyUI and comfy-cli! 🎉