comfyui-browser: an index over a folder, and an install command pointing somewhere else

The repository is the plugin; the hosted product is the pitch

The first line of the readme sends anyone who wants a user-friendly agentic solution to a hosted site from the same team, and that tells you where this repository sits. It is the ComfyUI custom node. The commercial product sits above it. So what you are evaluating is a local browser and manager, not an assistant, and the feature list matches that exactly. You browse and manage the images, videos and workflows sitting in the output folder. You add the workflows you care about to a saved collection so you can switch between them instead of hunting for files. You synchronise that collection anywhere with Git, and you subscribe to other people's workflow sources the same way. You search your workflows by keyword. Two custom nodes come with it: one that lets you select any input of the current graph, and a small plot node where you pick inputs and fill in values. The design bet underneath all of this is unglamorous and correct. ComfyUI's output directory is a folder with no index, and a workflow is a JSON graph with a filename you chose once. This adds the index and the naming, and then uses Git, which every developer already has, as the sync and sharing layer instead of inventing an account system.

Four configurable paths, and one of them follows the host's own flag

A configuration file in the extension's directory overrides four paths, and the defaults are worth reading closely because they are not all relative to the same thing. Three are relative to the extension's own install directory, which the tool works out for itself: the collections folder, the download logs folder, and the sources folder. The fourth, outputs, is relative to the ComfyUI installation instead. And then there is the part that is genuinely well designed: outputs has no path of its own to default to. It follows whatever the output directory argument was when ComfyUI was started, and if that argument was not passed, it follows ComfyUI's own default. So if you already move your outputs with the host application's flag, this browser moves with it and you configure nothing. If you override outputs here, you are taking over a link the tool would otherwise have maintained for you.

The default configuration values are:

For a first-time setup, the one setting that matters is collections, because that is where your saved workflows live. Everything else is either logs or a cache of things you did not write.

Where `[comfyui-browser]` is the automatically determined path of your
`comfyui-browser` installation, and `[comfyui]` is the automatically determined
path of your `comfyui` server.  Notably, the `outputs` directory defaults to the
`--output-directory` argument to `comfyui` itself, or the default path that
`comfyui` wishes to use for the `--output-directory` argument.

### Manually

Clone this repo into the `custom_nodes` folder and restart the ComfyUI.

The manual install command clones a different repository

There are two documented ways in. The first is through the extension manager that ComfyUI users already have installed: search for the package in the custom node installer and install it. The second is a manual clone into the custom nodes folder followed by a restart, and the command for that names a repository under a different account than the one you are currently reading. The invitation to submit workflow sources points at the same other account as the assigned maintainer. That pattern reads like a project that was renamed or moved between accounts, with one copy of the instructions left behind after the move. The practical consequence is not cosmetic. If you follow the manual command verbatim you install somebody else's copy of the plugin, and if the two copies have diverged, which is exactly what happens when a fork stops tracking its origin, you will spend your afternoon debugging a version you did not intend to run and whose readme is not the one you just read. Two habits avoid the whole problem. Install through the manager, which resolves the repository you are pointed at. Or, if you want the manual route, clone the URL of the repository you are actually reading rather than the one in the command. It is also worth noting that the project ships a Chinese translation of its readme, so the instructions exist in two languages and only one of them was updated.

Two frontends, one loaded by ComfyUI and one loaded as an iframe

The project structure explains something about the code that a feature list would not. A Python file at the root is the backend server. A web directory holds the frontend that ComfyUI itself loads, containing a Svelte bundle and a small entry script whose job is to talk to ComfyUI. Then there is a second, separate Svelte directory holding the frontend that appears inside a modal as an iframe. Two frontends in a project this size is unusual, and the development instructions say why. To work on it you start the backend by running ComfyUI itself with a flag that enables cross-origin headers, then start the Svelte app on its own development server and point it at the running ComfyUI instance through a query parameter, with one local port serving the development server and the other the model server. So the interface you use and the interface you develop are two builds of the same source served two different ways. The benefit is hot reload on a UI that otherwise lives inside another application's frame, which is genuinely painful to work with otherwise. The cost is that development requires modifying how the host application starts, and the cross-origin flag is not optional in that setup.

Five dependencies listed twice, two of them heavy for a file browser

The package manifest and the requirements file name the same five libraries: a progress bar, a dataframe library, the array library, an HTTP client, and a retrying transport. Listing them in two files is a small maintenance smell, since the two can drift and there is no mechanism described that keeps them in step. The more interesting question is what a file browser and a Git synchroniser are doing with a dataframe library and an array library. Nothing in the documentation describes an operation that needs either; they read like defaults that arrived with the first commit rather than like requirements somebody derived. That matters more than it would in a standalone application, because a ComfyUI custom node shares one Python environment with the host application and with every other node the user has installed. A plugin's dependency list is a cost imposed on everyone in that environment, and a plugin that pulls a dataframe library to do nothing visible in the readme is imposing it for no stated reason. The other metadata detail worth checking is the licence, which is given as a path to a file rather than as an identifier, while the repository's own licence field is not set. A licence file is present, so the intent is clear, and a tool reading the metadata programmatically sees nothing.

The registry block is filled in except for the two fields that matter

The manifest ends with a section for the Comfy Registry, and a comment in the file explains that the publisher identifier is what that registry uses. The display name is filled in. The publisher identifier and the icon are both empty strings. Read one way this is consistent with how the plugin is actually installed, because the manager installs from a Git repository rather than resolving a registry entry, so the manifest's registry block describes a path this project is not currently on. Read another way it means the manifest is not the source of truth for how the plugin is found: it is found by name, through a search in the manager, against a repository URL the user never sees. For anyone auditing or repackaging this, the practical reading is that the registry block is aspirational while everything else in the manifest is functional. The version is a plain one point zero, and there are no tags published, so the version number has no release history behind it and nothing to compare it against. A checklist item for your own review, then: if you are distributing this to other people, the publisher identifier is the field that has to be filled in before a registry listing will work, and it is empty.

The maintainer asks for Windows testing because they do not have a Windows machine

One line in the development notes carries more information about the state of the project than any feature in the list: please try to test on Windows, because the author only has Linux and macOS. That is worth taking at face value rather than reading past, because a ComfyUI custom node is a Python backend plus a browser frontend and the frontend is where platform differences show up. The project also asks for contributions in a second form, since the issue tracker doubles as the way to submit a workflow source through a template for adding a new workflow repository. That is an unusual review burden, because what arrives is data rather than code: a graph someone else will load and run, with no test to write against it. The changelog lives in its own file, and the checked list at the bottom of the readme shows the last two items ticked: synchronising collections to a remote Git repository, and adding an external repository to the local collections. The one still open is searching workflows by node name and model name, which is the search feature's obvious next step given that keyword search already exists. There are no tagged releases, and the last commit on record is dated 9 August 2026.

Editorial conclusion

Use it if your ComfyUI output directory has become the place your work actually lives, which is the normal state of affairs after a few months of generating, because a folder of images with no index and a set of workflow files with no names is exactly the problem it solves and nothing else does it well. Do not use it expecting an agent or an assistant; the readme's first line points at a hosted product from the same team for that, and this repository is the local plugin underneath. Two things to check before you rely on it. Install through the extension manager rather than by hand, because the manual command in the documentation clones a different repository URL from the one you are reading and a rename is the probable reason. And decide where your collections directory lives before you start saving work into it, because that is the setting that decides whether your saved workflows are portable or trapped inside an install directory that a reinstall will replace. On platform coverage, take the maintainer's own note seriously and test on anything that is not Linux or macOS.

Frequently asked questions

Official sources

1. Issues
2. README
3. talesofai/comfyui-browser on GitHub