IconMind vs SVG Repo

SVG Repo aggregates. Its collections come from hundreds of sources with hundreds of authors, which is why it can answer almost any query and why the results for a single query rarely look like each other. Search for server and you get a flat illustration, a thin outline, a filled glyph and a 3D render on the same screen.

That is fine when the icon lands on a slide, a blog post or an empty state. It is a problem in a toolbar, because the eye reads inconsistency before it reads meaning. Two icons drawn by two people at two stroke weights look like a bug even when both are good drawings.

IconMind is one set with one grid. Every icon is 24 by 24, every anchor sits on a half unit, every angle is 0, 45 or 90 degrees, and every stroke terminal is round. That constraint is what makes 2,287 icons look like they were drawn in one sitting.

SVG Repo labels each icon with the licence of the collection it came from. Much of it is CC0 or MIT and needs nothing from you. Some of it is CC BY and requires attribution wherever the icon appears, which for a shipped application means a credits screen and a record of which icon came from where.

That bookkeeping is manageable for one icon and unpleasant for two hundred. It is also the part teams skip, and skipping it is a licence violation rather than a style preference.

IconMind is MIT for the entire set, including every duotone cell, the sprite sheet and the generated components. One line in a licence file covers all of it, and it does not change when the set grows.

CC BY is not a difficult licence, but it does ask for something specific: credit the author, name the licence, and link both, wherever the work appears. For a blog post that is a caption. For an application it means a credits screen and a record of which icon came from which collection, kept up to date as icons change.

The usual failure is not malice. Someone downloads twelve icons in an afternoon, the tabs close, and six months later nobody can say which of them needed credit. At that point the only honest fix is to find them all again.

If you are going to mix sources, keep a short file in the repository next to the icons: file name, source URL, licence, author. It takes a minute per icon and it is the difference between an audit that takes an hour and one that takes a week.

IconMind avoids the question rather than answering it. Every cell is MIT, no attribution is required, and that stays true as the set grows.

Downloading SVGs means owning them. They land in a folder, someone writes a wrapper to give them a size and a colour, and from then on updating means downloading again and diffing by hand. Nothing is versioned, so there is no way to tell what changed between the icons you shipped in March and the ones you shipped in June.

A package removes that work. Import the icon, pass a weight, pass a class, and the bundler drops everything you did not use. Upgrading is one number in a manifest, and the changelog says what moved.

You can still have the files if you want them. Every IconMind cell exists as a plain SVG on disk and in the published sprite, so a design tool or a build script can read them directly without touching a framework.

Search SVG Repo for an agent and you get a person in a suit, an insurance mascot and a few robots. Search for retrieval, reranker, context window or tool call and the results stop being about software at all. The vocabulary of AI products is younger than most of the collections in the index.

IconMind spends 664 of its 2,287 icons there, arranged as families rather than singles: agent, agent run, agent handoff, agent memory; model, model training, model eval, model registry; retrieval, chunking, embedding, reranker, vector database. Those are the screens the set exists for.

Where SVG Repo wins outright is everything else. Anything decorative, illustrative, brand-related or physical is better served by an index of half a million drawings than by a set that deliberately refuses whole categories.

Keep the interface consistent and let the illustrations be borrowed. Navigation, toolbars, tables, form controls and empty-state chrome should come from one set. A hero illustration, a marketing spot or a one-off graphic can come from anywhere, because nothing sits next to it at 20 px.

If a borrowed icon has to sit in the interface, match it: 24 by 24 box, 2 px stroke, round caps and joins, and no fill. That is enough for most icons to pass unnoticed.

And check the licence before it ships, not after. The field is on the SVG Repo page for every icon, and it is the one piece of this that cannot be fixed later.

A general index searches the words attached to a drawing by whoever uploaded it. That is why searching an aggregator for a technical term returns either nothing or something from an unrelated field.

IconMind indexes four fields per icon: name, a description written for the search box, tags, and aliases for the words people actually type. Search llm and get model, search vector db and get vector-database, search handoff and get the agent family. Each of the 669 tag pages is also a page on the site, so the search works from Google as well as from the search box.

For assistants there is an MCP server. Given a description it searches the real metadata and returns slugs that exist, which is the difference between an import that compiles and a plausible guess that does not.

SVG Repo will hand you an SVG and, for many icons, a PNG. That covers the case where the icon is going into a document or a slide.

IconMind renders in the browser from the same source: SVG, PNG at 16 through 512, WebP, JPEG, and a multi-resolution ICO for favicons, with the colour you picked applied before the render. Nothing is pre-baked except the PNG previews search engines index, so the file you download matches what is on screen.

There are also whole-set downloads: every cell as a file, the sprite sheet, and per-variant zips, all under one MIT licence with no per-file bookkeeping.

An aggregator cannot promise consistency because it does not own the drawings. A set can, and the honest way to make that promise is to let people verify it.

Every IconMind icon is a declaration compiled by a validator that refuses geometry it cannot draw correctly: anchors off the half-unit grid, angles that are not 0, 45 or 90 degrees, strokes that would merge at the bold weight, icons that do not fill the same optical box as their neighbours. A nightly job rasterises all 13,722 cells and fails the build if any two render alike.

The commands are in the repository and take a few minutes on a laptop. That is the practical difference between a set and a folder of downloads: you can run the check yourself rather than trusting a screenshot.

For a one-off, it wins and it is not close. A stethoscope, a tractor, a national flag, a brand mark, a piece of decorative art: half a million drawings will have something and 2,287 will not.

For anything illustrative, it wins too. IconMind is a 24 px interface set and its rules stop it from ever producing an illustration.

For a shipped interface, the calculation flips. Coherence, one licence, versioned packages and a set that grows in the same hand are worth more than coverage you will not use, and that is the whole argument for a set over an index.

The SVG Repo figures are its own published counts as of 31 August 2026, and the licence description comes from the per-icon fields on the site, which name the source collection and its terms.

The IconMind figures are the set itself: 2,287 icons, 13,722 cells across two variants and three weights, 664 icons in the AI, agent, MCP and RAG families, and one MIT licence covering all of it. Package sizes come from the npm registry on the same date, read from the published tarball: 5.6 MB unpacked across 6,867 files.

The consistency claims are the ones worth checking rather than believing. The validator, the duplicate scan and the audit all run from the repository in a few minutes on a laptop, and they either pass or they name the icon that failed.