Find a Rust library or app for your machine learning job, with real usage data for each project.
Browse the catalogue · State of ML in Rust · Tier rules · Add a project
It covers crates on crates.io and apps outside it, including burn, candle, ort, polars, qdrant, rig and tantivy.
- Run and serve models
- Build LLM apps and agents
- Train models
- Speech and audio
- Text, embeddings and search
- Images and vision
- Video
- Documents and OCR
- Tabular data and classic ML
- Math and science
- Tensors and GPU compute
- Datasets and model formats
Or search everything. Search runs in your browser.
- What it does: a short description and the project's README.
- Whether people use it: downloads and dependents, shown first.
- Whether it is alive: release and commit activity, unpatched advisories and deprecation notices.
- Whether it fits: compatibility facts and similar projects. You can compare up to 4 projects.
- Where each number comes from: the raw value, a source link and the fetch date.
Each project gets one tier from published rules over measured data.
| Tier | Meaning |
|---|---|
| Featured | Shown first. It passes every adoption, health and docs gate. |
| Listed | Shown in the normal listing. |
| New | Its first release is recent. |
| Archived | No longer maintained, or ended by its author. |
- No LLM decides a tier, writes a claim on a project or topic page, or merges anything.
- The state of ML in Rust is the one page AI models write. It says so, and no tier reads it.
- GitHub stars alone never earn a tier. GitHub hides who starred, so stars cannot be screened for fakes.
- Every tier change appears on the changes page and in its feed.
The ML Rust organization runs this catalogue and maintains some of the listed projects.
- Its own projects pass the same rules as every other project.
- Featured for its own projects also needs approval from a maintainer outside the organization.
- The home page labels a hand-picked list of its projects. No tier reads that list.
Open the issue chooser and pick a form.
| Form | Use it to |
|---|---|
| Submit | Propose a library or app. |
| Correctness report | Report a wrong fact on an entry. |
| Topic change | Replace the topics of an entry. |
| Product or demo | Set an entry as a product or a demo. |
| Supersedes or fork | Report that one entry replaces another. |
| Maintainer statement | State that a project is maintained or finished. |
| Featured demotion | Approve a demotion that the rules propose. |
| Featured approval | Approve Featured for an org-affiliated entry. |
A person makes every approval and merges every change:
- The bot checks the form and posts a comment. An advisory link check adds text to that comment only.
- A maintainer listed in
data/maintainers.tomladdsapproved. The bot refuses an approval from a bot account or the submitter. - If the issue body changes after the check, the bot removes
approved. - The bot opens a data PR. A person reviews and merges it, and the site rebuilds.
Each judgment appears on its project page with a link to its issue.
| URL | Holds |
|---|---|
/llms.txt |
Where an AI agent starts. It links the small files below. |
/state/llms.txt |
The state report as plain text. |
/topics/<job>/llms.txt |
One line per project in a job. |
/projects/<slug>/llms.txt |
The facts of one project. |
/data/catalogue.json |
Every project and signal, as JSON. It is over 8 MB, so it is for scripts, not models. |
/changes/feed.xml |
Tier changes and new projects, as an Atom feed. |
Install Rust 1.95 or later. Run each command on its own line. CI adds --locked to the cargo builds.
cargo clippy --workspace --all-targets -- -D warnings
cargo nextest run --workspace --no-fail-fast
cargo fmt --all --check
cargo deny --exclude-dev check
cargo deny --exclude-dev --manifest-path crates/catalogue-search-wasm/Cargo.toml --target wasm32-unknown-unknown check
scripts/ci/build_search_wasm.sh OUT_DIR
shellcheck scripts/ci/*.sh deploy/*.sh
bash deploy/test_update.sh
cargo run -p catalogue -- validate --data data
cargo run -p catalogue -- build --no-tiersvalidatechecks every data file and every rule across files. Run it after any change underdata/.build_search_wasm.shbuilds the search wasm files. It needs the wasm-bindgen CLI at the versionCargo.lockpins.build --no-tiersreplacesdist/with a site without tier data. It does not write the search wasm files. The build needs either--reportor--no-tiers.
| Path | Holds |
|---|---|
crates/catalogue-core |
Data model, validation and tier rules. No IO. |
crates/catalogue-signals |
Reads the crates.io database dump, fetches the APIs and keeps signal history. |
crates/catalogue-bot |
Parses issue forms, writes check comments, and plans label and data changes. No IO. |
crates/catalogue-github |
The bot's GitHub App client. |
crates/catalogue-triage |
The bot's advisory link check. |
crates/catalogue-state |
Writes and checks the state report: what models read, the prompts and the checks on each answer. |
crates/catalogue-site |
Site templates, page data and refusal checks. No IO. |
crates/catalogue-search-wasm |
The browser search runtime. |
crates/catalogue |
The catalogue CLI. |
data/ |
Entries, judgments, maintainers, topics, categories, synonyms, discovery rules, tier thresholds, deprecated requirements, backends and the showcase. |
deploy/ |
systemd units and scripts for the bot VM. |
- The site deploys from
mainthrough.github/workflows/pages.yml. - A push deploys in minutes with the newest scores.
- A push that changes
data/fetches new entries only, then applies the rules to every entry. - A push that changes only the scoring crates fetches nothing, then applies the rules to every entry.
- The weekly scan runs on Sunday at 22:23 UTC. It fetches what changed.
- When the last scan is over 8 days old, the next push or manual run scans instead, unless the manual run asks for
allorchanged. - To deploy by hand, run
gh workflow run pages.yml. Add-f mode=new,none,changedorallto score first.alltakes over an hour. - The Pages workflow needs the
SIGNALS_TOKENsecret: a fine-grained token with public repositories read-only. - The bot runs on a VM. Run
sudo bash deploy/install.shon it. The script prints each input still missing, with the command that adds it.
- Code: Apache License 2.0 (
LICENSE). No crate is published to crates.io. - Data: CC-BY-4.0 (
data/LICENSE) coversdata/.