Find the Python dependencies that have no prebuilt wheel for your target platform, before your Docker build does.

Point it at your lockfile and the platforms you ship to (ARM servers, Alpine images, Windows on ARM, free-threaded Python 3.14t). It tells you which packages pip or uv will have to compile from source, and which can't be installed at all.

📊 New: wheel coverage for the top 5,000 PyPI packages, by platform and Python version, updated nightly.

Install and run (under 30 seconds)

With uv:

uvx wheelgap

Or with pipx:

pipx run wheelgap

Or install it: pip install wheelgap.

Run either one in a project directory. It finds uv.lock, pylock.toml, poetry.lock, pdm.lock, Pipfile.lock, requirements.lock or requirements.txt on its own.

Usage

See the recipes for multi-arch images, monorepos with several lockfiles, and a free-threading readiness report you can use in scripts.

# Will my Alpine ARM image compile anything?
wheelgap -p 3.13 -t alpine-arm64

# Is my stack ready for free-threaded Python 3.14?
wheelgap -p 3.14t -t all

# Several Pythons at once, Linux on both architectures
wheelgap uv.lock -p 3.12,3.13,3.14 -t linux-x86_64,linux-aarch64

# No -p? A .python-version file next to uv.lock (written by uv or pyenv) sets the Python
wheelgap uv.lock -t linux-x86_64

# Check a Pipenv project's default and development dependencies
wheelgap Pipfile.lock -p 3.13 -t linux-x86_64,linux-aarch64

# Check a PDM project's locked distribution files
wheelgap pdm.lock -p 3.13 -t linux-x86_64,linux-aarch64

# No lockfile? Check packages directly (latest version, or pin one)
wheelgap psycopg2 uvloop numpy==2.3.4 -t windows,win-arm64

# Take the target from the final stage's image (e.g. python:3.13-alpine
# or ghcr.io/astral-sh/uv:0.13.7-python3.13-trixie-slim), following
# named stages like `FROM base AS runtime` back to the image they start from.
# A python X.Y tag without a Debian release (python:3.13-slim) is read as
# the release Docker Hub currently builds it on (trixie, glibc 2.41, for 3.9
# and later); patch tags like 3.12.7-slim keep the --glibc default. Pin
# -bookworm or -trixie for an answer that doesn't move
wheelgap --dockerfile Dockerfile

# Machine-readable, for scripts and CI
wheelgap --json > wheelgap.json

# Why is it a gap? List each rejected wheel and the reason
wheelgap -t alpine-arm64 --explain

Projects usually ship to the same platforms, so you can write the defaults down once in pyproject.toml (next to the lockfile, or in the current directory). Command-line flags still win.

[tool.wheelgap]
python = ["3.13", "3.14t"]
targets = ["linux-x86_64", "linux-aarch64", "musl-aarch64"]
glibc = "2.31"

For uv projects, wheelgap also reads [tool.uv] required-environments as default targets:

[tool.uv]
required-environments = ["sys_platform == 'linux' and platform_machine == 'aarch64'"]

Target selection uses -t first, then --dockerfile, then [tool.wheelgap] targets, then uv's required-environments, then the built-in defaults. Each marker is evaluated against the eight common native platforms using the selected Python versions. A platform is included if any marker matches for any selected Python, and the resulting platforms are checked for all selected Pythons. Linux markers select both glibc and musl targets, since markers cannot distinguish the libc; the example above selects linux-aarch64 and musl-aarch64. Use -t to choose one explicitly. An inferred selection is explained on stderr, including with --json or --markdown. The inferred targets replace the built-in defaults, so a project that only requires Intel macOS wheels is then checked for Intel macOS only: add [tool.wheelgap] targets or -t to check more. An absent or empty setting keeps the built-in defaults, and so does one matching none of the targets above (with a note). An invalid marker is an error. Only required-environments is used here: environments, extra Linux architectures, mobile and Pyodide targets are not inferred.

Each cell in the table is one of:

Packages that are fine everywhere are hidden; --all shows them.

With several targets, the summary also counts the gaps on each target that has any (musl-aarch64 1 from source, 1 missing). --markdown lists the same counts. --json has them for every target, with zeros where there are no gaps and also when only one target is checked, as summary.per_target, keyed by target label: {"linux-aarch64": {"source": 0, "missing": 0, "unknown": 0}, "musl-aarch64": {"source": 1, "missing": 1, "unknown": 0}}.

--explain lists, for each gap, the wheels that were rejected and why, closest misses first:

armlib 1.4.0 on musl-aarch64: ● source
  armlib-1.4.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl  manylinux_2_17_aarch64 is a glibc wheel; musl (Alpine) needs musllinux

For uv.lock inputs, --explain also shows why the package is installed at all: the shortest dependency path from your project (or from a dependency group, shown as (dev group)) to the gap package. With extra/group selection it's a path the selection installs it through, not always the shortest. Paths are worked out per target, since markers can pull a package in on one platform only. With --json --explain, each package gets a matching via object keyed by target, like rejected: "via": {"win-arm64": ["myapp", "uvicorn", "uvloop"]}. Other inputs have no dependency graph, so they show no path (and "via": {}).

The shape of --json output is described by docs/wheelgap-output.schema.json. It matches the wheelgap version it ships with, and later versions may add keys. In --json output, results, rejected and via are all keyed by target label: just the platform ("linux-x86_64") when one Python is checked, or Python and platform ("3.13 linux-x86_64") with several. The top-level targets list gives each target's label, python and platform, so a script can look the key up instead of building it:

import json

data = json.load(open("wheelgap.json"))
key = next(t["label"] for t in data["targets"] if t["python"] == "3.13" and t["platform"] == "linux-x86_64")

uvloop 0.21.0 on win-arm64: ● source
  via: myapp → uvicorn → uvloop

If you think wheelgap got a package wrong, that output (or --json --explain) is all a bug report needs, with no need to share your lockfile.

Exit codes: 0 no gaps, 1 gaps found, 2 bad input. --exit-zero exits 0 even when gaps are found; bad input still exits 2. --fail-on missing (or fail-on = "missing" in [tool.wheelgap]) accepts source builds and only exits 1 when a package can't be installed at all.

If an input can't be read (a missing lockfile, an unknown target, a bad [tool.wheelgap] value), --json prints {"error": {"message": "...", "type": "TargetError"}} on stdout instead of results, next to the usual wheelgap: error: line on stderr and exit code 2. Command-line usage errors from argparse are still plain text on stderr.

Use --ignore psycopg2 for accepted gaps (repeatable or comma-separated, with case-insensitive package-name matching and equivalent hyphens, underscores and dots): matched packages are excluded from the table, counts and exit status, and listed as ignored in text (including quiet mode), JSON and Markdown. Set ignore = ["psycopg2"] in [tool.wheelgap] to use it by default; any --ignore flags replace the configured list rather than adding to it, and --ignore "" clears it for one run.

To accept a gap on one target only, add the target after a colon: --ignore psycopg2:musl-x86_64 (or ignore = ["psycopg2:musl-x86_64"]) still checks psycopg2 everywhere else. The target takes anything -t does (alpine, linux/amd64, all), one per entry (--ignore psycopg2:alpine,psycopg2:win-arm64). An unknown target is an error (exit 2), and so is a second target after a comma (psycopg2:alpine,win-arm64), which would otherwise read as a package name. The package keeps its real status in the report (● source (ignored) in the table), but a gap on that target doesn't count towards the summary or the exit status, isn't explained by --explain, and is a skipped testcase in --junit. Reports list it as ignored: psycopg2 (musl-x86_64), and --json puts it in ignored_targets (ignored keeps only packages ignored everywhere). A target that isn't checked in this run gets a note on stderr.

Use -q / --quiet (or quiet = true in [tool.wheelgap]) to show only the gap summary and package names. Clean runs without ignored packages are silent, JSON and Markdown output are unchanged, and quiet text output cannot be combined with --explain (--json --explain and --markdown --explain work).

Check the dependencies you actually install

With no selection options, wheelgap keeps its conservative behavior: all project/workspace extras and dependency groups count. For uv.lock inputs, select a deployment's dependencies explicitly:

# Match a production uv sync --no-dev --extra api
wheelgap uv.lock --no-dev --extra api -t linux-aarch64
# Include every project extra, or add a dependency group
wheelgap uv.lock --all-extras
wheelgap uv.lock --group docs
# Check only a group, without project dependencies or default groups
wheelgap uv.lock --only-group docs

--extra, --group and --only-group are repeatable. Once any selector is used, the project's runtime dependencies and uv's default groups (dev unless [tool.uv].default-groups changes it) count, unless --only-group is given; project extras are included only when requested. --no-dev removes dev, while --only-group excludes project dependencies and default groups. Combine --only-group with --group to add more groups, but not with --extra/--all-extras. An explicit dev group conflicts with --no-dev. Unknown extra/group names are errors. Command-line selectors need a uv.lock input, or for groups a Poetry 2.x poetry.lock or a pdm.lock, or for --extra/--all-extras a pyproject.toml (see below); configured groups apply to uv.lock, Poetry 2.x poetry.lock and pdm.lock inputs, and configured extras only to uv.lock inputs.

poetry.lock (Poetry 2.x) inputs accept --group, --only-group and --no-dev with Poetry's meanings, and configured groups apply to them too. main and every group not marked optional = true in pyproject.toml count; --group adds a group (like poetry install --with), --no-dev leaves out dev, and --only-group checks only the named groups (main only if named). A package counts if any of its groups is selected, with the markers of those groups. --extra isn't supported for poetry.lock yet, and a command-line selection on a Poetry 1.x lock (no per-package groups) is an error; configured groups leave a 1.x lock unfiltered.

pdm.lock inputs take the same selectors with PDM's meanings, and configured groups apply to them too. Every group the lock was made for (its [metadata] groups: default, the dev groups and any optional-dependencies groups locked with pdm lock -G) counts, as with a plain pdm install; --no-dev leaves out dev (like pdm install --without dev), and --only-group checks only the named groups (default only if named), so PDM's --prod is --only-group default here. --group only checks that the group is locked, since it already counts. A group that's in pyproject.toml but not in the lock is an error (lock it with pdm lock -G <name>), as in PDM. A package counts if any of its groups is selected. --extra isn't supported for pdm.lock, and a command-line selection on a lock without package groups (one locked without the inherit_metadata strategy) is an error; configured groups leave such a lock unfiltered.

Default selections can also be configured:

[tool.wheelgap]
extras = ["api"]
groups = ["docs"]

CLI --extra and --group replace their corresponding configured lists. --only-group suppresses both configured lists, and --no-dev removes a configured dev group. Empty lists explicitly select no extras or additional groups. With selection enabled, rooted workspaces start at the root project (members outside the selection aren't checked, and a note on stderr names them); virtual workspaces start at their members. Dependency-requested extras are followed transitively, including local workspace dependencies, rather than enabling every extra of every dependency.

Unresolved marker conditions (including uv's synthetic conflict markers and unknown OS releases) remain conservative: wheelgap can still report a possible gap rather than silently hide one. Selection itself needs no network. Index refreshes (--online, or a target outside the lockfile's requires-python) only look up packages in the selected graph. Configured extras only apply to uv.lock inputs, and configured groups to uv.lock, Poetry 2.x poetry.lock and pdm.lock inputs; other inputs are checked as usual.

Check a library's direct dependencies

Libraries usually don't commit a lockfile. Pass the library's pyproject.toml to check the dependencies it declares in [project] dependencies:

wheelgap pyproject.toml -t win-arm64 -t pyodide
# Add [project.optional-dependencies] groups
wheelgap pyproject.toml --extra test
wheelgap pyproject.toml --all-extras

Each requirement is read like a line of a requirements file, so ranges and bare names use the newest matching release and markers are respected. With several -p versions, that release isn't filtered by Requires-Python (#84), so ranges and bare names can show false gaps. Only direct dependencies are checked: their own dependencies aren't, and wheelgap always prints a note saying so. Use a lockfile for the full picture. If [project] requires-python leaves out a -p Python (every clause counts, as pip applies them), that target is still checked and a note names it; an invalid requires-python is ignored with a note. [dependency-groups] isn't read yet, so --group, --only-group and --no-dev are errors here, and a dependencies listed in [project] dynamic is an input error (exit 2), as are dynamic optional-dependencies with --extra/--all-extras and Poetry dependencies kept in [tool.poetry.dependencies] (check poetry.lock instead). wheelgap . in a directory with no lockfile doesn't switch to pyproject.toml on its own; the error suggests it. check_lockfile() and wheelgap diff don't read pyproject.toml.

Targets

Run wheelgap --list-targets to see target aliases and compatibility limits, or add --json for a machine-readable list.

If you know the deployment distro rather than its libc version, use --distro:

wheelgap uv.lock -p 3.13 -t linux-aarch64 --distro ubuntu-22.04
wheelgap uv.lock -p 3.13 -t musl-aarch64 --distro alpine

It supplies a glibc or musl version only: it does not select Python versions or platforms. Keep using -t for those platforms, including musl-* for Alpine. You can also set distro = "ubuntu-22.04" in [tool.wheelgap].

For each libc, an explicit --glibc / --musl wins, then --distro, then a version from --dockerfile, then a glibc / musl config value, then the config distro, then the built-in default. A distro only supplies its own libc; the other libc keeps its existing default. Names are case-insensitive; unknown names are errors. For older Alpine, pass --musl 1.1 explicitly. The text --list-targets output also lists distro names; its JSON output remains a list of platforms.

Only x86 Linux takes manylinux wheels older than glibc 2.17: every other architecture's first manylinux was manylinux2014 (glibc 2.17), so pip accepts nothing older there, and neither does wheelgap.

iOS and Android (PEP 730 and PEP 738, Python 3.13+) are checked with -t mobile or by name, for apps built with tools like BeeWare's Briefcase. Mobile dependencies are installed by a cross-platform pip run that can only use wheels, so a package with only an sdist shows as ✗ missing. --ios and --android-api (or ios = "15.0" and android-api = 28 in [tool.wheelgap]) set your app's minimum iOS version and Android API level. wheelgap checks PyPI only: Briefcase and Chaquopy also install from their own package repositories (BeeWare's on anaconda.org, and chaquo.com), which have mobile builds of packages like numpy, so a ✗ missing there may still install.

Pyodide (Python in the browser) only installs wheels: there's no compiler, so a package with only an sdist shows as ✗ missing, not ● source. Pyodide wheels are tied to one ABI per Python version and don't work across them. Pyodide also ships its own builds of many popular packages (numpy, pydantic-core and others) that aren't on PyPI. Point --pyodide-lock at the release's pyodide-lock.json (a path or https:// URL, also pyodide-lock = "..." in [tool.wheelgap]) and those count as ✓ bundled, with a note when Pyodide's version differs from the one you pinned (shown with -q too, and "bundled_version_differs": true in --json). That still passes: pyodide.loadPackage gives you Pyodide's version, though micropip.install of your exact pin would fail.

wheelgap -p 3.12 -t pyodide --pyodide-lock https://cdn.jsdelivr.net/pyodide/v0.27.7/full/pyodide-lock.json

Docker platform names also work with -t, for example -t linux/amd64,linux/arm64 (also linux/arm64/v8, linux/arm/v7, linux/ppc64le, linux/s390x, linux/riscv64 and linux/386, also spelled linux/i386).

-t all checks the eight common native platforms, and -t everything adds the armv7l, ppc64le, s390x, riscv64 and i686 Linux targets (glibc and musl) and win32. Pyodide, iOS and Android are only checked when you name them. Python versions take a trailing t for free-threaded builds (3.14t), which correctly rejects abi3 wheels. Without -p, wheelgap reads a .python-version file next to the lockfile (or pyproject.toml) if there is one, and otherwise uses the Python it is running on.

In CI (GitHub Action)

- uses: actions/checkout@v7
- uses: wheelgap/wheelgap@v0.9.0
  with:
    args: uv.lock -p 3.13 -t linux-x86_64,linux-aarch64

The job fails when a dependency would build from source on a target you ship, for example when a dependency bump drops ARM wheels. The gap table also goes to the run's summary page (set summary: false to turn that off).

Outside the Action, --markdown prints the same report as Markdown, for a job summary (wheelgap --markdown >> "$GITHUB_STEP_SUMMARY") or a PR comment. With --explain, the rejected wheels go in a collapsed section. --junit PATH also writes a JUnit XML file (one testsuite per target) that GitLab, Jenkins, Azure Pipelines and CircleCI can show as a test report, alongside whatever output mode you use.

With pre-commit

repos:
  - repo: https://github.com/wheelgap/wheelgap
    rev: v0.9.0
    hooks:
      - id: wheelgap

The hook runs when uv.lock, poetry.lock, pdm.lock, pylock.toml, Pipfile.lock, or a requirements*.txt or requirements*.lock file changes. Put your targets in [tool.wheelgap] in pyproject.toml so the hook checks the right platforms. Pass extra options with args:, for example args: [--fail-on, missing].

From Python

import wheelgap

targets = wheelgap.make_targets(python="3.13", platforms=["linux-aarch64", "musl-aarch64"])

result = wheelgap.check_package("numpy", "2.3.4", targets)  # looks the files up on PyPI
result.ok  # True if every target has a wheel
result.gaps  # the targets that don't
result.to_dict()  # {"name": ..., "results": {"3.13 musl-aarch64": "wheel", ...}}

for r in wheelgap.check_lockfile("uv.lock", targets):  # markers and the uv graph respected
    print(r.name, r.version, [t.platform for t in r.gaps])

check_lockfile(..., offline=True) never contacts an index: it raises wheelgap.sources.SourceError if an input has no locked file list and skips the re-check for Python versions outside the lock's requires-python (so those versions may show false gaps). It cannot be combined with online=True.

The package ships a py.typed marker (PEP 561), so mypy, pyright and other type checkers use its annotations.

check_packages([...]) checks many packages in parallel, and check_package(..., files=[...]) works offline from filenames you already have.

To count packages that Pyodide ships itself, as --pyodide-lock does, pass pyodide_locks=[wheelgap.pyodide.load_pyodide_lock("pyodide-lock.json")] (a path or https URL, one lock per Python) to check_packages(), check_package() or check_lockfile(). They show as bundled on pyodide targets for the lock's Python, and the result's bundled_version is the version Pyodide ships. A different version from your pin or range still counts as bundled (ok stays True, as in the CLI), so check bundled_version_differs; to_dict() includes both as "bundled" and "bundled_version_differs". A lock that no pyodide target can use raises PyodideLockError, as on the command line.

Scanning the top PyPI packages

python -m wheelgap.scan --top 5000 --out coverage.json

This checks the latest release of the most-downloaded packages (from top-pypi-packages) against the eight common platforms (and Pyodide), for Python 3.13, 3.14, 3.14t, 3.15 and 3.15t by default, and writes one JSON file with per-platform totals. It's the data behind the coverage tracker on the roadmap.

Add --pyodide-lock PATH_OR_URL once per Pyodide release (each is for one Python) to count the packages Pyodide ships itself as bundled on that Python's pyodide target; the per-target summary counts them too, and pyodide_locks in the JSON lists the locks used. Pyodide targets with no lock get a note on stderr.

Comparing two lockfiles

wheelgap diff old/uv.lock uv.lock -p 3.13 -t linux-aarch64,musl-aarch64
wheelgap diff origin/main:uv.lock uv.lock -p 3.13 -t linux-aarch64   # the base branch's lockfile, no checkout needed

This checks both lockfiles and lists the packages that got worse on a target (a wheel that became a source build, say) and the new packages with gaps, so a dependency bump that drops a platform fails CI. It exits 1 if there are any, 0 if not, and 2 on bad input. A package that can no longer be checked counts as worse: unknown, or now installed from git, a path, a URL or a private index (a new one of those counts as a new gap). Packages are matched by name, so a requirements-file path with no name (./pkg, or -e ./pkg without #egg=) shows up as a new package rather than as the one it replaced. Either input can be a git REF:PATH (origin/main:uv.lock, HEAD~1:poetry.lock) when no file or folder of that name exists: wheelgap reads it with git into a temporary folder under the same file name (with the pyproject.toml beside it at that ref, if any), so its format is detected and read as usual. PATH is relative to the repository root: use REF:./uv.lock from a subfolder. In a shallow CI checkout, fetch the base ref first. A ref or path git can't find exits 2 with git's message, and so does a missing git. The same -p and -t apply to both lockfiles, and [tool.wheelgap] and .python-version aren't read yet, so pass them explicitly. --json prints one object with worse (each package's old_versions and new_versions lists and, per target, old_status and new_status), new (each new package's versions and per-target status), unchecked and summary counts; a skipped cell has the status skipped and a skip_reason (git, url, private index, ...). diff is a subcommand only as the first argument when there's no file or folder named diff in the current directory (then it's read as an input, with a note; python -m wheelgap.diff always compares, and a pinned diff==2023.12.6 checks the PyPI package of that name). python -m wheelgap.diff works the same way.

With an AI coding agent

wheelgap ships a check-wheels skill that teaches coding agents to run it, read its JSON and suggest fixes. It follows the open Agent Skills format.

Claude Code (install from this repository as a plugin):

/plugin marketplace add wheelgap/wheelgap
/plugin install wheelgap@wheelgap

On Claude Code 2.1.275 or later, /plugin install wheelgap --marketplace wheelgap/wheelgap does both in one step.

Codex, Cursor, GitHub Copilot and others: copy integrations/claude-code/skills/check-wheels/ into your project's .agents/skills/ folder. For a Cursor project rule instead, copy integrations/cursor/wheelgap.mdc to .cursor/rules/.

As an MCP server: wheelgap-mcp lets any agent that speaks the Model Context Protocol call wheelgap as a tool, with no shell access needed. It runs over stdio and has five read-only tools: check_package (one package, latest or a given version), check_lockfile, check_dockerfile, diff_lockfiles and list_targets. check_lockfile gives the same answer as wheelgap PATH --json, [tool.wheelgap] settings included: it returns the packages with gaps, the ones it skipped (git, path and private-index packages, which aren't checked) and any notes. check_dockerfile gives the answer of wheelgap PATH --dockerfile DOCKERFILE --json (PATH defaults to the Dockerfile's folder), plus the image it read (python, platform, glibc, musl), so an agent can ask before switching python:3.13-slim to -alpine or adding --platform=linux/arm64; a Dockerfile wheelgap can't read is a tool error. diff_lockfiles takes an old and a new lockfile (plus python and targets) and answers as wheelgap diff OLD NEW --json does, with ok added: the packages that got worse on a target and the new ones with gaps, so an agent that bumps dependencies can save the old lockfile under its own name (mkdir -p /tmp/old && git show HEAD:uv.lock > /tmp/old/uv.lock) and ask whether the upgrade lost a wheel. On pyodide targets, check_package and check_lockfile take an optional pyodide_lock (a path or https URL of a pyodide-lock.json, as --pyodide-lock does), so packages Pyodide ships itself show as bundled. Tool calls send progress notifications when the client asks for them and don't stop the server answering ping. A cancelled call sends no more progress and no result, but its index lookups still finish before the next call starts. Like wheelgap, it needs nothing beyond packaging. For Claude Code:

claude mcp add wheelgap -- uvx --from wheelgap wheelgap-mcp

Other clients take the same command (uvx) and arguments in their MCP settings; the MCP setup guide has a copy-paste snippet for Claude Code, Cursor, VS Code, Codex and Claude Desktop. Pass check_lockfile, check_dockerfile and diff_lockfiles an absolute path or a file:// URI, since the server's working directory may not be your project. If the client shares its roots, a relative path such as uv.lock is resolved against the first root that contains it.

Contributing to wheelgap with an agent? See AGENTS.md.

How it works

• Version ranges such as numpy>=1.26,<2 use the newest matching release with a non-yanked file whose Requires-Python permits the selected Python. This applies when every target uses the same Python major/minor version (including a mix of regular and free-threaded builds). Missing or invalid Requires-Python metadata is treated as unrestricted. Runs checking several Python versions still choose the newest release matching the range without filtering by Python; per-Python resolution remains planned in #84. Exact pins and bare package names keep their existing behavior.
• Lockfiles are read offline. uv.lock, pylock.toml, poetry.lock and pdm.lock already record every file published for each pinned version (except a uv.lock or pylock.toml whose requires-python leaves out the Python you check; see below). wheelgap parses the wheel filenames' tags with packaging and checks them against the exact tag set each target accepts. That is the same logic pip and uv use to pick a wheel.
• requirements.txt, requirements.lock and bare package names are looked up on the PyPI JSON API. Feed it a fully pinned file (from Rye, uv pip compile, pip-compile or uv export): wheelgap checks the lines you give it and doesn't resolve dependencies. Rye's default (non-universal) locks are resolved for the machine that created them and carry no platform markers. Constraint files (-c FILE, -cFILE, --constraint FILE or --constraint=FILE) restrict unpinned requirements without adding packages. Exact constraints select a version; ranges are intersected. A constraint wheelgap can't apply leaves the package unknown rather than guessed: one with an environment marker (it can't yet be applied per target), a URL or path, or two different exact versions. Constraint files given as a URL or an environment variable (-c https://…, -c ${PIP_CONSTRAINT}) are skipped.
• Pipfile.lock reads default, develop and any custom package categories (such as docs or tests) and looks up distribution filenames on PyPI for their pinned versions. Environment markers are preserved; Git, path, file and editable dependencies are skipped as non-index packages. Packages pinned to an index that isn't PyPI are skipped too (private index), so their names aren't sent to pypi.org. A PyPI mirror counts as a private index here, since wheelgap can't tell it apart from an internal one, unless WHEELGAP_INDEX_URL points at the JSON API of that same index (for a source https://mirror.example.com/simple, set WHEELGAP_INDEX_URL=https://mirror.example.com/pypi): then those packages are looked up there, on the index the lock already names. It must be the same repository, not just the same host: <repo>/simple matches <repo>/pypi with the same scheme, host and port, so another repository on the same Artifactory or Nexus server stays skipped. The index has to serve PyPI's JSON API, including pinned versions (<WHEELGAP_INDEX_URL>/<name>/<version>/json); on a mirror that only serves the simple API, such as the Tsinghua or Aliyun mirrors, those packages come back as lookup errors like not on index instead. --index-url lookups don't use this.
• Environment markers are respected. A pywin32 ; sys_platform == 'win32' dependency won't show up as a Linux gap. For uv.lock the dependency graph is walked per target, so transitive platform-specific packages are handled too. On win32, linux-i686, musl-i686, linux-armv7l and musl-armv7l a platform_machine comparison counts as unknown, so the package is still checked. platform_machine reports the machine, not the Python: 32-bit Python on 64-bit Windows says AMD64, a linux/386 container on an x86_64 host says x86_64, and 32-bit Raspberry Pi OS on a Pi 4 or 5 (which boots a 64-bit kernel) or a linux/arm/v7 container on an arm64 host can say aarch64.
• --online re-checks lockfile packages against PyPI. Packages the lock says came from another index keep their locked files, so their names are never sent to PyPI. Use it when a project has uploaded new wheels for an existing version since you locked, or if your uv.lock was generated with environments restrictions. A uv.lock or pylock.toml whose requires-python leaves out a target's Python (say ==3.13.*, checked with -p 3.14t) may be missing wheels for it, because uv only keeps the wheels that can serve requires-python. Without --online, wheelgap then also looks up its PyPI packages on the index, keeps the lock's own files if a lookup fails, and prints a note (with -q, only when some packages couldn't be re-checked).
• --offline never contacts an index, for CI and coding agents without network. Inputs that need a lookup (requirements files, pyproject.toml, Pipfile.lock, package names, lockfile entries with no file list) make it exit 2 with an error naming the packages, so use a lockfile instead. A --pyodide-lock URL is refused too: pass a downloaded copy. The requires-python re-check above is skipped, with a note that those Pythons may show false gaps. It can't be combined with --online. Set offline = true in [tool.wheelgap] to make it the project default; --online on the command line overrides it for one run.
• Cache: lookups of a pinned version (name==1.2.3) are saved for 24 hours, so a repeat run needn't ask PyPI again. They're kept in wheelgap/pinned-v1 inside ~/.cache on Linux (or $XDG_CACHE_HOME), ~/Library/Caches on macOS and %LOCALAPPDATA% on Windows. Each entry holds only the release's file names and yanked flags, named by a hash of the index URL (without any credentials in it), the package and the version; on Linux and macOS, wheelgap creates the folder and its entries readable by you only. --online always asks the index again and saves the fresh answer, so later runs use it; --no-cache neither reads nor saves the cache for one run. Entries older than 24 hours are deleted, at most once a day, and wheelgap --clear-cache deletes all of them (add inputs to check them right after). Both touch only wheelgap's own entry files, and never follow a link. Unpinned names, version ranges and --index-url lookups aren't cached, and neither are lookups from the Python API.
• Private indexes: --index-url URL (or index-url in [tool.wheelgap]) sends the PyPI lookups described above to a simple index that serves the JSON form (PEP 691) instead. The URL must be https (plain http only for localhost). Credentials can go in the URL passed to --index-url (https://user:token@host/simple) and never appear in output; in CI, --index-url "$PRIVATE_INDEX" keeps them out of your repo. They're refused in [tool.wheelgap], since pyproject.toml is committed. An index that only serves HTML (PEP 503) gets a clear error.

Why this exists

Missing wheels show up late, in the most expensive place:

• Your python:3.13-alpine image on an ARM runner takes 25 minutes because two transitive dependencies don't publish musllinux_aarch64 wheels, and QEMU is compiling them.
• A build fails with error: Microsoft Visual C++ 14.0 or greater is required on a teammate's Snapdragon laptop.
• You try free-threaded Python 3.14t and half your stack quietly falls back to source builds, because abi3 wheels don't apply there.

Sites like pyreadiness.org and free-threaded wheels track the top few hundred packages. wheelgap checks your packages, at your pinned versions, for your targets, in about a second, and it can run in CI.

Roadmap

wheelgap is growing into caniuse for Python wheels: an open map of which packages have wheels for which platforms. The CLI stays focused, and around it come a nightly tracker of the most-used PyPI packages, a community knowledge base of workarounds, and campaigns to help upstream projects ship missing wheels. Read the roadmap and join the discussion →

Next up for the CLI, each with a first step written up and no PR yet:

• "nearest version with a wheel" hints (#20)
• version ranges: picking the newest release that supports each target's Python (#84)

The roadmap and help wanted labels have more.

Contributing

Contributions are welcome, and pull requests usually get a review within hours.

• New here? Issues labelled good first issue say which files to change and what to test.
• Done one or two? help wanted issues are bigger, and a good next step.

CONTRIBUTING.md has the 3-command dev setup, or you can start in the browser with no setup:

Questions are welcome in Discussions.

Thanks

wheelgap is created and maintained by @patrickballard, with a lot of help from these people. Thank you, all of you!

License

MIT