withcache

A tiny, operator-curated artifact cache for a small lab, for the big vendor downloads you re-pull constantly (CUDA, ROCm, DOCA, firmware, drivers), fronted by transparent curl/wget shims so existing scripts use it with no changes.

Think of it as "ccache for HTTP artifacts, without a proxy."

curl -fsSL https://the/origin/cuda.tar.gz -o cuda.tar.gz     # your script, unchanged
   └─ curlwithcache shim ─ WITHCACHE_SERVER set?
        ├─ cached  → served from the cache-host (fast, local)
        └─ miss/unset/unreachable → runs the real curl, exactly as written

Artifacts are cached by their origin URL as a key; the shim opts in by re-pointing the URL at the cache. No transparent proxy, no TLS interception, no client CA. The URL is a lookup key, not a connection target.

By default a miss is auto-fetched: the request falls through to origin (so the caller gets its file straight away), and the cache-host pulls the same artifact in the background, so the next request hits. Run with --curate to require a human instead, who reviews the miss list in a small web UI and presses Download (or pre-seeds via Add from URI). Either way the cache-host is the only box that needs internet egress (and any vendor credentials), and clients never write to it.

Why not just curl + a caching proxy?

For https:// (i.e. every vendor download) a forward proxy can't cache without SSL-bump / MITM: curl tunnels TLS end-to-end via CONNECT, so the proxy only sees ciphertext. The shim sidesteps that entirely by re-pointing the URL to the cache instead of intercepting the connection. And no proxy offers the optional operator-curated model (--curate: a miss queue a human approves).

Components

The cache-host and the Python shims are stdlib-only (no third-party runtime deps); the native shim is a dependency-free static binary.

Install

The cache-host and Python shims (works on any box with Python):

pipx install withcache    # or: uv tool install withcache  /  pip install withcache
# provides: curlwithcache  wgetwithcache  withcache-server

The native shim (no Python needed, for minimal/distroless boxes; ~200 KB static musl binary). Grab it from the Releases page; one binary serves both tools by the name it's invoked as:

curl -L .../releases/.../withcache-shim-x86_64-linux-musl -o /usr/local/bin/curlwithcache
chmod +x /usr/local/bin/curlwithcache

The Python shim is also the tested oracle and install-time fallback for platforms without a prebuilt binary; a differential test asserts the binary and the Python plan() rewrite argv identically.

Deploy the cache-host

export WITHCACHE_ADMIN_PASSWORD=change-me    # protects the operator UI
podman compose -f deploy/compose.yml up -d   # or: docker compose -f ...
# operator UI:  http://withcache-server:8081/

Or without containers:

WITHCACHE_ADMIN_PASSWORD=change-me withcache-server --data-dir ./data --port 8081

Data (blobs + cache.db + session-secret) lives in the /data volume (or --data-dir). Artifacts are immutable per version, so there's no cache invalidation. --workers N sets the number of concurrent download workers, --curate switches from auto-fetch to operator-approved pulls, and --max-bytes (e.g. 50G) caps the cache: when full it refuses new fills (no auto-eviction), and you free space by deleting artifacts in the UI.

Use the shims (transparent curl / wget)

Every approach is the same two ingredients: (1) point at the cache with WITHCACHE_SERVER, and (2) make curl/wget resolve to the shim. They differ only in how widely the system curl/wget is shadowed. Pick the least invasive one that fits.

Safety: with WITHCACHE_SERVER unset the shim is a pure pass-through (it just execs the real tool, zero network/parsing), so even the system-wide setup is harmless wherever the cache isn't configured. Worst case is always "no caching, curl still works."

These all use command -v curlwithcache, so they work whether you installed the native binary or the Python launcher (both land under that name).

1. No shadowing: call the shims by name (least invasive)

Nothing is renamed; you opt in per command. Good for trying it out or a script you can edit.

export WITHCACHE_SERVER=http://withcache-server:8081
curlwithcache -fsSL https://the/origin/cuda.tar.gz -o cuda.tar.gz
wgetwithcache https://the/origin/rocm.tar.gz

2. This shell only: shadow curl/wget for the session

Put curl/wget symlinks in a dir and prepend it to PATH in the current shell. Reversible by just closing the shell.

mkdir -p ~/.withcache/bin
ln -sf "$(command -v curlwithcache)" ~/.withcache/bin/curl
ln -sf "$(command -v wgetwithcache)" ~/.withcache/bin/wget

export WITHCACHE_SERVER=http://withcache-server:8081
export PATH="$HOME/.withcache/bin:$PATH"
hash -r                       # forget any cached curl/wget location

command -v curl               # -> ~/.withcache/bin/curl  (verify it's the shim)
curl -fsSL https://the/origin/cuda.tar.gz -o cuda.tar.gz   # existing scripts, unchanged
wget https://the/origin/rocm.tar.gz                        # still saved as rocm.tar.gz

3. Your user: make it the default for your shells (persistent)

Create the symlinks once, then add the two exports to your shell rc. Affects all your future interactive shells; undo by deleting the block.

mkdir -p ~/.withcache/bin
ln -sf "$(command -v curlwithcache)" ~/.withcache/bin/curl
ln -sf "$(command -v wgetwithcache)" ~/.withcache/bin/wget

cat >> ~/.bashrc <<'EOF'

# withcache: transparent curl/wget caching
export WITHCACHE_SERVER=http://withcache-server:8081
export PATH="$HOME/.withcache/bin:$PATH"
EOF

4. One project only: scope it with direnv

Drop an .envrc in a project tree (requires direnv); caching applies only inside that directory.

# .envrc
export WITHCACHE_SERVER=http://withcache-server:8081
PATH_add ~/.withcache/bin        # assumes the symlinks from approach 2/3 exist

Then direnv allow.

5. The whole machine: every user, every shell (most invasive)

Install the shim as curl/wget in /usr/local/bin (ahead of /usr/bin on the default PATH) and set the server globally. This also catches build tools and package managers that shell out to curl/wget.

sudo ln -sf "$(command -v curlwithcache)" /usr/local/bin/curl
sudo ln -sf "$(command -v wgetwithcache)" /usr/local/bin/wget

# A login-shell env file (covers interactive logins; daemons started outside a
# login shell won't see it; set WITHCACHE_SERVER in their unit if you need it).
echo 'export WITHCACHE_SERVER=http://withcache-server:8081' \
  | sudo tee /etc/profile.d/withcache.sh >/dev/null

On minimal/distroless hosts use the native shim binary here: same symlink, no Python required.

Verify / turn it off

command -v curl                       # which curl is in effect (the shim, or the real one)
export REAL_CURL=/usr/bin/curl        # optional: pin the wrapped tool (also $REAL_WGET)

unset WITHCACHE_SERVER                 # instantly back to plain curl (pass-through)
rm ~/.withcache/bin/curl ~/.withcache/bin/wget   # remove shadowing entirely

How it works: the shim scans for the URL, asks the cache, and execs the real tool:

1. Find the real curl/wget on $PATH (skipping itself; $REAL_CURL/$REAL_WGET override).
2. With WITHCACHE_SERVER set, find the URL (the scheme:// arg, or --url).
3. Probe the cache with that same tool (curl -I / wget --spider).
   • Hit → re-point only the URL at http://server/b/<base64(origin)>/<basename> and exec the real tool (so -o, -O, -L, --retry, … all still apply, and the file is named after the artifact).
   • Miss / unreachable → exec the real tool with your arguments untouched (origin); the miss is recorded for the operator.
4. With no WITHCACHE_SERVER, it does zero network/parsing, just execs the real tool.

Notes & limits (all degrade gracefully; worst case is "no caching, curl still works"):

• Needs the wrapped tool present (it shims it). Adds ~Python-startup latency per call.
• URLs hidden in a -K/-i config file or piped via stdin aren't seen → those calls pass through uncached.
• Per-tool env override: CURLWITHCACHE_SERVER / WGETWITHCACHE_SERVER beat WITHCACHE_SERVER.

Operator UI

http://withcache-server:8081/ (Bootstrap 5 + Bootstrap Icons + HTMX, bundled offline; matches bty's chrome for a consistent trio) is a five-page dashboard:

• Cached (landing): URL, size, hits (times served) and misses (times requested before it was cached), SHA-256, fetched-at, each with Delete to free space.
• Streams: in-flight stream-through-and-store fetches serving bytes to a client while writing to disk.
• Downloads: live progress bars, queued/running/completed/cancelled/failed, Cancel, and Clear finished. Downloads run in a background worker pool, not in the request, so large pulls never block, modelled on bty's job managers. The subnav Fetch form pre-seeds an artifact before anyone misses it.
• Misses: auto-fetched by default, or (under --curate) each with Download (queues a background pull) and Dismiss.
• Catalog: image catalog fetched from a nosi-style catalog.toml (URL configurable via $WITHCACHE_CATALOG_URL or the subnav Set&fetch input); pre-seed by URL via the "Add image from oras" input.

Auth

Single-tenant session-cookie auth (modelled on bty's approach, env password instead of PAM). The read path (/blob, /b/…, /healthz) is open so shims never log in; the operator surface (/, /admin/*) is gated.

Cache keys & signed URLs

The key is scheme://host/path with the query string dropped by default, so CDN/presigned URLs (whose tokens change every request) still match by path. Pass --keep-query to the server for query-sensitive keys. Package-manager repos (.deb/.rpm) are GPG-signed and verified by the client regardless of transport, so caching them this way is safe.

Consume from another tool (the client library)

A tool that already knows its download URLs (e.g. an installer or a provisioner) can prefer the cache without shelling out to a shim or re-implementing the /b/ scheme. withcache.client is stdlib-only, so importing it adds no dependencies:

from withcache import client

# "use the cache when it's warm, the origin otherwise"
url = client.serve_url("http://cache:8081", origin) or origin

is_cached() is a graceful HEAD (a miss, timeout, or unreachable cache all return False, so you fall back to the origin), and it doubles as a warm-up: the probe records the miss and, in auto-fetch mode, enqueues the fill, so the next call flips to the cache. The encoding is shared with the shims and server, so consumers stay in lockstep with the cache-host.

Pull an oras:// artifact (oras + client together)

For a registry blob, pair withcache.oras (resolve the reference to a blob URL

• bearer) with client.serve_url (prefer the cache when warm). A runnable example is in examples/pull_oras_blob.py:

from withcache import client, oras

resolved = oras.resolve_ref("oras://ghcr.io/<owner>/<repo>@sha256:<digest>")
url, headers = resolved.blob_url, dict(resolved.headers)
url = client.serve_url(server, url, headers=headers) or url   # cache when warm

Tests

python -m unittest discover -s tests   # stdlib only, no test deps