diff --git a/docs/proposals/index.rst b/docs/proposals/index.rst
index 31408767..f5712deb 100644
--- a/docs/proposals/index.rst
+++ b/docs/proposals/index.rst
@@ -4,5 +4,6 @@ Fromager Enhancement Proposals
 .. toctree::
    :maxdepth: 1
 
+   multiple-cache-wheel-servers
    new-patcher-config
    new-resolver-config
diff --git a/docs/proposals/multiple-cache-wheel-servers.md b/docs/proposals/multiple-cache-wheel-servers.md
new file mode 100644
index 00000000..34931a2e
--- /dev/null
+++ b/docs/proposals/multiple-cache-wheel-servers.md
@@ -0,0 +1,131 @@
+# Multiple cache wheel servers
+
+- Author: Christian Heimes
+- Created: 2026-04-16
+- Status: Open
+
+## What
+
+Support multiple cache wheel server URLs instead of a single
+`--cache-wheel-server-url`. Caches are checked in order, similar to pip's
+`--extra-index-url` or uv's `unsafe-best-match` index strategy.
+
+## Why
+
+In downstream, each variant and CPU architecture has its own build cache in
+a GitLab PyPI registry. During bootstrap and build, fromager uses a
+read/write cache in GitLab to avoid rebuilding wheels that were already built
+for the current architecture. After a successful bootstrap and build, wheels
+for all architectures are copied into a variant-specific cache in Pulp.
+
+Currently fromager only supports a single `--cache-wheel-server-url`. To
+pull from both the arch-specific GitLab cache **and** the variant-wide Pulp
+cache, we need ordered multi-cache support.
+
+All cache indexes are fully trusted and controlled by us, so security
+concerns around index mixing (as with pip's `--extra-index-url`) do not
+apply.
+
+## Goals
+
+- Accept multiple cache wheel server URLs on the CLI and in
+  `WorkContext`.
+- Check caches in order: local wheel server first, then each cache URL in
+  the order specified.
+- Remain backwards-compatible: a single `--cache-wheel-server-url` keeps
+  working.
+- Integrate with `build_environment.py` so build-dependency installation
+  also searches multiple caches.
+
+## Current state
+
+`bootstrap`, `build-sequence`, and `build-parallel` each accept a single
+`-c` / `--cache-wheel-server-url` option. The cache URL is not stored on
+`WorkContext`; it is threaded through function arguments.
+
+`wheels.get_wheel_server_urls()` builds an ordered list for resolution:
+
+1. Per-package variant `wheel_server_url` (if set, used exclusively)
+2. `ctx.wheel_server_url` (local wheel server)
+3. `cache_wheel_server_url` (single cache, appended last)
+
+Build-dependency installation (`build_environment.BuildEnvironment`) only
+uses the local wheel server (`--index-url`). The cache is not passed to
+pip/uv when installing build dependencies.
+
+`Bootstrapper._download_wheel_from_cache()` checks the single cache URL
+and downloads the first matching wheel.
+
+## Proposed changes
+
+### 1. CLI: accept multiple `--cache-wheel-server-url` values
+
+Make the existing `-c` / `--cache-wheel-server-url` click option
+repeatable (`multiple=True`). The internal parameter name changes from
+`cache_wheel_server_url` (singular) to `cache_wheel_server_urls`
+(sequence). A single `-c URL` keeps working.
+
+### 2. Make `cache_wheel_server_urls` arguments a sequence
+
+All internal call sites that currently pass a single
+`cache_wheel_server_url: str | None` change to accept a sequence.
+`wheels.get_wheel_server_urls()` appends all cache URLs in order instead
+of a single one.
+
+Affected modules: `commands/bootstrap.py`, `commands/build.py`,
+`bootstrapper.py`, `bootstrap_requirement_resolver.py`, `wheels.py`.
+
+### 3. Bootstrapper: iterate over caches
+
+`Bootstrapper._download_wheel_from_cache()` iterates over all cache URLs
+instead of checking a single one. The first cache that has a matching
+wheel wins (first-match semantics, same as `--extra-index-url`).
+
+### 4. Build environment: pass caches to pip/uv
+
+Store `cache_wheel_server_urls` on `WorkContext` and extend
+`pip_wheel_server_args` to emit `--extra-index-url` for each cache URL.
+`BuildEnvironment.install()` then picks up caches automatically.
+
+Set `UV_INDEX_STRATEGY=unsafe-best-match` in the build environment so uv
+considers all indexes equally and picks the best matching version across
+all of them, rather than stopping at the first index that has any version
+of a package.
+
+`WorkContext` already carries the local wheel server URL, so the cache
+URLs belong there too -- they share the same lifetime.
+
+### 5. Resolution order
+
+The final resolution order for wheel lookups:
+
+1. **Per-package variant `wheel_server_url`** -- if set, used exclusively
+   (no change).
+2. **Local wheel server** (`ctx.wheel_server_url`) -- just-built wheels.
+3. **Cache 1** (first `-c` URL) -- e.g., variant-wide Pulp cache
+   (read-only).
+4. **Cache 2** (second `-c` URL) -- e.g., arch-specific GitLab PyPI
+   registry (read/write).
+5. ... additional caches in CLI order.
+
+This matches the use case: check local builds first, then the Pulp cache
+with already-published wheels, then the arch-specific GitLab cache.
+
+## Example usage
+
+```bash
+fromager \
+    --variant cuda \
+    bootstrap \
+    -c https://pulp.example/content/rhel-ai-cuda/simple \
+    -c https://gitlab.example/api/v4/projects/123/packages/pypi/simple \
+    -r requirements.txt
+```
+
+## Migration
+
+- The CLI flag name stays the same (`--cache-wheel-server-url` / `-c`).
+- Existing invocations with zero or one `-c` flag work without changes.
+- The singular parameter name `cache_wheel_server_url` is renamed to
+  `cache_wheel_server_urls` in internal APIs. This is a breaking change
+  for any downstream code calling these functions directly.
diff --git a/docs/spelling_wordlist.txt b/docs/spelling_wordlist.txt
index dbec132b..1a5a5f54 100644
--- a/docs/spelling_wordlist.txt
+++ b/docs/spelling_wordlist.txt
@@ -2,6 +2,7 @@ args
 backend
 backends
 backoff
+bootstrapper
 canonicalize
 canonicalized
 changelog
@@ -34,6 +35,7 @@ lexicographically
 libcurl
 linter
 localhost
+lookups
 matcher
 mypy
 namespace
@@ -73,6 +75,7 @@ txt
 unshare
 url
 urls
+uv
 vendored
 vendoring
 versionless