diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
new file mode 100644
index 0000000..87afbc1
--- /dev/null
+++ b/.github/workflows/benchmark.yml
@@ -0,0 +1,29 @@
+name: Benchmark this PR
+
+on:
+  pull_request:
+    branches: [main]
+    paths-ignore: ['docs/**']
+
+permissions:
+  pull-requests: write
+  contents: read
+
+jobs:
+  benchmark:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    strategy:
+      fail-fast: false
+      matrix:
+        version: ["1", "lts"]
+    env:
+      # Force consistent Julia depot path for self-hosted runners
+      JULIA_DEPOT_PATH: ~/.julia
+    steps:
+      - uses: MilesCranmer/AirspeedVelocity.jl@action-v1
+        with:
+          julia-version: ${{ matrix.version }}
+          script: "benchmark/benchmarks.jl"
+          extra-pkgs: ""
+          job-summary: true
diff --git a/benchmark/benchmarks.jl b/benchmark/benchmarks.jl
new file mode 100644
index 0000000..d37efde
--- /dev/null
+++ b/benchmark/benchmarks.jl
@@ -0,0 +1,53 @@
+using FunctionWrappersWrappers, BenchmarkTools
+
+const SUITE = BenchmarkGroup()
+
+# FunctionWrappersWrapper wraps one function across multiple argument signatures
+argtypes = (Tuple{Float64, Float64}, Tuple{Int, Int})
+rettypes = (Float64, Int)
+
+# =============================================================================
+# Construction
+# =============================================================================
+
+SUITE["construct"] = BenchmarkGroup()
+
+SUITE["construct"]["default"] = @benchmarkable FunctionWrappersWrapper(
+    +, $argtypes, $rettypes
+)
+SUITE["construct"]["dictcache"] = @benchmarkable FunctionWrappersWrapper(
+    +, $argtypes, $rettypes; cache = DictCache()
+)
+SUITE["construct"]["no_policy"] = @benchmarkable FunctionWrappersWrapper(
+    +, $argtypes, $rettypes; policy = AllowAll()
+)
+
+# =============================================================================
+# Calls and metadata
+# =============================================================================
+
+SUITE["call"] = BenchmarkGroup()
+
+fww = FunctionWrappersWrapper(+, argtypes, rettypes)
+fww_dict = FunctionWrappersWrapper(+, argtypes, rettypes; cache = DictCache())
+
+SUITE["call"]["float_args"] = @benchmarkable $fww(1.0, 2.0)
+SUITE["call"]["int_args"] = @benchmarkable $fww(1, 2)
+SUITE["call"]["dictcache"] = @benchmarkable $fww_dict(1.0, 2.0)
+
+SUITE["call"]["unwrap"] = @benchmarkable unwrap($fww)
+SUITE["call"]["wrapped_signatures"] = @benchmarkable wrapped_signatures($fww)
+SUITE["call"]["wrapped_return_types"] = @benchmarkable wrapped_return_types($fww)
+
+# =============================================================================
+# Vector of wrappers
+# =============================================================================
+
+SUITE["vector"] = BenchmarkGroup()
+
+fww1 = FunctionWrappersWrapper(sin, (Tuple{Float64},), (Float64,))
+FWW1 = typeof(fww1)
+fww_vec = FWW1[sin, tan, log]
+
+SUITE["vector"]["getindex_call"] = @benchmarkable $fww_vec[1](0.5)
+SUITE["vector"]["construct"] = @benchmarkable $FWW1[sin, tan, log]