diff --git a/internal/config/loader.go b/internal/config/loader.go
index 9b2d2397d..3ceda2a4a 100644
--- a/internal/config/loader.go
+++ b/internal/config/loader.go
@@ -13,6 +13,7 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/smart-mcp-proxy/mcpproxy-go/internal/secret"
@@ -432,6 +433,19 @@ func atomicWriteFile(path string, data []byte, perm os.FileMode) error {
 	return atomicWriteFileMode(path, data, perm, false)
 }
 
+// skipWriteSync drops the fsync from atomicWriteFileMode. Only tests set it,
+// through DisableWriteSyncForTesting; production writes always sync.
+var skipWriteSync atomic.Bool
+
+// DisableWriteSyncForTesting makes config file writes skip the fsync before
+// the atomic rename for the rest of the process. The written bytes are still
+// visible to every reader in the process; only crash durability is lost, which
+// no unit test exercises. Call it from TestMain in packages whose tests write
+// configs heavily: FlushFileBuffers on the Windows CI runner costs tens of
+// milliseconds per write and dominated internal/runtime's test time. Never
+// call it from production code.
+func DisableWriteSyncForTesting() { skipWriteSync.Store(true) }
+
 // atomicWriteFileMode is atomicWriteFile; with exactMode the temp file is
 // chmod-ed to perm before the rename, so the result carries perm itself and
 // not perm filtered by the process umask (which OpenFile applies).
@@ -473,8 +487,10 @@ func atomicWriteFileMode(path string, data []byte, perm os.FileMode, exactMode b
 	}
 
 	// Fsync to ensure data is on disk
-	if err := tmpFile.Sync(); err != nil {
-		return fmt.Errorf("failed to sync temp file: %w", err)
+	if !skipWriteSync.Load() {
+		if err := tmpFile.Sync(); err != nil {
+			return fmt.Errorf("failed to sync temp file: %w", err)
+		}
 	}
 
 	// Close temp file
diff --git a/internal/runtime/main_test.go b/internal/runtime/main_test.go
new file mode 100644
index 000000000..95580a4bd
--- /dev/null
+++ b/internal/runtime/main_test.go
@@ -0,0 +1,29 @@
+package runtime
+
+import (
+	"os"
+	"testing"
+
+	"go.etcd.io/bbolt"
+
+	"github.com/smart-mcp-proxy/mcpproxy-go/internal/config"
+	"github.com/smart-mcp-proxy/mcpproxy-go/internal/storage"
+)
+
+// TestMain turns off fsync for config writes and bbolt commits. Nearly every
+// test here builds a Runtime, which writes the config file, commits to
+// config.db and creates a Bleve index; on the windows-latest runner those
+// flushes cost tens of milliseconds each and the package took ~470s against
+// go test's default 10m timeout (TestConfigCommit_CommitPathsConverge alone
+// ran 92s). Skipping the flush changes nothing a test can observe in-process
+// (writes stay visible to every reader); only crash durability is lost, and no
+// test here simulates a crash.
+//
+// bbolt.DefaultOptions is what Bleve's scorch index copies for its root.bolt;
+// storage opens config.db with its own options, hence the separate switch.
+func TestMain(m *testing.M) {
+	config.DisableWriteSyncForTesting()
+	storage.DisableSyncForTesting()
+	bbolt.DefaultOptions.NoSync = true
+	os.Exit(m.Run())
+}
diff --git a/internal/storage/bbolt.go b/internal/storage/bbolt.go
index 1842fdbc8..44786d622 100644
--- a/internal/storage/bbolt.go
+++ b/internal/storage/bbolt.go
@@ -9,6 +9,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"sync/atomic"
 	"time"
 
 	"go.etcd.io/bbolt"
@@ -36,6 +37,19 @@ type BoltDB struct {
 	logger *zap.SugaredLogger
 }
 
+// noSync opens databases with bbolt's NoSync. Only tests set it, through
+// DisableSyncForTesting; production databases always sync their commits.
+var noSync atomic.Bool
+
+// DisableSyncForTesting makes every database opened afterwards in this
+// process skip the fsync on commit (bbolt Options.NoSync). Committed data is
+// still visible to every reader of the open database; only crash durability is
+// lost, which no unit test exercises. Call it from TestMain in packages whose
+// tests open many databases or commit heavily: an fsync per commit on the
+// Windows CI runner dominated internal/runtime's test time. Never call it from
+// production code.
+func DisableSyncForTesting() { noSync.Store(true) }
+
 // NewBoltDB creates a new BoltDB instance
 func NewBoltDB(dataDir string, logger *zap.SugaredLogger) (*BoltDB, error) {
 	dbPath := filepath.Join(dataDir, "config.db")
@@ -45,6 +59,7 @@ func NewBoltDB(dataDir string, logger *zap.SugaredLogger) (*BoltDB, error) {
 	// owner-only (0600).
 	db, err := bbolt.Open(dbPath, 0600, &bbolt.Options{
 		Timeout: 10 * time.Second,
+		NoSync:  noSync.Load(),
 	})
 	if err != nil {
 		logger.Warnf("Failed to open database on first attempt: %v", err)
diff --git a/internal/storage/main_test.go b/internal/storage/main_test.go
new file mode 100644
index 000000000..61dc03395
--- /dev/null
+++ b/internal/storage/main_test.go
@@ -0,0 +1,15 @@
+package storage
+
+import (
+	"os"
+	"testing"
+)
+
+// TestMain opens every test database with bbolt NoSync: an fsync per commit
+// made this package take ~150s on the windows-latest runner. Committed data
+// stays visible to every reader of the open database; only crash durability is
+// lost, and no test here simulates a crash. See DisableSyncForTesting.
+func TestMain(m *testing.M) {
+	DisableSyncForTesting()
+	os.Exit(m.Run())
+}