diff --git a/.gitignore b/.gitignore
index 76cec07..dcecdc1 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,3 +1,4 @@
+**/.gdbinit
 !*/
 difftest/
 !/nemu/*
diff --git a/flake.nix b/flake.nix
index d5c35f1..74de402 100644
--- a/flake.nix
+++ b/flake.nix
@@ -93,7 +93,7 @@
         devShells.npc = pkgs.mkShell.override { stdenv = pkgs.ccacheStdenv; } {
           inherit (self.checks.${system}.pre-commit-check) shellHook;
           CHISEL_FIRTOOL_PATH = "${nixpkgs-circt162.legacyPackages.${system}.circt}/bin";
-          NPC_IMAGES_DIR="${self.packages.${system}.am-kernels-npc}/share/am-kernels";
+          NPC_IMAGES_PATH = "${self.packages.${system}.rv32Cross.am-kernels-npc}/share/am-kernels";
           packages = with pkgs; [
             clang-tools
             cmake
@@ -109,6 +109,7 @@
 
           nativeBuildInputs = with pkgs; [
             cmake
+            ninja
             sbt
             nvboard
             nixpkgs-circt162.legacyPackages.${system}.circt
@@ -117,10 +118,11 @@
             flex
             bison
             verilator
-            self.packages.${system}.am-kernels-npc
+            self.packages.${system}.rv32Cross.am-kernels-npc
           ];
 
           buildInputs = with pkgs; [
+            spdlog
             nvboard
             openssl
             libllvm
@@ -128,10 +130,6 @@
             readline
             mini-gdbstub
           ] ++ self.checks.${system}.pre-commit-check.enabledPackages;
-
-          cmakeFlags = [
-            "-DDIFFTEST_LIB:string=${self.packages.${system}.nemu-lib}/lib/riscv32-nemu-interpreter-so"
-          ];
         };
       }
     );
diff --git a/npc/CMakeLists.txt b/npc/CMakeLists.txt
index 4b4424e..2903475 100644
--- a/npc/CMakeLists.txt
+++ b/npc/CMakeLists.txt
@@ -41,6 +41,7 @@ if(BUILD_SIM_NVBOARD_TARGET)
   find_package(SDL2_image REQUIRED)
 endif()
 find_package(CLI11 CONFIG REQUIRED)
+find_package(spdlog REQUIRED)
 
 option(ENABLE_SDB "Enable simple debugger" OFF)
 
@@ -58,9 +59,6 @@ set(CHISEL_OUTPUT_VERILATOR_CONF ${CHISEL_OUTPUT_DIR}/conf.vlt)
 set(CHISEL_OUTPUT_TOPMODULE ${CHISEL_OUTPUT_DIR}/${TOPMODULE}.sv)
 set(CHISEL_EMIT_ARGS "--target-dir ${CHISEL_OUTPUT_DIR}")
 
-# -- Find difftest binaries
-file(GLOB_RECURSE DIFFTEST_BINARY_FILES "${DIFFTEST_RESOURCE_DIR}/*.bin")
-
 # -- Build NVBoard executable
 if(BUILD_SIM_NVBOARD_TARGET)
   add_subdirectory(csrc_nvboard)
diff --git a/npc/csrc/Flow/CMakeLists.txt b/npc/csrc/Flow/CMakeLists.txt
index d64cfca..2af6a5b 100644
--- a/npc/csrc/Flow/CMakeLists.txt
+++ b/npc/csrc/Flow/CMakeLists.txt
@@ -1,6 +1,6 @@
 include(ChiselBuild)
-add_executable(V${TOPMODULE} config.cpp main.cpp)
-target_link_libraries(V${TOPMODULE} PRIVATE devices gdbstub)
+add_executable(V${TOPMODULE} config.cpp gdbstub_wrapper.cpp main.cpp)
+target_link_libraries(V${TOPMODULE} PRIVATE devices gdbstub spdlog::spdlog)
 target_include_directories(V${TOPMODULE} PRIVATE ${CMAKE_SOURCE_DIR}/include)
 
 verilate(
@@ -28,8 +28,8 @@ foreach(DIFFTEST_BINARY_FILE IN LISTS DIFFTEST_BINARY_FILES)
   unset(FILENAME)
 endforeach()
 
-add_library(${TOPMODULE} SHARED config.cpp main.cpp)
-target_link_libraries(${TOPMODULE} PRIVATE devices gdbstub)
+add_library(${TOPMODULE} SHARED config.cpp gdbstub_wrapper.cpp)
+target_link_libraries(${TOPMODULE} PRIVATE devices gdbstub spdlog::spdlog)
 target_include_directories(${TOPMODULE} PRIVATE ${CMAKE_SOURCE_DIR}/include)
 set_property(TARGET PROPERTY POSITION_INDEPENDENT_CODE ON)
 target_link_options(${TOPMODULE} PRIVATE -Wl,-E)
diff --git a/npc/csrc/Flow/config.cpp b/npc/csrc/Flow/config.cpp
index 71e042f..035fe24 100644
--- a/npc/csrc/Flow/config.cpp
+++ b/npc/csrc/Flow/config.cpp
@@ -2,8 +2,10 @@
 
 void Config::cli_parse(int argc, char **argv) {
   CLI::App app;
+  app.add_option("-l,--listen", gdbsocket,
+                 "Listen to debugger at this address");
+  app.add_flag("-g", do_debug, "Listen for gdb");
   app.add_option("-m,--memory", memory_file, "Content of memory")
-      ->required()
       ->check(CLI::ExistingFile);
   app.add_flag("!--no-bin", memory_file_binary,
                "Memory file is in text format");
@@ -12,11 +14,6 @@ void Config::cli_parse(int argc, char **argv) {
   app.add_option("--diff-lib", lib_ref,
                  "Dynamic library file of difftest reference")
       ->check(CLI::ExistingFile);
-  app.add_flag("--mtrace", do_mtrace, "Enable memory tracing");
-  app.add_option(
-         "--mtrace-range", mtrace_ranges,
-         "Specify memory tracing range (default: 0x80000000-0x8fffffff)")
-      ->delimiter(',');
 
   try {
     app.parse(argc, argv);
diff --git a/npc/csrc/Flow/gdbstub_wrapper.cpp b/npc/csrc/Flow/gdbstub_wrapper.cpp
new file mode 100644
index 0000000..970ce55
--- /dev/null
+++ b/npc/csrc/Flow/gdbstub_wrapper.cpp
@@ -0,0 +1,196 @@
+extern "C" {
+#include <gdbstub.h>
+}
+#include "components.hpp"
+#include <VFlow.h>
+#include <config.hpp>
+#include <cstdint>
+#include <cstdlib>
+#include <devices.hpp>
+#include <types.h>
+#include <vl_wrapper.hpp>
+#include <vpi_user.h>
+#include <vpi_wrapper.hpp>
+
+using Registers = _RegistersVPI<uint32_t, 32>;
+using VlModule = VlModuleInterfaceCommon<VFlow, Registers>;
+
+// SDB::SDB<NPC::npc_interface> sdb_dut;
+bool g_skip_memcheck = false;
+VlModule *top;
+Registers *regs;
+vpiHandle pc = nullptr;
+
+const size_t PMEM_START = 0x80000000;
+const size_t PMEM_END = 0x87ffffff;
+
+extern "C" {
+/* === Memory Access === */
+using MMap = MemoryMap<Memory<128 * 1024>, Devices::DeviceMap>;
+void *pmem_get() {
+  static Devices::DeviceMap devices{
+      new Devices::Serial(0x10000000, 0x1000),
+      new Devices::RTC(0x10001000, 0x1000),
+  };
+  static auto pmem = new MemoryMap<Memory<128 * 1024>, Devices::DeviceMap>(
+      std::make_unique<Memory<128 * 1024>>(config.memory_file, true, PMEM_START,
+                                           PMEM_END),
+      std::make_unique<Devices::DeviceMap>(devices));
+  return pmem;
+}
+
+int pmem_read(int raddr) {
+  void *pmem = pmem_get();
+  auto mem = static_cast<MMap *>(pmem);
+  // TODO: Do memory difftest at memory read and write to diagnose at a finer
+  // granularity
+  mem->trace(raddr, true, regs->get_pc());
+  if (g_skip_memcheck)
+    return mem->read(PMEM_START);
+  return mem->read(raddr);
+}
+
+void pmem_write(int waddr, int wdata, char wmask) {
+  void *pmem = pmem_get();
+  auto mem = static_cast<MMap *>(pmem);
+  mem->trace((std::size_t)waddr, false, regs->get_pc(), wdata);
+  return mem->write((std::size_t)waddr, wdata, wmask);
+}
+
+/* === For gdbstub === */
+
+int npc_read_mem(void *args, size_t addr, size_t len, void *val) {
+  void *pmem = pmem_get();
+  auto mmap = static_cast<MMap *>(pmem);
+  return mmap->copy_to(addr, (uint8_t *)val, len);
+}
+
+int npc_write_mem(void *args, size_t addr, size_t len, void *val) {
+  void *pmem = pmem_get();
+  auto mmap = static_cast<MMap *>(pmem);
+  return mmap->copy_from(addr, (uint8_t *)val, len);
+}
+
+int npc_read_reg(void *args, int regno, size_t *value) {
+  if (regno == 32)
+    *value = regs->get_pc();
+  else
+    *value = (*regs)[regno];
+  return 0;
+}
+
+int npc_write_reg(void *args, int regno, size_t value) { return 1; }
+
+inline void breakpoint_to_action(const Breakpoint *bp, gdb_action_t *res) {
+  if (bp == nullptr) {
+    res->reason = gdb_action_t::ACT_NONE;
+    return;
+  }
+  switch (bp->type) {
+  case BP_SOFTWARE:
+    res->reason = gdb_action_t::ACT_BREAKPOINT;
+    break;
+  case BP_ACCESS:
+    res->reason = gdb_action_t::ACT_WATCH;
+    break;
+  case BP_WRITE:
+    res->reason = gdb_action_t::ACT_WWATCH;
+    break;
+  case BP_READ:
+    res->reason = gdb_action_t::ACT_RWATCH;
+    break;
+  }
+  res->data = bp->addr;
+}
+
+void npc_cont(void *args, gdb_action_t *res) {
+  DbgState *dbg = (DbgState *)args;
+  const Breakpoint *stopped_at = nullptr;
+  stopped_at = top->cont(*dbg->bp);
+  breakpoint_to_action(stopped_at, res);
+}
+
+void npc_stepi(void *args, gdb_action_t *res) {
+  DbgState *dbg = (DbgState *)args;
+  const Breakpoint *stopped_at = nullptr;
+  stopped_at = top->stepi(*dbg->bp);
+  breakpoint_to_action(stopped_at, res);
+}
+
+bool npc_set_bp(void *args, size_t addr, bp_type_t type) {
+  DbgState *dbg = (DbgState *)args;
+  for (const auto &bp : *dbg->bp) {
+    if (bp.addr == addr && bp.type == type) {
+      return true;
+    }
+  }
+  dbg->bp->push_back({.addr = addr, .type = type});
+  return true;
+}
+
+bool npc_del_bp(void *args, size_t addr, bp_type_t type) {
+  DbgState *dbg = (DbgState *)args;
+  for (auto it = dbg->bp->begin(); it != dbg->bp->end(); it++) {
+    if (it->addr == addr && it->type == type) {
+      std::swap(*it, *dbg->bp->rbegin());
+      dbg->bp->pop_back();
+      return true;
+    }
+  }
+  return false;
+}
+
+void npc_on_interrupt(void *args) { ; }
+
+void npc_init(void *args) {
+  DbgState *dbg = (DbgState *)args;
+  void *mem = pmem_get();
+  dbg->bp = new std::vector<Breakpoint>;
+
+  top = new VlModule;
+  regs = new Registers("TOP.Flow.reg_0.regFile_", "TOP.Flow.pc.out");
+  top->setup(config.wavefile, regs);
+  top->reset_eval(10);
+}
+
+bool npc_do_difftest = true;
+
+static gdbstub_t gdbstub_priv;
+arch_info_t npc_isa_arch_info{
+    .target_desc = strdup(TARGET_RV32), .reg_num = 32, .reg_byte = 4};
+
+size_t npc_dbg_state_size = sizeof(DbgState);
+
+} // extern "C"
+
+int gdbstub_loop() {
+  DbgState dbg;
+
+  target_ops npc_gdbstub_ops = {.cont = npc_cont,
+                                .stepi = npc_stepi,
+                                .read_reg = npc_read_reg,
+                                .write_reg = npc_write_reg,
+                                .read_mem = npc_read_mem,
+                                .write_mem = npc_write_mem,
+                                .set_bp = npc_set_bp,
+                                .del_bp = npc_del_bp,
+                                .on_interrupt = NULL};
+
+  if (!config.do_debug) {
+    gdb_action_t res;
+    npc_init(&dbg);
+    npc_cont(&dbg, &res);
+    return !(res.reason == gdb_action_t::ACT_SHUTDOWN);
+  }
+
+  if (!gdbstub_init(&gdbstub_priv, &npc_gdbstub_ops,
+                    (arch_info_t)npc_isa_arch_info, NULL,
+                    config.gdbsocket.c_str())) {
+    return EINVAL;
+  }
+  npc_init(&dbg);
+
+  bool success = gdbstub_run(&gdbstub_priv, &dbg);
+  // gdbstub_close(&gdbstub_priv);
+  return !success;
+}
diff --git a/npc/csrc/Flow/main.cpp b/npc/csrc/Flow/main.cpp
index 069abe4..ff35334 100644
--- a/npc/csrc/Flow/main.cpp
+++ b/npc/csrc/Flow/main.cpp
@@ -1,180 +1,13 @@
-extern "C" {
-#include <gdbstub.h>
-}
-// #include "VFlow___024root.h"
-#include "components.hpp"
-#include <VFlow.h>
 #include <config.hpp>
-#include <cstdint>
-#include <cstdlib>
-#include <devices.hpp>
-#include <memory>
+#include <spdlog/cfg/env.h>
+#include <spdlog/spdlog.h>
 #include <types.h>
-#include <vector>
-#include <vl_wrapper.hpp>
-#include <vpi_user.h>
-#include <vpi_wrapper.hpp>
 
-using Registers = _RegistersVPI<uint32_t, 32>;
-using VlModule = VlModuleInterfaceCommon<VFlow, Registers>;
-
-// SDB::SDB<NPC::npc_interface> sdb_dut;
-bool g_skip_memcheck = false;
-VlModule *top;
-Registers *regs;
-vpiHandle pc = nullptr;
-
-const size_t PMEM_START = 0x80000000;
-const size_t PMEM_END = 0x87ffffff;
-
-struct DbgState {
-  std::vector<Breakpoint> *bp;
-};
-
-extern "C" {
-
-/* === Memory Access === */
-using MMap = MemoryMap<Memory<128 * 1024>, Devices::DeviceMap>;
-void *pmem_get() {
-  static Devices::DeviceMap devices{
-      new Devices::Serial(0x10000000, 0x1000),
-      new Devices::RTC(0x10001000, 0x1000),
-  };
-  static auto pmem = new MemoryMap<Memory<128 * 1024>, Devices::DeviceMap>(
-      std::make_unique<Memory<128 * 1024>>(
-          "/nix/store/"
-          "nv2c00y24qwz1jihfbaip9n1lskbzyb3-am-kernel-riscv32-none-elf-2024-07-"
-          "10/share/am-kernels/add.bin",
-          config.memory_file_binary, PMEM_START, PMEM_END),
-      std::make_unique<Devices::DeviceMap>(devices), config.mtrace_ranges);
-  return pmem;
-}
-
-int pmem_read(int raddr) {
-  void *pmem = pmem_get();
-  auto mem = static_cast<MMap *>(pmem);
-  // TODO: Do memory difftest at memory read and write to diagnose at a finer
-  // granularity
-  if (config.do_mtrace)
-    mem->trace(raddr, true, regs->get_pc());
-  if (g_skip_memcheck)
-    return mem->read(PMEM_START);
-  return mem->read(raddr);
-}
-
-void pmem_write(int waddr, int wdata, char wmask) {
-  void *pmem = pmem_get();
-  auto mem = static_cast<MMap *>(pmem);
-  if (config.do_mtrace)
-    mem->trace((std::size_t)waddr, false, regs->get_pc(), wdata);
-  return mem->write((std::size_t)waddr, wdata, wmask);
-}
-
-/* === For gdbstub === */
-
-int npc_read_mem(void *args, size_t addr, size_t len, void *val) {
-  void *pmem = pmem_get();
-  auto mmap = static_cast<MMap *>(pmem);
-  mmap->copy_to(addr, (uint8_t *)val, len);
-  return 0;
-}
-
-int npc_write_mem(void *args, size_t addr, size_t len, void *val) {
-  void *pmem = pmem_get();
-  auto mmap = static_cast<MMap *>(pmem);
-  mmap->copy_from(addr, (uint8_t *)val, len);
-  return 0;
-}
-
-int npc_read_reg(void *args, int regno, size_t *value) {
-  if (regno == 32)
-    *value = regs->get_pc();
-  else
-    *value = (*regs)[regno];
-  return 0;
-}
-
-int npc_write_reg(void *args, int regno, size_t value) { return 1; }
-
-void npc_cont(void *args, gdb_action_t *res) {
-  DbgState *dbg = (DbgState *)args;
-  *res = top->eval(*dbg->bp);
-}
-
-void npc_stepi(void *args, gdb_action_t *res) {
-  DbgState *dbg = (DbgState *)args;
-  *res = top->eval(*dbg->bp);
-}
-
-bool npc_set_bp(void *args, size_t addr, bp_type_t type) {
-  DbgState *dbg = (DbgState *)args;
-  for (const auto &bp : *dbg->bp) {
-    if (bp.addr == addr && bp.type == type) {
-      return true;
-    }
-  }
-  dbg->bp->push_back({.addr = addr, .type = type});
-  return true;
-}
-
-bool npc_del_bp(void *args, size_t addr, bp_type_t type) {
-  DbgState *dbg = (DbgState *)args;
-  for (auto it = dbg->bp->begin(); it != dbg->bp->end(); it++) {
-    if (it->addr == addr && it->type == type) {
-      std::swap(*it, *dbg->bp->rbegin());
-      dbg->bp->pop_back();
-      return true;
-    }
-  }
-  return false;
-}
-
-void npc_on_interrupt(void *args) { ; }
-
-void npc_init(void *args) {
-  DbgState *dbg = (DbgState *)args;
-  dbg->bp = new std::vector<Breakpoint>;
-
-  top = new VlModule;
-  regs = new Registers("TOP.Flow.reg_0.regFile_", "TOP.Flow.pc.out");
-  top->setup(config.wavefile, regs);
-  top->reset_eval(10);
-}
-
-static gdbstub_t gdbstub_priv;
-arch_info_t isa_arch_info = {
-    .target_desc = strdup(TARGET_RV32), .reg_num = 32, .reg_byte = 4};
-
-size_t argsize = sizeof(DbgState);
-
-} // extern "C"
-
-int gdbstub_loop() {
-  DbgState dbg;
-
-  target_ops npc_gdbstub_ops = {.cont = npc_cont,
-                                .stepi = npc_stepi,
-                                .read_reg = npc_read_reg,
-                                .write_reg = npc_write_reg,
-                                .read_mem = npc_read_mem,
-                                .write_mem = npc_write_mem,
-                                .set_bp = npc_set_bp,
-                                .del_bp = npc_del_bp,
-                                .on_interrupt = NULL};
-
-  if (!gdbstub_init(&gdbstub_priv, &npc_gdbstub_ops, (arch_info_t)isa_arch_info,
-                    strdup("/tmp/gdbstub-npc.sock"))) {
-    return EINVAL;
-  }
-  npc_init(&dbg);
-
-  bool success = gdbstub_run(&gdbstub_priv, &dbg);
-  // gdbstub_close(&gdbstub_priv);
-  return !success;
-}
+int gdbstub_loop();
 
 int main(int argc, char **argv, char **env) {
+  spdlog::cfg::load_env_levels();
   config.cli_parse(argc, argv);
-
+  spdlog::debug("Configuration parsed");
   return gdbstub_loop();
 }
diff --git a/npc/csrc/devices/CMakeLists.txt b/npc/csrc/devices/CMakeLists.txt
index fa33794..3480105 100644
--- a/npc/csrc/devices/CMakeLists.txt
+++ b/npc/csrc/devices/CMakeLists.txt
@@ -1,3 +1,4 @@
 add_library(devices serial.cpp rtc.cpp)
 
 target_include_directories(devices PUBLIC include)
+target_link_libraries(devices PRIVATE spdlog::spdlog)
diff --git a/npc/csrc/devices/include/devices.hpp b/npc/csrc/devices/include/devices.hpp
index 4954ac0..0470a44 100644
--- a/npc/csrc/devices/include/devices.hpp
+++ b/npc/csrc/devices/include/devices.hpp
@@ -1,14 +1,12 @@
-#include <array>
-#include <cassert>
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
 #include <initializer_list>
 #include <iostream>
-#include <iterator>
 #include <map>
 #include <memory>
-#include <unordered_map>
+#include <spdlog/logger.h>
+#include <spdlog/sinks/stdout_color_sinks.h>
 #include <utility>
 
 namespace Devices {
@@ -55,6 +53,7 @@ public:
 
 class DeviceMap {
   std::map<uint64_t, Device *> addr_to_device;
+  std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("devicemap");
 
 public:
   DeviceMap(std::initializer_list<Device *> devices) {
@@ -66,8 +65,8 @@ public:
     auto it = addr_to_device.upper_bound(addr);
     if (it == addr_to_device.begin() ||
         (--it)->second->addr + it->second->len <= addr) {
-      std::cerr << "Use of a unintialized device at memory addr: 0x" << std::hex
-                << addr << std::dec << std::endl;
+      logger->error("Accessed an unintialized device at memory addr: 0x{:x}",
+                    addr);
       return false;
     }
     auto &device = it->second;
diff --git a/npc/csrc_nvboard/Keyboard/main.cpp b/npc/csrc_nvboard/Keyboard/main.cpp
index 8f5ea74..372bbaa 100644
--- a/npc/csrc_nvboard/Keyboard/main.cpp
+++ b/npc/csrc_nvboard/Keyboard/main.cpp
@@ -49,11 +49,10 @@ int main(int argc, char **argv, char **env) {
   while (true) {
     nvboard_update();
     cycle(top, [&] {
-      if (keycode != top->io_ps2_data){
+      if (keycode != top->io_ps2_data) {
         keycode = top->io_ps2_data;
         printf("%d\n", keycode);
-      }
-    });
+      });
   }
   delete top;
-}
\ No newline at end of file
+  }
diff --git a/npc/include/components.hpp b/npc/include/components.hpp
index bf16895..cfd05f1 100644
--- a/npc/include/components.hpp
+++ b/npc/include/components.hpp
@@ -3,23 +3,21 @@
 #define _NPC_COMPONENTS_H_
 #include "types.h"
 #include <array>
-#include <cmath>
 #include <cstdint>
 #include <cstdlib>
-#include <exception>
 #include <filesystem>
 #include <fstream>
 #include <iostream>
-#include <map>
-#include <memory>
-#include <sstream>
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include <spdlog/spdlog.h>
 #include <stdexcept>
 #include <string>
-#include <vector>
 
 template <typename T, std::size_t nr> class _RegistersBase {
   std::array<T, nr> regs;
   T pc;
+  std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("registers");
+
   virtual T fetch_pc() const;
   virtual T fetch_reg(std::size_t id) const;
 
@@ -33,12 +31,6 @@ public:
   }
 };
 
-// class MemoryFile {
-//   std::filesystem::path filepath;
-//   public:
-
-// };
-
 template <std::size_t n> class Memory {
   paddr_t pmem_start, pmem_end;
 
@@ -83,15 +75,13 @@ public:
 template <typename Mem, typename DevMap> class MemoryMap {
   std::unique_ptr<Mem> ram;
   std::unique_ptr<DevMap> devices;
-  const std::vector<std::array<uint64_t, 2>> &trace_ranges;
+  std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("mmap");
 
 public:
-  MemoryMap(std::unique_ptr<Mem> &&ram, std::unique_ptr<DevMap> &&devices,
-            const std::vector<std::array<uint64_t, 2>> &trace_ranges)
-      : ram(std::move(ram)), devices(std::move(devices)),
-        trace_ranges(trace_ranges) {}
+  MemoryMap(std::unique_ptr<Mem> &&ram, std::unique_ptr<DevMap> &&devices)
+      : ram(std::move(ram)), devices(std::move(devices)) {}
+
   void write(paddr_t waddr, word_t wdata, char wmask) {
-    // printf("waddr: 0x%x\n", waddr);
     size_t len = (wmask & 1) + ((wmask & 2) >> 1) + ((wmask & 4) >> 2) +
                  ((wmask & 8) >> 3);
     if (ram->in_pmem(waddr)) {
@@ -99,47 +89,38 @@ public:
     } else if (devices->handle(waddr, (uint8_t *)&wdata, len, true)) {
     }
   }
+
   word_t read(paddr_t raddr) const {
     word_t res = 0;
-    // printf("raddr: 0x%x, in_pmem: %d\n", raddr, ram->in_pmem(raddr));
     if (ram->in_pmem(raddr)) {
       ram->transfer(raddr, (uint8_t *)&res, 4, false);
     } else if (devices->handle(raddr, (uint8_t *)&res, 4, false)) {
     }
     return res;
   }
-  void copy_to(paddr_t addr, uint8_t *buf, size_t len) const {
+
+  int copy_to(paddr_t addr, uint8_t *buf, size_t len) const {
     if (ram->in_pmem(addr)) {
       ram->transfer(addr, buf, len, false);
+      return 0;
     } else {
-      std::cerr << "0x" << std::hex << addr << " not in pmem" << std::endl;
+      return EINVAL;
     }
   }
-  void copy_from(paddr_t addr, uint8_t *buf, size_t len) {
+
+  int copy_from(paddr_t addr, uint8_t *buf, size_t len) {
     if (ram->in_pmem(addr)) {
       ram->transfer(addr, buf, len, true);
+      return 0;
     } else {
-      std::cerr << "0x" << std::hex << addr << " not in pmem" << std::endl;
+      return EINVAL;
     }
   }
+
   void *get_pmem() { return ram->mem.data(); }
+
   void trace(paddr_t addr, bool is_read, word_t pc = 0, word_t value = 0) {
-    for (auto &r : trace_ranges) {
-      if (r[0] <= addr && r[1] >= addr) {
-        std::stringstream os;
-        if (pc != 0)
-          os << "0x" << std::hex << pc << " ";
-        if (is_read)
-          os << "[R] "
-             << "0x" << addr << ": 0x" << this->read(addr);
-        else
-          os << "[W] " << value << " -> "
-             << "0x" << addr;
-        os << std::dec << std::endl;
-        std::cout << os.rdbuf();
-        break;
-      }
-    }
+    logger->trace("[{}] 0x{:x}", is_read ? 'R' : 'W', this->read(addr));
   }
 };
 #endif
diff --git a/npc/include/config.hpp b/npc/include/config.hpp
index c57ca2e..8c12f31 100644
--- a/npc/include/config.hpp
+++ b/npc/include/config.hpp
@@ -3,17 +3,14 @@
 #include <CLI/App.hpp>
 #include <CLI/CLI.hpp>
 #include <CLI/Validators.hpp>
-#include <cstddef>
 #include <filesystem>
-#include <utility>
 
 struct Config {
   std::filesystem::path memory_file;
+  std::string gdbsocket = "gdbstub-npc.sock";
+  bool do_debug{false};
   bool interactive{false};
   bool memory_file_binary = {true};
-  bool do_mtrace{false};
-  std::vector<std::array<std::size_t, 2>> mtrace_ranges{
-      {0x80000000, 0x8ffffffff}};
   std::filesystem::path wavefile;
   std::filesystem::path lib_ref;
   void cli_parse(int argc, char **argv);
diff --git a/npc/include/types.h b/npc/include/types.h
index e991d8a..fcc14e4 100644
--- a/npc/include/types.h
+++ b/npc/include/types.h
@@ -1,6 +1,11 @@
 #ifndef _NPC_TYPES_H__
 #define _NPC_TYPES_H__
+#ifdef __cplusplus
+extern "C" {
+#endif
+#include <gdbstub.h>
 #include <inttypes.h>
+#include <stddef.h>
 
 typedef uint32_t word_t;
 typedef int32_t sword_t;
@@ -16,9 +21,19 @@ typedef uint32_t paddr_t;
 #define FMT_ADDR "0x%08x"
 typedef uint16_t ioaddr_t;
 
+struct Breakpoint {
+  size_t addr;
+  bp_type_t type;
+};
+
+#ifdef __cplusplus
+}
+#endif
+
 #ifdef __cplusplus
 #include <map>
 #include <string>
+#include <vector>
 
 const std::map<std::string, int> riscv32_regs_by_name{
     {"$0", 0},  {"ra", 1},  {"sp", 2},   {"gp", 3},   {"tp", 4},  {"t0", 5},
@@ -27,12 +42,10 @@ const std::map<std::string, int> riscv32_regs_by_name{
     {"s2", 18}, {"s3", 19}, {"s4", 20},  {"s5", 21},  {"s6", 22}, {"s7", 23},
     {"s8", 24}, {"s9", 25}, {"s10", 26}, {"s11", 27}, {"t3", 28}, {"t4", 29},
     {"t5", 30}, {"t6", 31}};
-#endif
 
-#include <gdbstub.h>
-struct Breakpoint {
-  size_t addr;
-  bp_type_t type;
+struct DbgState {
+  std::vector<Breakpoint> *bp;
 };
+#endif
 
 #endif
diff --git a/npc/include/vl_wrapper.hpp b/npc/include/vl_wrapper.hpp
index 02a6ac5..aa8e798 100644
--- a/npc/include/vl_wrapper.hpp
+++ b/npc/include/vl_wrapper.hpp
@@ -2,6 +2,7 @@
 #define _NPC_TRACER_H_
 #include "components.hpp"
 #include "types.h"
+#include "verilated.h"
 #include <filesystem>
 #include <gdbstub.h>
 #include <sys/types.h>
@@ -47,7 +48,7 @@ public:
     registers = r;
   }
 
-  void eval() {
+  void eval(void) {
     if (this->is_posedge()) {
       posedge_cnt++;
     }
@@ -57,35 +58,34 @@ public:
     if (tracer)
       tracer->update();
   }
-  void eval(int n) {
-    for (int i = 0; i < n; i++) {
-      this->eval();
+
+  const Breakpoint *stepi(const std::vector<Breakpoint> &breakpoints) {
+    this->eval();
+    this->eval();
+    size_t pc = registers->get_pc();
+    for (const auto &bp : breakpoints) {
+      if (pc == bp.addr) {
+        return &bp;
+      }
     }
+    return nullptr;
   }
 
-  gdb_action_t eval(const std::vector<Breakpoint> &breakpoints) {
-    gdb_action_t res;
+  const Breakpoint *cont(const std::vector<Breakpoint> &breakpoints) {
+    const Breakpoint *res = nullptr;
     do {
-      this->eval();
-      size_t pc = registers->get_pc();
-      for (const auto &bp : breakpoints) {
-        if (pc == bp.addr) {
-          res.data = bp.addr;
-          switch (bp.type) {
-          default:
-            res.reason = gdb_action_t::ACT_BREAKPOINT;
-          }
-          return res;
-        }
-      }
-    } while (true);
+      res = stepi(breakpoints);
+    } while (res == nullptr);
+    return res;
   }
 
   void reset_eval(int n) {
     extern bool g_skip_memcheck;
     g_skip_memcheck = true;
     this->reset = 1;
-    this->eval(n);
+    do {
+      this->eval();
+    } while (--n);
     this->reset = 0;
     g_skip_memcheck = false;
   }
diff --git a/npc/include/vpi_wrapper.hpp b/npc/include/vpi_wrapper.hpp
index 4aaaeb8..25f26fb 100644
--- a/npc/include/vpi_wrapper.hpp
+++ b/npc/include/vpi_wrapper.hpp
@@ -1,12 +1,16 @@
 #ifndef _NPC_VPI_WRAPPER_H_
 #define _NPC_VPI_WRAPPER_H_
 #include <components.hpp>
+#include <spdlog/logger.h>
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include <spdlog/spdlog.h>
 #include <vpi_user.h>
 
 template <typename T, std::size_t nr>
 class _RegistersVPI : public _RegistersBase<T, nr> {
   std::array<vpiHandle, nr> reg_handles;
   vpiHandle pc_handle;
+
   T vpi_get(vpiHandle vh) const {
     s_vpi_value v;
     v.format = vpiIntVal;