diff --git a/.gitignore b/.gitignore
index 130dd0d91..7ec0ccb3d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,6 +4,7 @@
 *.clang-tidy
 *.deb
 *.a
+*.lib
 *.wasm
 
 build
@@ -28,6 +29,8 @@ doc/xml/
 *.tar.gz
 *.raw
 *.so
+*.dll
+*.exe
 *.dylib
 *.dtb
 .DS_Store
diff --git a/src/Makefile b/src/Makefile
index 560df269e..f58b2d5c7 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -37,6 +37,7 @@ GRPC_CPP_PLUGIN=$(shell which grpc_cpp_plugin)
 release?=no
 sanitize?=no
 coverage?=no
+nothreads?=no
 
 COVERAGE_TOOLCHAIN?=gcc
 COVERAGE_OUTPUT_DIR?=coverage
@@ -50,18 +51,19 @@ endif
 
 # Mac OS X specific setup
 ifeq ($(TARGET_OS),Darwin)
+PICCFLAGS=-fPIC
 SOLDFLAGS=-bundle -undefined dynamic_lookup
-LIBLDFLAGS=-dynamiclib -undefined dynamic_lookup
+LIBLDFLAGS=-dynamiclib
 EXELDFLAGS=
+PTHREAD_CFLAGS=
+PTHREAD_LDFLAGS=-lpthread
 CC=clang
 CXX=clang++
-AR_EXEC=libtool -static
+AR=libtool -static -o
 INCS=
 
-PTHREAD_LIB=-lpthread
-
 ifeq ($(MACOSX_DEPLOYMENT_TARGET),)
-export MACOSX_DEPLOYMENT_TARGET := $(shell sw_vers -productVersion | sed -r  "s/([[:digit:]]+)\.([[:digit:]]+)\..+/\1.\2.0/")
+export MACOSX_DEPLOYMENT_TARGET := $(shell sw_vers -productVersion | sed -E "s/([[:digit:]]+)\.([[:digit:]]+)\..+/\1.\2.0/")
 endif
 
 # Homebrew installation
@@ -95,15 +97,17 @@ PROFILE_DATA=default.profdata
 else
 
 # Linux specific setup
-SOLDFLAGS=-shared -fPIC -pthread $(GCLDFLAGS)
-LIBLDFLAGS=$(SOLDFLAGS) -Wl,-z,defs
+PICCFLAGS=-fPIC
+SOLDFLAGS=-shared $(PICCFLAGS) $(GCLDFLAGS)
+LIBLDFLAGS=$(SOLDFLAGS)
 EXELDFLAGS=$(GCLDFLAGS)
+PTHREAD_CFLAGS=-pthread
+PTHREAD_LDFLAGS=-pthread -lpthread
 CC=gcc
 CXX=g++
-AR_EXEC=ar rcs
+AR=ar rcs
 INCS=
 
-PTHREAD_LIB=-lpthread
 BOOST_INC=
 GRPC_PROTOBUF_INC=$(shell pkg-config --cflags-only-I grpc++ protobuf)
 GRPC_PROTOBUF_LIB=$(shell pkg-config --libs grpc++ protobuf)
@@ -125,7 +129,7 @@ LUACARTESI_GRPC_LIBS=$(GRPC_PROTOBUF_LIB)
 LUACARTESI_JSONRPC_LIBS=
 REMOTE_CARTESI_MACHINE_LIBS=$(GRPC_PROTOBUF_LIB)
 JSONRPC_REMOTE_CARTESI_MACHINE_LIBS=
-TEST_MACHINE_C_API_LIBS=$(GRPC_PROTOBUF_LIB) $(PTHREAD_LIB)
+TEST_MACHINE_C_API_LIBS=$(GRPC_PROTOBUF_LIB)
 HASH_LIBS=
 
 #DEFS+= -DMT_ALL_DIRTY
@@ -248,8 +252,16 @@ EMPTY:=
 SPACE:=$(EMPTY) $(EMPTY)
 CLANG_TIDY_HEADER_FILTER=$(PWD)/($(subst $(SPACE),|,$(LINTER_HEADERS)))
 
-CXXFLAGS+=$(OPTFLAGS) -std=gnu++17 -fvisibility=hidden -fPIC -MMD $(CC_MARCH) $(INCS) $(GCFLAGS) $(UBFLAGS) $(DEFS) $(WARNS)
-CFLAGS+=$(OPTFLAGS) -std=gnu99 -fvisibility=hidden -fPIC -MMD $(CC_MARCH) $(INCS) $(GCFLAGS) $(UBFLAGS) $(DEFS) $(WARNS)
+ifeq ($(nothreads),no)
+CFLAGS+=$(PTHREAD_CFLAGS)
+CXXFLAGS+=$(PTHREAD_CFLAGS)
+LDFLAGS+=$(PTHREAD_LDFLAGS)
+else
+DEFS+=-DNO_THREADS
+endif
+
+CXXFLAGS+=$(OPTFLAGS) -std=gnu++17 -fvisibility=hidden -MMD $(PICCFLAGS) $(CC_MARCH) $(INCS) $(GCFLAGS) $(UBFLAGS) $(DEFS) $(WARNS)
+CFLAGS+=$(OPTFLAGS) -std=gnu99 -fvisibility=hidden -MMD $(PICCFLAGS) $(CC_MARCH) $(INCS) $(GCFLAGS) $(UBFLAGS) $(DEFS) $(WARNS)
 LDFLAGS+=$(UBFLAGS)
 
 COVERAGE_WORKLOAD=\
@@ -275,14 +287,14 @@ $(error invalid value for COVERAGE_TOOLCHAIN: $(COVERAGE_TOOLCHAIN))
 endif
 endif
 
-CXXFLAGS+=$(MYCXXFLAGS)
-CFLAGS+=$(MYCFLAGS)
+CXXFLAGS+=$(MYCXXFLAGS) $(MYDEFS)
+CFLAGS+=$(MYCFLAGS) $(MYDEFS)
 LDFLAGS+=$(MYLDFLAGS)
 SOLDFLAGS+=$(MYSOLDFLAGS)
 LIBLDFLAGS+=$(MYLIBLDFLAGS)
 EXELDFLAGS+=$(MYEXELDFLAGS)
 
-all: luacartesi libcartesi.a libcartesi_jsonrpc.a grpc hash c-api jsonrpc-remote-cartesi-machine
+all: libcartesi.a libcartesi_jsonrpc.a c-api luacartesi jsonrpc-remote-cartesi-machine grpc hash
 
 .PHONY: all generate use clean test lint format format-lua check-format check-format-lua luacartesi grpc hash c-api compile_flags.txt
 
@@ -386,10 +398,10 @@ libcartesi_jsonrpc: libcartesi_jsonrpc.a $(LIBCARTESI_JSONRPC)
 libcartesi_jsonrpc.so: $(LIBCARTESI_JSONRPC)
 
 libcartesi.a: $(LIBCARTESI_OBJS)
-	$(AR_EXEC) $@ $(LIBCARTESI_OBJS)
+	$(AR) $@ $(LIBCARTESI_OBJS)
 
 libcartesi_jsonrpc.a: $(LIBCARTESI_JSONRPC_OBJS)
-	$(AR_EXEC) $@ $(LIBCARTESI_JSONRPC_OBJS)
+	$(AR) $@ $(LIBCARTESI_JSONRPC_OBJS)
 
 $(LIBCARTESI): $(LIBCARTESI_OBJS)
 	$(CXX) -o $@ $(LIBCARTESI_OBJS) $(LIBCARTESI_LIBS) $(LDFLAGS) $(LIBCARTESI_LDFLAGS) $(LIBLDFLAGS)
diff --git a/src/json-util.h b/src/json-util.h
index ae61a166c..54f9ce670 100644
--- a/src/json-util.h
+++ b/src/json-util.h
@@ -41,9 +41,7 @@ std::string to_string(const char *s);
 
 // Generate a new optional-like type
 template <int I, typename T>
-struct new_optional : public std::optional<T> {
-    using std::optional<T>::optional;
-};
+struct new_optional : public std::optional<T> {};
 
 // Optional-like type used by parse_args function to identify an optional parameter
 template <typename T>
diff --git a/src/machine-c-api.cpp b/src/machine-c-api.cpp
index 8724ca54a..fc80358d2 100644
--- a/src/machine-c-api.cpp
+++ b/src/machine-c-api.cpp
@@ -17,7 +17,7 @@
 #include <any>
 #include <cstring>
 #include <exception>
-#include <future>
+#include <functional>
 #include <ios>
 #include <optional>
 #include <regex>
@@ -66,8 +66,6 @@ int cm_result_failure(char **err_msg) try { throw; } catch (std::exception &e) {
         return CM_ERROR_LENGTH_ERROR;
     } catch (std::out_of_range &ex) {
         return CM_ERROR_OUT_OF_RANGE;
-    } catch (std::future_error &ex) {
-        return CM_ERROR_FUTURE_ERROR;
     } catch (std::logic_error &ex) {
         return CM_ERROR_LOGIC_ERROR;
     } catch (std::bad_optional_access &ex) {
diff --git a/src/machine-c-api.h b/src/machine-c-api.h
index c290b60d9..4cae7c3cf 100644
--- a/src/machine-c-api.h
+++ b/src/machine-c-api.h
@@ -55,7 +55,6 @@ typedef enum { // NOLINT(modernize-use-using)
     CM_ERROR_DOMAIN_ERROR,
     CM_ERROR_LENGTH_ERROR,
     CM_ERROR_OUT_OF_RANGE,
-    CM_ERROR_FUTURE_ERROR,
     CM_ERROR_LOGIC_ERROR,
     CM_LOGIC_ERROR_END,
     // Bad optional access error
diff --git a/src/machine.cpp b/src/machine.cpp
index 05cd1e2a5..6de5ee927 100644
--- a/src/machine.cpp
+++ b/src/machine.cpp
@@ -15,15 +15,11 @@
 //
 
 #include <boost/range/adaptor/sliced.hpp>
-#include <chrono>
 #include <cinttypes>
 #include <cstdio>
 #include <cstring>
-#include <future>
 #include <iomanip>
 #include <iostream>
-#include <mutex>
-#include <thread>
 
 #include "clint-factory.h"
 #include "dtb.h"
@@ -1448,7 +1444,7 @@ bool machine::verify_dirty_page_maps(void) const {
 }
 
 static uint64_t get_task_concurrency(uint64_t value) {
-    const uint64_t concurrency = value > 0 ? value : std::max(std::thread::hardware_concurrency(), 1U);
+    const uint64_t concurrency = value > 0 ? value : std::max(os_get_concurrency(), UINT64_C(1));
     return std::min(concurrency, static_cast<uint64_t>(THREADS_MAX));
 }
 
@@ -1467,68 +1463,55 @@ bool machine::update_merkle_tree(void) const {
         // For each PMA, we launch as many threads (n) as defined on concurrency
         // runtime config or as the hardware supports.
         const uint64_t n = get_task_concurrency(m_r.concurrency.update_merkle_tree);
-        // The update_page_node_hash function in the machine_merkle_tree is not thread
-        // safe, so we protect it with a mutex
-        std::mutex updatex;
-        // Each thread is launched as a future, whose value tells if the
-        // computation succeeded
-        std::vector<std::future<bool>> futures;
-        futures.reserve(n);
-        for (uint64_t j = 0; j < n; ++j) {
-            futures.emplace_back(std::async((n == 1) ? std::launch::deferred : std::launch::async,
-                [&](int j) -> bool {
-                    auto scratch = unique_calloc<unsigned char>(PMA_PAGE_SIZE, std::nothrow_t{});
-                    if (!scratch) {
-                        return false;
-                    }
-                    machine_merkle_tree::hasher_type h;
-                    // Thread j is responsible for page i if i % n == j.
-                    for (uint64_t i = j; i < pages_in_range; i += n) {
-                        const uint64_t page_start_in_range = i * PMA_PAGE_SIZE;
-                        const uint64_t page_address = pma->get_start() + page_start_in_range;
-                        const unsigned char *page_data = nullptr;
-                        // Skip any clean pages
-                        if (!pma->is_page_marked_dirty(page_start_in_range)) {
-                            continue;
-                        }
-                        // If the peek failed, or if it returned a page for update but
-                        // we failed updating it, the entire process failed
-                        if (!peek(*pma, *this, page_start_in_range, &page_data, scratch.get())) {
+        const bool succeeded = os_parallel_for(n, [&](int j, const parallel_for_mutex &mutex) -> bool {
+            auto scratch = unique_calloc<unsigned char>(PMA_PAGE_SIZE, std::nothrow_t{});
+            if (!scratch) {
+                return false;
+            }
+            machine_merkle_tree::hasher_type h;
+            // Thread j is responsible for page i if i % n == j.
+            for (uint64_t i = j; i < pages_in_range; i += n) {
+                const uint64_t page_start_in_range = i * PMA_PAGE_SIZE;
+                const uint64_t page_address = pma->get_start() + page_start_in_range;
+                const unsigned char *page_data = nullptr;
+                // Skip any clean pages
+                if (!pma->is_page_marked_dirty(page_start_in_range)) {
+                    continue;
+                }
+                // If the peek failed, or if it returned a page for update but
+                // we failed updating it, the entire process failed
+                if (!peek(*pma, *this, page_start_in_range, &page_data, scratch.get())) {
+                    return false;
+                }
+                if (page_data) {
+                    const bool is_pristine = std::all_of(page_data, page_data + PMA_PAGE_SIZE,
+                        [](unsigned char pp) -> bool { return pp == '\0'; });
+
+                    if (is_pristine) {
+                        // The update_page_node_hash function in the machine_merkle_tree is not thread
+                        // safe, so we protect it with a mutex
+                        const parallel_for_mutex_guard lock(mutex);
+                        if (!m_t.update_page_node_hash(page_address,
+                                machine_merkle_tree::get_pristine_hash(machine_merkle_tree::get_log2_page_size()))) {
                             return false;
                         }
-                        if (page_data) {
-                            const bool is_pristine = std::all_of(page_data, page_data + PMA_PAGE_SIZE,
-                                [](unsigned char pp) -> bool { return pp == '\0'; });
-
-                            if (is_pristine) {
-                                const std::lock_guard<std::mutex> lock(updatex);
-                                if (!m_t.update_page_node_hash(page_address,
-                                        machine_merkle_tree::get_pristine_hash(
-                                            machine_merkle_tree::get_log2_page_size()))) {
-                                    return false;
-                                }
-                            } else {
-                                hash_type hash;
-                                m_t.get_page_node_hash(h, page_data, hash);
-                                {
-                                    const std::lock_guard<std::mutex> lock(updatex);
-                                    if (!m_t.update_page_node_hash(page_address, hash)) {
-                                        return false;
-                                    }
-                                }
+                    } else {
+                        hash_type hash;
+                        m_t.get_page_node_hash(h, page_data, hash);
+                        {
+                            // The update_page_node_hash function in the machine_merkle_tree is not thread
+                            // safe, so we protect it with a mutex
+                            const parallel_for_mutex_guard lock(mutex);
+                            if (!m_t.update_page_node_hash(page_address, hash)) {
+                                return false;
                             }
                         }
                     }
-                    return true;
-                },
-                j));
-        }
-        // Check if any thread failed
-        bool succeeded = true;
-        for (auto &f : futures) {
-            succeeded = succeeded && f.get();
-        }
-        // If so, we also failed
+                }
+            }
+            return true;
+        });
+        // If any thread failed, we also failed
         if (!succeeded) {
             m_t.end_update(gh);
             return false;
diff --git a/src/os.cpp b/src/os.cpp
index d302c9b27..0236f6684 100644
--- a/src/os.cpp
+++ b/src/os.cpp
@@ -41,10 +41,17 @@
 #include <iostream>
 #include <string>
 #include <system_error>
+#include <vector>
 
 #include "os.h"
 #include "unique-c-ptr.h"
 
+#ifdef HAVE_THREADS
+#include <future>
+#include <mutex>
+#include <thread>
+#endif
+
 #if defined(HAVE_TTY) || defined(HAVE_MMAP) || defined(HAVE_TERMIOS) || defined(_WIN32)
 #include <fcntl.h> // open
 #endif
@@ -63,6 +70,14 @@
 
 #ifdef _WIN32
 
+#ifndef WIN32_LEAN_AND_MEAN
+#define WIN32_LEAN_AND_MEAN
+#endif
+
+#ifndef _CRT_SECURE_NO_WARNINGS
+#define _CRT_SECURE_NO_WARNINGS
+#endif
+
 #include <direct.h> // mkdir
 #include <io.h>     // _write/_close
 #include <windows.h>
@@ -514,4 +529,40 @@ int64_t os_now_us() {
     return static_cast<int64_t>(std::chrono::duration_cast<std::chrono::microseconds>(end - start).count());
 }
 
+uint64_t os_get_concurrency() {
+#ifdef HAVE_THREADS
+    return std::thread::hardware_concurrency();
+#else
+    return 1;
+#endif
+}
+
+bool os_parallel_for(uint64_t n, const std::function<bool(uint64_t j, const parallel_for_mutex &mutex)> &task) {
+#ifdef HAVE_THREADS
+    if (n > 1) {
+        std::mutex mutex;
+        const parallel_for_mutex for_mutex = {[&] { mutex.lock(); }, [&] { mutex.unlock(); }};
+        std::vector<std::future<bool>> futures;
+        futures.reserve(n);
+        for (uint64_t j = 0; j < n; ++j) {
+            futures.emplace_back(std::async(std::launch::async, task, j, for_mutex));
+        }
+        // Check if any thread failed
+        bool succeeded = true;
+        for (auto &f : futures) {
+            succeeded = succeeded && f.get();
+        }
+        // Return overall status
+        return succeeded;
+    }
+#endif
+    // Run without extra threads when concurrency is 1 or as fallback
+    const parallel_for_mutex for_mutex{[] {}, [] {}};
+    bool succeeded = true;
+    for (uint64_t j = 0; j < n; ++j) {
+        succeeded = succeeded && task(j, for_mutex);
+    }
+    return succeeded;
+}
+
 } // namespace cartesi
diff --git a/src/os.h b/src/os.h
index a6d67aeb0..7440ef85e 100644
--- a/src/os.h
+++ b/src/os.h
@@ -19,6 +19,7 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <functional>
 
 /// \file
 /// \brief System-specific OS handling operations
@@ -55,6 +56,38 @@ void os_unmap_file(unsigned char *host_memory, uint64_t length);
 /// \brief Get time elapsed since its first call with microsecond precision
 int64_t os_now_us();
 
+/// \brief Get the number of concurrent threads supported by the OS
+uint64_t os_get_concurrency();
+
+/// \brief Mutex for os_parallel_for()
+struct parallel_for_mutex {
+    std::function<void()> lock;
+    std::function<void()> unlock;
+};
+
+/// \brief Mutex guard for os_parallel_for()
+struct parallel_for_mutex_guard {
+    parallel_for_mutex_guard(const parallel_for_mutex &mutex) : mutex(mutex) {
+        mutex.lock();
+    }
+    ~parallel_for_mutex_guard() {
+        mutex.unlock();
+    }
+
+    parallel_for_mutex_guard() = delete;
+    parallel_for_mutex_guard(const parallel_for_mutex_guard &) = default;
+    parallel_for_mutex_guard(parallel_for_mutex_guard &&) = default;
+    parallel_for_mutex_guard &operator=(const parallel_for_mutex_guard &) = delete;
+    parallel_for_mutex_guard &operator=(parallel_for_mutex_guard &&) = delete;
+
+private:
+    parallel_for_mutex mutex;
+};
+
+/// \brief Runs a for loop in parallel using up to n threads
+/// \return True if all thread tasks succeeded
+bool os_parallel_for(uint64_t n, const std::function<bool(uint64_t j, const parallel_for_mutex &mutex)> &task);
+
 } // namespace cartesi
 
 #endif
diff --git a/src/test-utils.h b/src/test-utils.h
index ee871ed1c..aac1bb5ff 100644
--- a/src/test-utils.h
+++ b/src/test-utils.h
@@ -31,8 +31,8 @@ using hash_type = cartesi::keccak_256_hasher::hash_type;
 namespace detail {
 
 constexpr int WORD_LOG2_SIZE = 3;
-constexpr int PAGE_LOG2_SIZE = 12;
-constexpr int PAGE_SIZE = (UINT64_C(1) << PAGE_LOG2_SIZE);
+constexpr int MERKLE_PAGE_LOG2_SIZE = 12;
+constexpr int MERKLE_PAGE_SIZE = (UINT64_C(1) << MERKLE_PAGE_LOG2_SIZE);
 
 static hash_type merkle_hash(cartesi::keccak_256_hasher &h, const std::string_view &data, int log2_size) {
     hash_type result;
@@ -89,7 +89,7 @@ static hash_type calculate_proof_root_hash(const cm_merkle_tree_proof *proof) {
 static hash_type calculate_emulator_hash(cm_machine *machine) {
     cartesi::back_merkle_tree tree(64, 12, 3);
     std::string page;
-    page.resize(detail::PAGE_SIZE);
+    page.resize(detail::MERKLE_PAGE_SIZE);
     cm_memory_range_descr_array *mrds = nullptr;
     auto mrds_deleter = [](cm_memory_range_descr_array **mrds) { cm_delete_memory_range_descr_array(*mrds); };
     std::unique_ptr<cm_memory_range_descr_array *, decltype(mrds_deleter)> auto_mrds(&mrds, mrds_deleter);
@@ -102,14 +102,14 @@ static hash_type calculate_emulator_hash(cm_machine *machine) {
     uint64_t last = 0;
     for (size_t i = 0; i < mrds->count; ++i) {
         const auto &m = mrds->entry[i];
-        tree.pad_back((m.start - last) >> detail::PAGE_LOG2_SIZE);
+        tree.pad_back((m.start - last) >> detail::MERKLE_PAGE_LOG2_SIZE);
         auto end = m.start + m.length;
-        for (uint64_t s = m.start; s < end; s += detail::PAGE_SIZE) {
+        for (uint64_t s = m.start; s < end; s += detail::MERKLE_PAGE_SIZE) {
             if (cm_read_memory(machine, s, reinterpret_cast<unsigned char *>(page.data()), page.size(), &err_msg) !=
                 0) {
                 throw std::runtime_error{err_msg};
             }
-            auto page_hash = merkle_hash(page, detail::PAGE_LOG2_SIZE);
+            auto page_hash = merkle_hash(page, detail::MERKLE_PAGE_LOG2_SIZE);
             tree.push_back(page_hash);
         }
         last = end;