Design goals

There are myriads of JSON libraries out there, and each may even have its reason to exist. Our class had these design goals:

Other aspects were not so important to us:

See the contribution guidelines for more information.


The single required source, file json.hpp is in the src directory or released here. All you need to do is add

#include "json.hpp"

// for convenience
using json = nlohmann::json;

to the files you want to use JSON objects. That's it. Do not forget to set the necessary switches to enable C++11 (e.g., -std=c++11 for GCC and Clang).

:beer: If you are using OS X and Homebrew, just type brew tap nlohmann/json and brew install nlohmann_json and you're set. If you want the bleeding edge rather than the latest release, use brew install nlohmann_json --HEAD.

If you are using the Meson Build System, then you can wrap this repo as a subproject.

If you are using Conan to manage your dependencies, merely add jsonformoderncpp/x.y.z@vthiery/stable to your's requires, where x.y.z is the release version you want to use. Please file issues here if you experience problems with the packages.

If you are using hunter on your project for external dependencies, then you can use the nlohman_json package. Please see the hunter project for any issues regarding the packaging.

:warning: Version 3.0.0 is currently under development. Branch develop is used for the ongoing work and is probably unstable. Please use the master branch for the last stable version 2.1.1.


Beside the examples below, you may want to check the documentation where each function contains a separate code example (e.g., check out emplace()). All example files can be compiled and executed on their own (e.g., file emplace.cpp).

JSON as first-class data type

Here are some examples to give you an idea how to use the class.

Assume you want to create the JSON object

  "pi": 3.141,
  "happy": true,
  "name": "Niels",
  "nothing": null,
  "answer": {
    "everything": 42
  "list": [1, 0, 2],
  "object": {
    "currency": "USD",
    "value": 42.99

With the JSON class, you could write:

// create an empty structure (null)
json j;

// add a number that is stored as double (note the implicit conversion of j to an object)
j["pi"] = 3.141;

// add a Boolean that is stored as bool
j["happy"] = true;

// add a string that is stored as std::string
j["name"] = "Niels";

// add another null object by passing nullptr
j["nothing"] = nullptr;

// add an object inside the object
j["answer"]["everything"] = 42;

// add an array that is stored as std::vector (using an initializer list)
j["list"] = { 1, 0, 2 };

// add another object (using an initializer list of pairs)
j["object"] = { {"currency", "USD"}, {"value", 42.99} };

// instead, you could also write (which looks very similar to the JSON above)
json j2 = {
  {"pi", 3.141},
  {"happy", true},
  {"name", "Niels"},
  {"nothing", nullptr},
  {"answer", {
    {"everything", 42}
  {"list", {1, 0, 2}},
  {"object", {
    {"currency", "USD"},
    {"value", 42.99}

Note that in all these cases, you never need to "tell" the compiler which JSON value you want to use. If you want to be explicit or express some edge cases, the functions json::array and json::object will help:

// a way to express the empty array []
json empty_array_explicit = json::array();

// ways to express the empty object {}
json empty_object_implicit = json({});
json empty_object_explicit = json::object();

// a way to express an _array_ of key/value pairs [["currency", "USD"], ["value", 42.99]]
json array_not_object = { json::array({"currency", "USD"}), json::array({"value", 42.99}) };

Serialization / Deserialization

To/from strings

You can create an object (deserialization) by appending _json to a string literal:

// create object from string literal
json j = "{ \"happy\": true, \"pi\": 3.141 }"_json;

// or even nicer with a raw string literal
auto j2 = R"(
    "happy": true,
    "pi": 3.141

Note that without appending the _json suffix, the passed string literal is not parsed, but just used as JSON string value. That is, json j = "{ \"happy\": true, \"pi\": 3.141 }" would just store the string "{ "happy": true, "pi": 3.141 }" rather than parsing the actual object.

The above example can also be expressed explicitly using json::parse():

// parse explicitly
auto j3 = json::parse("{ \"happy\": true, \"pi\": 3.141 }");

You can also get a string representation (serialize):

// explicit conversion to string
std::string s = j.dump();    // {\"happy\":true,\"pi\":3.141}

// serialization with pretty printing
// pass in the amount of spaces to indent
std::cout << j.dump(4) << std::endl;
// {
//     "happy": true,
//     "pi": 3.141
// }

To/from streams (e.g. files, string streams)

You can also use streams to serialize and deserialize:

// deserialize from standard input
json j;
std::cin >> j;

// serialize to standard output
std::cout << j;

// the setw manipulator was overloaded to set the indentation for pretty printing
std::cout << std::setw(4) << j << std::endl;

These operators work for any subclasses of std::istream or std::ostream. Here is the same example with files:

// read a JSON file
std::ifstream i("file.json");
json j;
i >> j;

// write prettified JSON to another file
std::ofstream o("pretty.json");
o << std::setw(4) << j << std::endl;

Please note that setting the exception bit for failbit is inappropriate for this use case. It will result in program termination due to the noexcept specifier in use.

Read from iterator range

You can also read JSON from an iterator range; that is, from any container accessible by iterators whose content is stored as contiguous byte sequence, for instance a std::vector<std::uint8_t>:

std::vector<std::uint8_t> v = {'t', 'r', 'u', 'e'};
json j = json::parse(v.begin(), v.end());

You may leave the iterators for the range [begin, end):

std::vector<std::uint8_t> v = {'t', 'r', 'u', 'e'};
json j = json::parse(v);

STL-like access

We designed the JSON class to behave just like an STL container. In fact, it satisfies the ReversibleContainer requirement.

// create an array using push_back
json j;

// also use emplace_back

// iterate the array
for (json::iterator it = j.begin(); it != j.end(); ++it) {
  std::cout << *it << '\n';

// range-based for
for (auto& element : j) {
  std::cout << element << '\n';

// getter/setter
const std::string tmp = j[0];
j[1] = 42;
bool foo =;

// comparison
j == "[\"foo\", 1, true]"_json;  // true

// other stuff
j.size();     // 3 entries
j.empty();    // false
j.type();     // json::value_t::array
j.clear();    // the array is empty again

// convenience type checkers

// create an object
json o;
o["foo"] = 23;
o["bar"] = false;
o["baz"] = 3.141;

// also use emplace
o.emplace("weather", "sunny");

// special iterator member functions for objects
for (json::iterator it = o.begin(); it != o.end(); ++it) {
  std::cout << it.key() << " : " << it.value() << "\n";

// find an entry
if (o.find("foo") != o.end()) {
  // there is an entry with key "foo"

// or simpler using count()
int foo_present = o.count("foo"); // 1
int fob_present = o.count("fob"); // 0

// delete an entry

Conversion from STL containers

Any sequence container (std::array, std::vector, std::deque, std::forward_list, std::list) whose values can be used to construct JSON types (e.g., integers, floating point numbers, Booleans, string types, or again STL containers described in this section) can be used to create a JSON array. The same holds for similar associative containers (std::set, std::multiset, std::unordered_set, std::unordered_multiset), but in these cases the order of the elements of the array depends how the elements are ordered in the respective STL container.

std::vector<int> c_vector {1, 2, 3, 4};
json j_vec(c_vector);
// [1, 2, 3, 4]

std::deque<double> c_deque {1.2, 2.3, 3.4, 5.6};
json j_deque(c_deque);
// [1.2, 2.3, 3.4, 5.6]

std::list<bool> c_list {true, true, false, true};
json j_list(c_list);
// [true, true, false, true]

std::forward_list<int64_t> c_flist {12345678909876, 23456789098765, 34567890987654, 45678909876543};
json j_flist(c_flist);
// [12345678909876, 23456789098765, 34567890987654, 45678909876543]

std::array<unsigned long, 4> c_array {{1, 2, 3, 4}};
json j_array(c_array);
// [1, 2, 3, 4]

std::set<std::string> c_set {"one", "two", "three", "four", "one"};
json j_set(c_set); // only one entry for "one" is used
// ["four", "one", "three", "two"]

std::unordered_set<std::string> c_uset {"one", "two", "three", "four", "one"};
json j_uset(c_uset); // only one entry for "one" is used
// maybe ["two", "three", "four", "one"]

std::multiset<std::string> c_mset {"one", "two", "one", "four"};
json j_mset(c_mset); // both entries for "one" are used
// maybe ["one", "two", "one", "four"]

std::unordered_multiset<std::string> c_umset {"one", "two", "one", "four"};
json j_umset(c_umset); // both entries for "one" are used
// maybe ["one", "two", "one", "four"]

Likewise, any associative key-value containers (std::map, std::multimap, std::unordered_map, std::unordered_multimap) whose keys can construct an std::string and whose values can be used to construct JSON types (see examples above) can be used to create a JSON object. Note that in case of multimaps only one key is used in the JSON object and the value depends on the internal order of the STL container.

std::map<std::string, int> c_map { {"one", 1}, {"two", 2}, {"three", 3} };
json j_map(c_map);
// {"one": 1, "three": 3, "two": 2 }

std::unordered_map<const char*, double> c_umap { {"one", 1.2}, {"two", 2.3}, {"three", 3.4} };
json j_umap(c_umap);
// {"one": 1.2, "two": 2.3, "three": 3.4}

std::multimap<std::string, bool> c_mmap { {"one", true}, {"two", true}, {"three", false}, {"three", true} };
json j_mmap(c_mmap); // only one entry for key "three" is used
// maybe {"one": true, "two": true, "three": true}

std::unordered_multimap<std::string, bool> c_ummap { {"one", true}, {"two", true}, {"three", false}, {"three", true} };
json j_ummap(c_ummap); // only one entry for key "three" is used
// maybe {"one": true, "two": true, "three": true}

JSON Pointer and JSON Patch

The library supports JSON Pointer (RFC 6901) as alternative means to address structured values. On top of this, JSON Patch (RFC 6902) allows to describe differences between two JSON values - effectively allowing patch and diff operations known from Unix.

// a JSON value
json j_original = R"({
  "baz": ["one", "two", "three"],
  "foo": "bar"

// access members with a JSON pointer (RFC 6901)
// "two"

// a JSON patch (RFC 6902)
json j_patch = R"([
  { "op": "replace", "path": "/baz", "value": "boo" },
  { "op": "add", "path": "/hello", "value": ["world"] },
  { "op": "remove", "path": "/foo"}

// apply the patch
json j_result = j_original.patch(j_patch);
// {
//    "baz": "boo",
//    "hello": ["world"]
// }

// calculate a JSON patch from two JSON values
json::diff(j_result, j_original);
// [
//   { "op":" replace", "path": "/baz", "value": ["one", "two", "three"] },
//   { "op": "remove","path": "/hello" },
//   { "op": "add", "path": "/foo", "value": "bar" }
// ]

Implicit conversions

The type of the JSON object is determined automatically by the expression to store. Likewise, the stored value is implicitly converted.

// strings
std::string s1 = "Hello, world!";
json js = s1;
std::string s2 = js;

// Booleans
bool b1 = true;
json jb = b1;
bool b2 = jb;

// numbers
int i = 42;
json jn = i;
double f = jn;

// etc.

You can also explicitly ask for the value:

std::string vs = js.get<std::string>();
bool vb = jb.get<bool>();
int vi = jn.get<int>();

// etc.

Arbitrary types conversions

Every type can be serialized in JSON, not just STL-containers and scalar types. Usually, you would do something along those lines:

namespace ns {
    // a simple struct to model a person
    struct person {
        std::string name;
        std::string address;
        int age;

ns::person p = {"Ned Flanders", "744 Evergreen Terrace", 60};

// convert to JSON: copy each value into the JSON object
json j;
j["name"] =;
j["address"] = p.address;
j["age"] = p.age;

// ...

// convert from JSON: copy each value from the JSON object
ns::person p {

It works, but that's quite a lot of boilerplate... Fortunately, there's a better way:

// create a person
ns::person p {"Ned Flanders", "744 Evergreen Terrace", 60};

// conversion: person -> json
json j = p;

std::cout << j << std::endl;
// {"address":"744 Evergreen Terrace","age":60,"name":"Ned Flanders"}

// conversion: json -> person
ns::person p2 = j;

// that's it
assert(p == p2);

Basic usage

To make this work with one of your types, you only need to provide two functions:

using nlohmann::json;

namespace ns {
    void to_json(json& j, const person& p) {
        j = json{ {"name",}, {"address", p.address}, {"age", p.age} };

    void from_json(const json& j, person& p) { ="name").get<std::string>();
        p.address ="address").get<std::string>();
        p.age ="age").get<int>();
} // namespace ns

That's all! When calling the json constructor with your type, your custom to_json method will be automatically called. Likewise, when calling get<your_type>(), the from_json method will be called.

Some important things:

How do I convert third-party types?

This requires a bit more advanced technique. But first, let's see how this conversion mechanism works:

The library uses JSON Serializers to convert types to json. The default serializer for nlohmann::json is nlohmann::adl_serializer (ADL means Argument-Dependent Lookup).

It is implemented like this (simplified):

template <typename T>
struct adl_serializer {
    static void to_json(json& j, const T& value) {
        // calls the "to_json" method in T's namespace

    static void from_json(const json& j, T& value) {
        // same thing, but with the "from_json" method

This serializer works fine when you have control over the type's namespace. However, what about boost::optional, or std::filesystem::path (C++17)? Hijacking the boost namespace is pretty bad, and it's illegal to add something other than template specializations to std...

To solve this, you need to add a specialization of adl_serializer to the nlohmann namespace, here's an example:

// partial specialization (full specialization works too)
namespace nlohmann {
    template <typename T>
    struct adl_serializer<boost::optional<T>> {
        static void to_json(json& j, const boost::optional<T>& opt) {
            if (opt == boost::none) {
                j = nullptr;
            } else {
              j = *opt; // this will call adl_serializer<T>::to_json which will
                        // find the free function to_json in T's namespace!

        static void from_json(const json& j, boost::optional<T>& opt) {
            if (j.is_null()) {
                opt = boost::none;
            } else {
                opt = j.get<T>(); // same as above, but with
                                  // adl_serializer<T>::from_json

How can I use get() for non-default constructible/non-copyable types?

There is a way, if your type is MoveConstructible. You will need to specialize the adl_serializer as well, but with a special from_json overload:

struct move_only_type {
    move_only_type() = delete;
    move_only_type(int ii): i(ii) {}
    move_only_type(const move_only_type&) = delete;
    move_only_type(move_only_type&&) = default;

    int i;

namespace nlohmann {
    template <>
    struct adl_serializer<move_only_type> {
        // note: the return type is no longer 'void', and the method only takes
        // one argument
        static move_only_type from_json(const json& j) {
            return {j.get<int>()};

        // Here's the catch! You must provide a to_json method! Otherwise you
        // will not be able to convert move_only_type to json, since you fully
        // specialized adl_serializer on that type
        static void to_json(json& j, move_only_type t) {
            j = t.i;

Can I write my own serializer? (Advanced use)

Yes. You might want to take a look at unit-udt.cpp in the test suite, to see a few examples.

If you write your own serializer, you'll need to do a few things:

Here is an example, without simplifications, that only accepts types with a size <= 32, and uses ADL.

// You should use void as a second template argument
// if you don't need compile-time checks on T
template<typename T, typename SFINAE = typename std::enable_if<sizeof(T) <= 32>::type>
struct less_than_32_serializer {
    template <typename BasicJsonType>
    static void to_json(BasicJsonType& j, T value) {
        // we want to use ADL, and call the correct to_json overload
        using nlohmann::to_json; // this method is called by adl_serializer,
                                 // this is where the magic happens
        to_json(j, value);

    template <typename BasicJsonType>
    static void from_json(const BasicJsonType& j, T& value) {
        // same thing here
        using nlohmann::from_json;
        from_json(j, value);

Be very careful when reimplementing your serializer, you can stack overflow if you don't pay attention:

template <typename T, void>
struct bad_serializer
    template <typename BasicJsonType>
    static void to_json(BasicJsonType& j, const T& value) {
      // this calls BasicJsonType::json_serializer<T>::to_json(j, value);
      // if BasicJsonType::json_serializer == bad_serializer ... oops!
      j = value;

    template <typename BasicJsonType>
    static void to_json(const BasicJsonType& j, T& value) {
      // this calls BasicJsonType::json_serializer<T>::from_json(j, value);
      // if BasicJsonType::json_serializer == bad_serializer ... oops!
      value = j.template get<T>(); // oops!

Binary formats (CBOR and MessagePack)

Though JSON is a ubiquitous data format, it is not a very compact format suitable for data exchange, for instance over a network. Hence, the library supports CBOR (Concise Binary Object Representation) and MessagePack to efficiently encode JSON values to byte vectors and to decode such vectors.

// create a JSON value
json j = R"({"compact": true, "schema": 0})"_json;

// serialize to CBOR
std::vector<std::uint8_t> v_cbor = json::to_cbor(j);

// 0xa2, 0x67, 0x63, 0x6f, 0x6d, 0x70, 0x61, 0x63, 0x74, 0xf5, 0x66, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x00

// roundtrip
json j_from_cbor = json::from_cbor(v_cbor);

// serialize to MessagePack
std::vector<std::uint8_t> v_msgpack = json::to_msgpack(j);

// 0x82, 0xa7, 0x63, 0x6f, 0x6d, 0x70, 0x61, 0x63, 0x74, 0xc3, 0xa6, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x00

// roundtrip
json j_from_msgpack = json::from_msgpack(v_msgpack);

Supported compilers

Though it's 2016 already, the support for C++11 is still a bit sparse. Currently, the following compilers are known to work:

I would be happy to learn about other compilers/versions.

Please note:

    APP_STL := c++_shared
    APP_CPPFLAGS += -frtti -fexceptions

The code compiles successfully with Android NDK, Revision 9 - 11 (and possibly later) and CrystaX's Android NDK version 10.

The following compilers are currently used in continuous integration at Travis and AppVeyor:

Compiler Operating System Version String
GCC 4.9.4 Ubuntu 14.04.5 LTS g++-4.9 (Ubuntu 4.9.4-2ubuntu1~14.04.1) 4.9.4
GCC 5.4.1 Ubuntu 14.04.5 LTS g++-5 (Ubuntu 5.4.1-2ubuntu1~14.04) 5.4.1 20160904
GCC 6.3.0 Ubuntu 14.04.5 LTS g++-6 (Ubuntu/Linaro 6.3.0-18ubuntu2~14.04) 6.3.0 20170519
GCC 7.1.0 Ubuntu 14.04.5 LTS g++-7 (Ubuntu 7.1.0-5ubuntu2~14.04) 7.1.0
Clang 3.5.0 Ubuntu 14.04.5 LTS clang version 3.5.0-4ubuntu2~trusty2 (tags/RELEASE_350/final)
Clang 3.6.2 Ubuntu 14.04.5 LTS clang version 3.6.2-svn240577-1~exp1 (branches/release_36)
Clang 3.7.1 Ubuntu 14.04.5 LTS clang version 3.7.1-svn253571-1~exp1 (branches/release_37)
Clang 3.8.0 Ubuntu 14.04.5 LTS clang version 3.8.0-2ubuntu3~trusty5 (tags/RELEASE_380/final)
Clang 3.9.1 Ubuntu 14.04.5 LTS clang version 3.9.1-4ubuntu3~14.04.2 (tags/RELEASE_391/rc2)
Clang 4.0.1 Ubuntu 14.04.5 LTS clang version 4.0.1-svn305264-1~exp1 (branches/release_40)
Clang 5.0.0 Ubuntu 14.04.5 LTS clang version 5.0.0-svn310902-1~exp1 (branches/release_50)
Clang Xcode 6.4 Darwin Kernel Version 14.3.0 (OSX 10.10.3) Apple LLVM version 6.1.0 (clang-602.0.53) (based on LLVM 3.6.0svn)
Clang Xcode 7.3 Darwin Kernel Version 15.0.0 (OSX 10.10.5) Apple LLVM version 7.3.0 (clang-703.0.29)
Clang Xcode 8.0 Darwin Kernel Version 15.6.0 Apple LLVM version 8.0.0 (clang-800.0.38)
Clang Xcode 8.1 Darwin Kernel Version 16.1.0 (macOS 10.12.1) Apple LLVM version 8.0.0 (clang-800.0.42.1)
Clang Xcode 8.2 Darwin Kernel Version 16.1.0 (macOS 10.12.1) Apple LLVM version 8.0.0 (clang-800.0.42.1)
Clang Xcode 8.3 Darwin Kernel Version 16.5.0 (macOS 10.12.4) Apple LLVM version 8.1.0 (clang-802.0.38)
Clang Xcode 9 beta Darwin Kernel Version 16.6.0 (macOS 10.12.5) Apple LLVM version 9.0.0 (clang-900.0.26)
Visual Studio 14 2015 Windows Server 2012 R2 (x64) Microsoft (R) Build Engine version 14.0.25420.1
Visual Studio 2017 Windows Server 2016 Microsoft (R) Build Engine version 15.1.1012.6693


The class is licensed under the MIT License:

Copyright © 2013-2017 Niels Lohmann

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.



I deeply appreciate the help of the following people.

Thanks a lot for helping out! Please let me know if I forgot someone.

Used third-party tools

The library itself contains of a single header file licensed under the MIT license. However, it is built, tested, documented, and whatnot using a lot of third-party tools and services. Thanks a lot!

Projects using JSON for Modern C++

The library is currently used in Apple macOS Sierra and iOS 10. I am not sure what they are using the library for, but I am happy that it runs on so many devices.


Execute unit tests

To compile and run the tests, you need to execute

$ make json_unit -Ctest
$ ./test/json_unit "*"

All tests passed (14504461 assertions in 48 test cases)

Alternatively, you can use CMake and run

mkdir build
cd build
cmake ..

For more information, have a look at the file .travis.yml.