Implementing a kernel

In most of the cases, the base kernel implementation is enough, and creating a kernel only means implementing the interpreter part.

The structure of your project should at least look like the following:

└── example/
    ├── src/
    │   ├── custom_interpreter.cpp
    │   ├── custom_interpreter.hpp
    │   └── main.cpp
    ├── share/
    │   └── jupyter/
    │       └── kernels/
    │           └── my_kernel/
    │               └──
    └── CMakeLists.txt

Implementing the interpreter

Let’s start by editing the custom_interpreter.hpp file, it should contain the declaration of your interpreter class:


#include "xeus/xinterpreter.hpp"

#include "nlohmann/json.hpp"

using xeus::xinterpreter;

namespace nl = nlohmann;

namespace custom
    class custom_interpreter : public xinterpreter


        custom_interpreter() = default;
        virtual ~custom_interpreter() = default;


        void configure_impl() override;

        nl::json execute_request_impl(int execution_counter,
                                      const std::string& code,
                                      bool silent,
                                      bool store_history,
                                      nl::json user_expressions,
                                      bool allow_stdin) override;

        nl::json complete_request_impl(const std::string& code,
                                       int cursor_pos) override;

        nl::json inspect_request_impl(const std::string& code,
                                      int cursor_pos,
                                      int detail_level) override;

        nl::json is_complete_request_impl(const std::string& code) override;

        nl::json kernel_info_request_impl() override;

        void shutdown_request_impl() override;



Almost all custom_interpreter methods return a nl::json instance. This is actually using nlohmann json which is a modern C++ implementation of a JSON datastructure.

Code Execution

Then, you would need to implement all of those methods one by one in the custom_interpreter.cpp file. The main method is of course the execute_request_impl which executes the code whenever the client is sending an execute request.

nl::json custom_interpreter::execute_request_impl(int execution_counter, // Typically the cell number
                                                  const std::string& /*code*/, // Code to execute
                                                  bool /*silent*/,
                                                  bool /*store_history*/,
                                                  nl::json /*user_expressions*/,
                                                  bool /*allow_stdin*/)
    // You can use the C-API of your target language for executing the code,
    // e.g. `PyRun_String` for the Python C-API
    //      `luaL_dostring` for the Lua C-API

    // Use this method for publishing the execution result to the client,
    // this method takes the ``execution_counter`` as first argument,
    // the data to publish (mime type data) as second argument and metadata
    // as third argument.
    // Replace "Hello World !!" by what you want to be displayed under the execution cell
    nl::json pub_data;
    pub_data["text/plain"] = "Hello World !!";
    publish_execution_result(execution_counter, std::move(pub_data), nl::json());

    // You can also use this method for publishing errors to the client, if the code
    // failed to execute
    // publish_execution_error(error_name, error_value, error_traceback);
    publish_execution_error("TypeError", "123", {"!@#$", "*(*"});

    nl::json result;
    result["status"] = "ok";
    return result;

The result and arguments of the execution request are described in the execute_request documentation.


The other methods are all optional, but we encourage you to implement them in order to have a fully-featured kernel.

Input request

For input request support, you would need to monkey-patch the language functions that prompt for a user input (input and raw_input in Python, in Lua etc) and call xeus::blocking_input_request instead. The second parameter should be set to False is what the user is typing should not be visible on the screen.

#include "xeus/xinput.hpp"

xeus::blocking_input_request("User name:", true);
xeus::blocking_input_request("Password:", false);


The configure_impl method allows you to perform some operations after the custom_interpreter creation and before executing any request. This is optional, but it can be useful, for example it is used in xeus-python for initializing the auto-completion engine.

void custom_interpreter::configure_impl()
    // Perform some operations

Code Completion

The complete_request_impl method allows you to implement the auto-completion logic for your kernel.

nl::json custom_interpreter::complete_request_impl(const std::string& code,
                                                   int cursor_pos)
    nl::json result;

    // Code starts with 'H', it could be the following completion
    if (code[0] == 'H')
        result["status"] = "ok";
        result["matches"] = {"Hello", "Hey", "Howdy"};
        result["cursor_start"] = 5;
        result["cursor_end"] = cursor_pos;
    // No completion result
        result["status"] = "ok";
        result["matches"] = {};
        result["cursor_start"] = cursor_pos;
        result["cursor_end"] = cursor_pos;

    return result;

The result and arguments of the completion request are described in the complete_request documentation.

Code Inspection

Allows the kernel user to inspect a variable/class/type in the code. It takes the code and the cursor position as arguments, it is up to the kernel author to extract the token at the given cursor position in the code in order to know for which name the user wants inspection.

nl::json custom_interpreter::inspect_request_impl(const std::string& code,
                                                  int /*cursor_pos*/,
                                                  int /*detail_level*/)
    nl::json result;

    if ("print") == 0)
        result["found"] = true;
        result["text/plain"] = "Print objects to the text stream file, [...]";
        result["found"] = false;

    result["status"] = "ok";
    return result;

The result and arguments of the inspection request are described in the inspect_request documentation.

Code Completeness

This request is never called from the Notebook or from JupyterLab clients, but it is called from the Jupyter console client. It allows the client to know if the user finished typing his code, before sending an execute request. For example, in Python, the following code is not considered as complete:

def foo:

So the kernel should return “incomplete” with an indentation value of 4 for the next line.

The following code is considered as complete:

def foo:

So the kernel should return “complete”.

nl::json custom_interpreter::is_complete_request_impl(const std::string& /*code*/)
    nl::json result;

    // if (is_complete(code))
    // {
        result["status"] = "complete";
    // }
    // else
    // {
    //    result["status"] = "incomplete";
    //    result["indent"] = 4;

    return result;

The result and arguments of the completness request are described in the is_complete_request documentation.

Kernel info

This request allows the client to get information about the kernel: language, language version, kernel version, etc.

nl::json custom_interpreter::kernel_info_request_impl()
    nl::json result;
    result["implementation"] = "my_kernel";
    result["implementation_version"] = "0.1.0";
    result["language_info"]["name"] = "python";
    result["language_info"]["version"] = "3.7";
    result["language_info"]["mimetype"] = "text/x-python";
    result["language_info"]["file_extension"] = ".py";
    return result;

The result and arguments of the kernel info request are described in the kernel_info_request documentation.

Kernel shutdown

This allows you to perform some operations before shutting down the kernel.

void custom_interpreter::shutdown_request_impl() {
    std::cout << "Bye!!" << std::endl;

Implementing the main entry

Now let’s edit the main.cpp file which is the main entry for the kernel executable.

#include <memory>

#include "xeus/xkernel.hpp"
#include "xeus/xkernel_configuration.hpp"

#include "custom_interpreter.hpp"

int main(int argc, char* argv[])
    // Load configuration file
    std::string file_name = (argc == 1) ? "connection.json" : argv[2];
    xeus::xconfiguration config = xeus::load_configuration(file_name);

    // Create interpreter instance
    using interpreter_ptr = std::unique_ptr<custom::custom_interpreter>;
    interpreter_ptr interpreter = interpreter_ptr(new custom::custom_interpreter());

    // Create kernel instance and start it
    xeus::xkernel kernel(config, xeus::get_user_name(), std::move(interpreter));

    return 0;

Kernel file

The kernel.json file is a json file used by Jupyter in order to retrieve all the available kernels.

It must be installed in the INSTALL_PREFIX/share/jupyter/kernels/my_kernel directory, we will see how to do that in the next chapter.

This json file contains:

  • display_name: the name that the Jupyter client should display in its interface (e.g. on the main JupyterLab page).

  • argv: the command that the Jupyter client needs to run in order to start the kernel. You should leave this value unchanged if you are not sure what you are doing.

  • language: the target language of your kernel.

You can edit the file as following. This file will be used by cmake for generating the actual kernel.json file which will be installed.

    "display_name": "my_kernel",
    "argv": [
    "language": "python"


You can provide logos that will be used by the Jupyter client. Those logos should be in files named logo-32x32.png and logo-64x64.png (32x32 and 64x64 being the size of the image in pixels), they should be placed next to the file.

Compiling and installing the kernel

Your CMakeLists.txt file should look like the following:

cmake_minimum_required(VERSION 3.4.3)

set(EXECUTABLE_NAME my_kernel)

# Configuration
# =============


# We generate the kernel.json file, given the installation prefix and the executable name
configure_file (

# Dependencies
# ============

# Be sure to use recent versions
set(xeus_REQUIRED_VERSION 0.19.1)
set(cppzmq_REQUIRED_VERSION 4.3.0)

find_package(xeus ${xeus_REQUIRED_VERSION} REQUIRED)
find_package(cppzmq ${cppzmq_REQUIRED_VERSION} REQUIRED)

# Flags
# =====



    if (HAS_CPP14_FLAG)
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++14")
        message(FATAL_ERROR "Unsupported compiler -- xeus requires C++14 support!")

# Target and link
# ===============

# my_kernel source files

# My kernel executable
add_executable(${EXECUTABLE_NAME} src/main.cpp ${MY_KERNEL_SRC} )
target_link_libraries(${EXECUTABLE_NAME} PRIVATE xeus)

if (APPLE)
    set_target_properties(${EXECUTABLE_NAME} PROPERTIES
    set_target_properties(${EXECUTABLE_NAME} PROPERTIES

set_target_properties(${EXECUTABLE_NAME} PROPERTIES

# Installation
# ============

# Install my_kernel

# Configuration and data directories for jupyter and my_kernel
set(XJUPYTER_DATA_DIR "share/jupyter" CACHE STRING "Jupyter data directory")

# Install Jupyter kernelspecs
set(MY_KERNELSPEC_DIR ${CMAKE_CURRENT_SOURCE_DIR}/share/jupyter/kernels)
        PATTERN "*.in" EXCLUDE)

Now you should be able to install your new kernel and use it with any Jupyter client.

For the installation you first need to install dependencies, the easier way is using conda:

conda install -c conda-forge cmake jupyter xeus xtl nlohmann_json cppzmq

Then create a build folder in the repository and build the kernel from there:

mkdir build
cd build
make install

That’s it! Now if you run the Jupyter Notebook interface you should be able to create a new Notebook selecting the my_kernel kernel. Congrats!

Writing unit-tests for your kernel

For writing unit-tests for you kernel, you can use the jupyter_kernel_test Python library. It allows you to test the results of the requests you send to the kernel.