Agentikernel

This is an AI agent, wrapped in a jupyter kernel.

It is made to interact with other kernels, through different levels of authorization :

• read kernel history,

• send code to kernel with user validation

• (send code to kernel without user validation) # not yet implemented

This is a subclass of pydantic-ai-kernel.

Getting Started

pip install agentikernel

Then any jupyter front-end could access the kernel, for example :

pip install jupyter-console
jupyter console --kernel agentikernel

for command line interface. But a jupyter notebook would also work.

Quick start

Create a config file following the scheme declared here. For example :

agent_name: agentik
system_prompt: You are a specialist in code, always ready to analyze code from jupyter kernels.
model:
  model_name: qwen3:1.7b
  model_type: openai
  model_provider:
    name: ollama
    params:
      base_url: http://localhost:11434/v1

a permanent config file can be placed in ~/.jupyter/jupyter_agentikernel_config.yaml.

Then start the kernel jupyter console --kernel agentikernel, and send :

%load_config path_to_config_file.yaml

The kernel is now initialized with an inference provider.

In an other window, start any jupyter kernel (default is python) :

jupyter console --ConnectionFileMixin.connection_file ./kernel_connection_file_test.json

It creates a connection file named kernel_connection_file_test.json

Within the agentikernel, run :

%add_kernel kernel_connection_file_test.json

Then, the agentikernel can use the python kernel in read-only mode. It means it can access the cells that have been running on the kernel.

Commands

All commands from pydantic-ai-kernel are inherited (/load_config, ...).

We've added the following ones :

• %add_kernel path_to_kernel_connection_file --label tool_label

Declares a tool to the pydantic-ai agent, that allows it to run code on the kernel. The kernel must already be running. The kernel connection file is created when you start any jupyter kernel, see jupyter_client. They are stored in a runtime directory, that you can access by runnning jupyter --paths ref - see the Runtime paths.

By default, the agent has read-only access to the kernel. But it can be set to write, by adding the flag: --write.

• %remove_kernel tool_label

Removes the connection with the kernel that has label 'tool_label'. And removes the tool to call it for the agent. The kernel is not stopped, we just undeclare it to the agent. Connection can be made again by running %add_kernel.