Support reading cell source only and clearing cell outputs

[koosty]

Problem

When updating a notebook cell, the LLM often needs to read the current cell content before making changes. However, if the cell has already been executed and contains large outputs—such as long text logs, tables, or embedded images—this output is included in the cell’s content. As a result, the LLM’s context window can quickly become saturated with irrelevant output data, reducing available space for meaningful reasoning or edits. Currently, there is no tool to clear a cell’s output before reading it, nor a way to read a cell’s source without its output.

Proposed Solution

• Enhance the read_cell MCP tool to accept an optional parameter (e.g., include_output: bool = True) so LLM can request only the source code/metadata of a cell without its output.
• Add a new tool, e.g., clear_cell_output, that removes the output of a specified cell (or all cells), enabling clean reads or reducing notebook size before LLM interaction.

These changes would allow LLM-driven agents to efficiently interact with notebooks by focusing only on relevant (executable) content and managing context length effectively.

Additional context

• Similar functionality exists in Jupyter frontends (e.g., “Cell → Clear Output” in JupyterLab)
• Jupyter’s internal model already distinguishes between source and outputs in cell objects, so this separation is natively supported.

[ChengJiale150]

Hello, @koosty ! In my #139 , I have change the default output of read_cells tool and rename it to read_notebook in detailed mode, now it can only return cell source and support pagination to avoid return too many token at once. Now the cell output will only return by read_cell tool with include_output.

[eleonorecharles]

Hello @koosty, thanks for opening the issue! In addition to the improvements suggested by @ChengJiale150, an additional tool clear_cell_output could be added thanks to the support of all JupyterLab commands (see #128).