Excel Agent Skills

25 Excel ops (workbook/sheet/data/format/chart/pivot/validation). No Microsoft Excel required (openpyxl).
Two ways to run: full server (stdio/SSE/HTTP) with all tool schemas in context, or the minimal path (slate docs plus run/xl_cli.py or run/*.sh) where you only load one op doc and the runner's stdout.

Token numbers

	Full (all 25 tools)	Minimal (slate + runner)
Baseline	~2379 tokens	~446 total, ~18 per op
Example (3 ops)	293 tokens	~54 tokens

Minimal path uses about 5× fewer tokens for the tool surface in that scenario. Full surface = discoverability and full param docs; minimal = less context, deterministic, easier to debug. When to use which: notes/when-full-surface.md.

Full methodology, per-tool table, and how to reproduce: analysis/CONTEXT-REPORT.md.

Minimal path (slate + runners)

Python 3.10+. Paths must be absolute.

./run/create_workbook.sh /tmp/my.xlsx
./run/create_worksheet.sh /tmp/my.xlsx Sheet2
# or
python3 run/xl_cli.py create_workbook /path/to/file.xlsx
python3 run/xl_cli.py create_worksheet /path/to/file.xlsx Sheet2

Each op has a short doc in slate/*.md. The scripts in run/ (run/<op>.sh and run/xl_cli.py) use the same logic as the server, just no server process.

Full server

Runs over stdio, SSE, or streamable HTTP. See TOOLS.md for how to wire it up.

Operations (25 total)


`create_workbook`	`create_worksheet`
`get_workbook_metadata`	`write_data_to_excel`
`read_data_from_excel`	`format_range`
`merge_cells`	`unmerge_cells`
`get_merged_cells`	`apply_formula`
`validate_formula_syntax`	`create_chart`
`create_pivot_table`	`create_table`
`copy_worksheet`	`delete_worksheet`
`rename_worksheet`	`copy_range`
`delete_range`	`validate_excel_range`
`get_data_validation_info`	`insert_rows`
`insert_columns`	`delete_sheet_rows`
`delete_sheet_columns`

Layout

Path	Purpose
`slate/`	~18 tokens per op, one .md per tool
`run/`	`xl_cli.py` + `run/<op>.sh`
`analysis/`	token counts, baselines, CONTEXT-REPORT
`notes/`	when full vs minimal makes sense