witan-xlsx-mcp

npx skills add witanlabs/witan-cli \
  --skill witan-xlsx-mcp

No xlsx_* tools in your tool list? The Witan MCP server isn't connected — point the user at https://api.witanlabs.com/mcp for setup steps.

Goal

Two jobs, one tool:

• Read & Analyze an existing workbook — find data, trace how figures are calculated, answer what-if questions — without guessing or corrupting it.
• Produce & Edit workbooks that are correct, polished, and idiomatic: formula-driven, sensibly formatted, and matched to the workbook's domain.

You are judged on correctness, layout, readability, and idiomatic style. A workbook with formula errors is not finished.

The tool

xlsx_exec runs sandboxed JavaScript against a workbook opened server-side; sibling tools xlsx_calc, xlsx_lint, and xlsx_render handle verification and previews (below). Pass the whole script as the code string — it's JSON-encoded, so no shell escaping ever applies:

// xlsx_exec { file_id: "file_…", code: <this script> }
const sheets = await xlsx.listSheets(wb)
const tsv = await xlsx.readRangeTsv(wb, { sheet: "Summary", from: {row:1,col:1}, to: {row:20,col:8} })
return { sheets, tsv }

• Globals — xlsx (the API) and wb (the open workbook, passed first to every call), plus input (the call's input argument, default {}) and print (like console.log, captured into the response's stdout). Top-level await works; no imports. return sets the response's result.
• filename instead of file_id starts from a brand-new empty workbook — pass exactly one of the two (.xlsx only). It always mints a fresh workbook (never looks one up by name) and names the file if you save.
• save: true persists. Without it every write is ephemeral — it applies in the server session, recalculates, then is discarded; so reads and what-ifs never risk the file, and each run starts clean. Saving a file_id run adds a revision (only if the script wrote cells); saving a filename run mints the new file.
• Read answers from result.touched["Sheet!A1"], never recompute them in JS — after a write the recalculated value is there. Calculating it yourself defeats the engine.

After reading this file, you MUST read references/api.md before your first xlsx_exec call — the function surface is large and not guessable. It replaces the server's manual: do not call read_witan_manual.

api.md groups functions under headings you can grep for — Reading, Searching, Tracing, Computing, Validating, Rendering, Charts, Conditional Formatting, Images, Writing — each with full signatures.

Files in and out

The user's local file is the source of truth; a file_id is a working copy you upload so the server can operate on it, not a place the user looks. Keep the round-trip invisible — never surface the local-vs-server split — and track which local path each file_id maps to.

• Upload — prepare_upload { filename }, then POST the raw bytes (not multipart) to upload_url; the JSON response has id (the file_id) and revision_id.
• Operate — pass file_id to any xlsx_* tool; omit revision_id for the current revision.
• Write back — after any save: true, GET the download_url in the response's output bundle and overwrite the user's local file without being asked. (prepare_download mints a fresh URL any other time.)
• Re-sync — if the user changed the file since you uploaded it, POST the current bytes via prepare_upload_revision before reading. If your working copy disagrees with the user, assume your copy is stale.
• org_id — omit it; it auto-resolves for single-org users. If the call fails with a candidate list, ask the user which org and pass it from then on.

Work efficiently (latency matters)

• Batch independent reads/writes into one xlsx_exec; code takes a whole script, so prefer one rich call over many small ones.
• Don't re-read what you already pulled earlier in the task — reuse it.
• Exception — what-if: there, deliberately split into separate xlsx_exec calls so you can review what you found before editing (see below).

Quality floor

Applies to every workbook you create or change. When editing an existing workbook, inspect or render it first and match its established conventions — the rules below are defaults for new or unspecified work, never a licence to restyle someone's model.

• Formulas, not values — every derived number is a formula so the sheet stays live.
  • ✅ setCells(wb, [{ address: "B10", formula: "=SUM(B2:B9)" }])
  • ❌ summing in JS and writing { address: "B10", value: 4200 }
• No magic numbers — reference an input cell, don't bury a constant.
  • ✅ =B5*(1+$B$6) ❌ =B5*1.05
• Separate inputs from logic — assumptions (rates, growth, multiples) live in their own labelled cells or an Inputs area, never inside formulas.
• Right references — absolute ($B$6) vs relative so fills and copies behave.
• Format to the data type — real dates with a date format (yyyy-mm-dd), not date-looking text; thousands separators for money/counts; sensible decimals on percentages; state units in the header when ambiguous (Revenue ($000s)).
• Lay it out for a human — styled header row; numbers right-aligned, labels left; sane column widths (cap autofit so nothing runs off-screen); modest row heights; whitespace between sections.
• Stay valid — give every Excel Table a unique, explicit name; prefix literal text starting with = using ' so it isn't read as a formula.

Domain conventions go beyond this floor. For financial models, valuations, projections, or IB work, read references/domains/financial-modelling.md.

Reading & understanding a workbook

• Get the lay of the land — listSheets (sheets, used ranges, cross-sheet deps), then describeSheet for a structure map, or readRangeTsv to dump a region cheaply.
• Find things — findCells / findRows are fuzzy and case-insensitive; pass synonym arrays or a RegExp. tableLookup reads a value by row + column label inside a detected table.
• Disambiguate — a label and the formula cell beside it often share text. Use context or read surrounding rows with readRangeTsv, then pick the formula cell, not the label.
• Trace calculations — getCellPrecedents / getCellDependents for local hops; traceToInputs / traceToOutputs for the full chain. Traces can be huge — filter by nearbyLabel, don't print them all.

What-if / sensitivity

For "what happens to Y if X changes?", use two separate xlsx_exec calls so you verify before editing:

1. Locate the output cell (first call) — search for what's asked about, review candidates, confirm you have the formula cell.
2. Trace + change + read (second call) — traceToInputs(output) to confirm X actually drives Y, then setCells to change X and read the new value from result.touched[output]. Missing from touched ⇒ it didn't recalc ⇒ wrong cell.
3. Report baseline → new, and check result.errors is empty (a new error means your edit broke something downstream).

For many values at once, use sweepInputs — all combinations in one call with structured before/after stats. Circular/iterative models: if iterative calc is enabled, setCells honours it and non-convergence shows up as errors; use xlsx_calc (with verify: true) for a standalone full-workbook pass.

Authoring a workbook

Build in this order — it minimises rework and round-trips:

1. Plan the structure — sheets, and where inputs vs calculations vs outputs live. Inputs first.
2. Write content in bulk — headers, data, and formulas in as few setCells / xlsx_exec calls as possible. Seed a formula once, then fill across the range rather than writing each cell.
3. Format — number formats, header styling, widths/heights, alignment, borders (Quality floor above; finance models also follow the domain reference).
4. Add interactivity as needed — data validation for categorical inputs, conditional formatting, then charts/tables.
5. Verify (below) before calling it done.

Prototype ephemerally (omit save) while iterating; add save: true only once the structure is right.

Verify before done — mandatory

A workbook with formula errors is not delivered. Before finishing any authoring or edit:

• Recalculate and check errors — xlsx_calc with verify: true must report zero errors. It lists every error with address and formula. Fix and repeat until clean. (Without verify: true it also persists a new revision — only do that deliberately.)
• Spot-check key ranges — readRangeTsv (with formulas) on totals, a few sample references, and edge rows.
• Confirm layout — xlsx_render (or previewStyles in-script) on the headline range; headers, merges, number formats, and charts look as intended.
• Lint for logic errors — xlsx_lint flags what calc can't: double-counting, approximate-match lookups on unsorted data, mixed currencies/units. Review and resolve or knowingly accept each finding.

Then report what you built (sheets / ranges), write the saved file back over the user's local copy, and for what-ifs report the baseline → new values.