🔌 KitchenSink4XL Community Edition

Landing page · llms.txt (machine-readable capability manifest for agents and LLM crawlers)

Read and edit Excel workbooks with your AI assistant, and know which numbers are calculated, cached or missing.

Read and edit real Excel workbooks from Claude Code, Codex CLI, Copilot CLI or any other MCP client that runs local tools. KitchenSink4XL connects your assistant to .xlsx files with tools for formulas, tables and data checks, and it tells you whether a value was calculated, cached or never computed before you put it in a report. Files are processed on your computer; the only thing that leaves it is what your AI app sends to its own provider. The Community edition is free under the AGPL. The Business edition adds a Windows installer, a signed update channel, a license your company can approve and support.

Works on: Windows, macOS and Linux for the file tools. Recalculation, pivots and the other Excel-powered features need Windows with Microsoft Excel installed.

Install

Pick the route for your AI app. The commands go in PowerShell on Windows or a terminal on macOS and Linux, not into an AI chat. The package routes need Python 3.12 or newer.

Claude Desktop

Install uv, then quit and reopen Claude Desktop. Download the .mcpb file from KitchenSink4XL releases. In Claude Desktop open Settings, then Extensions, then Advanced settings, then Install extension, and choose the file. The bundle fetches the Python package the first time it starts, so the first launch needs a network connection. Restart your session and check that the tools show as connected.

Claude Code or Codex CLI

Install uv, then run the line for your app and restart your session:

claude mcp add xl -s user -- uvx kitchensink4xl

codex mcp add xl -- uvx kitchensink4xl

Any other local MCP client

Use uvx as the command and kitchensink4xl as its argument, or install the package and use kitchensink4xl as the server command:

pip install kitchensink4xl

Then follow your client's guide for adding a local MCP server. Installing the package on its own does not connect it to an AI app.

Business edition

Compare the editions on the pricing page. Already purchased? Your Windows installer and download link are in your license portal.

What it can do

69 tools with every pack enabled, including the pack controls. The default lite surface has 40.

• See whether a value was calculated, cached or missing before you use it.
• Read a range or summarize a large table without pulling every row.
• Write formulas, values and number formats into an existing workbook.
• Add or remove rows and columns while supported references update.
• Sort, filter and format tables for a client or a colleague.
• Import and export CSV and JSON.
• Check a workbook for structural problems before sharing it.
• Use installed Excel on Windows for recalculation, pivots and PDF export.

What is available depends on the packs you enable and the applications installed. The full tool reference is below.

Business edition

Need a license your company can approve and a setup someone supports? The Business edition pairs these tools with a Windows installer, a signed update channel and support under the Business terms. Update checks tell you when a covered release is available; nothing installs on its own. Compare the options on the pricing page. The Community edition stays free under the AGPL, including business use that meets its terms.

Privacy Policy

The tools run on your computer, and KitchenSink4AI receives no documents and no usage data from them. Your AI app may send prompts, file contents and tool results to its own provider under that app's settings and terms. Installing downloads packages, and the Community version check contacts PyPI unless you disable it; those requests carry connection details such as your network address and never your documents. Cloud folders and backups follow their own settings. The Privacy Policy covers the product, purchases and support records.

Not affiliated with, endorsed by, or sponsored by Microsoft Corporation. Microsoft and Excel are trademarks of the Microsoft group of companies.

The number that looks right and isn't

Here is the dirty secret of every tool that reads Excel files: a spreadsheet remembers its answers, not its thinking. When something changes upstream, the old answers just sit there, looking exactly like real numbers. Most AI tools will read one anyway and confidently report it, and you will make a decision based on it. This server refuses to bluff. Every number it hands your AI comes with the truth attached: freshly calculated, remembered from the last save, or missing, with the affected cells named. And when it really matters, it asks Excel itself to recalculate and prove it.

What happens to column F when you insert column D

Insert one column in a real workbook and you find out how much of a spreadsheet lives to the right of where you clicked. Formulas, totals, charts, that table on sheet three that quietly feeds the summary: all of them have opinions about where their cells just went.

Most tools do the cheap version: move the cells, save the file, hope. The damage never shows up when the edit runs. It shows up three days later, in a meeting, in the one total that quietly stopped including its last row.

This server moves everything that should move, and checks the result against what Excel itself would have done. When an edit cannot be done safely, it says no out loud instead of guessing quietly. A refusal costs you a minute. A quiet guess costs you the deal.

Against the rest of the aisle

Feature grids flatter everybody. Here it is instead as the requests an agent actually gets:

Capability survey as of 2026-09-04, compiled from public repositories, documentation, and issue trackers. n/d: not documented in the sources surveyed, which is not a claimed absence. The most popular Excel MCP server is haris-musa/excel-mcp-server, 25 tools on openpyxl, last released 2026-04-12; its own open issues #94 and #118 ask for computed values instead of formula text. openpyxl's documentation states that it does not manage dependencies such as formulae, tables and charts when rows or columns are inserted or deleted. * The rest of the aisle: negokaz/excel-mcp-server (Go, seven tools, paginated reads, live editing on Windows), sbroenne/mcp-server-excel (C#, 31 tools over 326 operations driving the real Excel application, the capability ceiling, but Windows with Excel and exclusive access to the file, and no file-based tier). Two newer servers, PSU3D0/agent-spreadsheet and logisky/logisheets-mcp, embed real calculation engines and are worth watching. Microsoft ships no Excel MCP server, and Claude for Excel is an Office add-in rather than an MCP server. † Star counts read from the GitHub API on 2026-09-04, shown for context: stars say how long a shelf has been in view, not what is on it. Corrections welcome: open an issue.

The packs

The stock is arranged in capability packs, the way a good shop groups what belongs together. A session opens on the lite core and switches on the pack a job needs with one call. Numbers below come straight from scripts/measure_surface.py, never hand-counted.

Quickstart: start lite, enable what you need

A session begins with the lite core. When a task needs more, the agent turns on the pack by name:

enable_tools(["design"])        # styles, conditional formatting, charts, names
enable_tools(["com"])           # real recalculation, real pivots, PDF export

The lite core carries no degraded stand-ins, so the lazy path is a dead end on purpose: a refusal for out-of-scope work names the exact pack and the exact call that unlocks it. Power users who want everything loaded from the start can pin it with KS4XL_MODE=full in the server environment, or a comma-separated pack list. Administrators can lock the selection with KS4XL_PACK_POLICY=locked.

Two numbers that matter

• 129 workbook operations, 67 tools. Many tools here are action multiplexers, so the tool count undersells the surface: manage_worksheet alone performs seven distinct operations, manage_table nine, validate runs nine correctness batteries. The operations figure comes from scripts/count_operations.py, which reads the dispatch values each tool actually validates, out of committed source, and never from hand-math. Its docstring carries the counting definition and what is deliberately excluded. A committed snapshot (scripts/operations_snapshot.json) plus a guard test make a drifting figure a test failure rather than a marketing decision.
• Tiered loading: starts at about 15.5k tokens, scales to everything. A fresh session loads the 40-tool lite core and turns on capability packs only when a task needs them, with one enable_tools call. Load every pack and the full surface measures about 24,200 tokens. All figures come from scripts/measure_surface.py; see Context cost.

Context cost (measured)

Most MCP servers move into your AI's context like a hoarder: everything, up front, whether the job needs it or not. This one starts light and only unpacks a shelf when the work calls for it. Here is the bill, measured by scripts/measure_surface.py:

No other Excel MCP server surveyed publishes what its own tool definitions cost to load. The figures above come from a measuring script that ships with the source, so you can check them yourself. Clients that defer tool schemas until first use pay close to zero until a tool is actually called.

Safety model

• Automatic timestamped backup before every mutation, in two rotating slots in a hidden .ks4xl-backups/ folder next to the workbook. manage_backups lists, restores, and prunes them. Exclude that folder from cloud sync tools: the slots churn on every edit and sync clients can hold locks that slow saves down. The backup slots hold what this server changed, and only that. An edit made in Excel or any other program is never captured, so prev restores the state before the last change made here, not before the last change made anywhere.
• Filter-hidden rows: query_range, export_range, and the aggregates read every row in the range, including rows an autofilter is hiding. Excel's own SUBTOTAL skips hidden rows; these functions do not, and a filtered sheet can therefore sum differently here than in Excel's own status bar.
• Saves are atomic and validated; a failed operation leaves the original byte-identical.
• A round-trip hazard scan runs before a mutating save. Parts the writer cannot preserve are named to the caller, and a loss is refused unless the caller explicitly allows the specific classes the gate named.
• Verify-after-write, with restore-from-backup on failure. verify_com=true on any mutating call adds the deep check: the produced file has to open in a real hidden Excel with no repair prompt, or the backup goes back and the save refuses. KS4XL_VERIFY_COM=1 makes that the default for every save.
• Structural edits are checked against an independent expectation of Excel's own behavior for formulas, tables, named ranges, charts, and cross-sheet references.
• Formula text is never inferred from a leading = character; cell type decides, so text that merely looks like a formula is not re-armed as one on copy, move, sort, or reference rewrite.
• Ambiguous targets are refused with the candidates listed, never guessed.
• The COM tier opens its own hidden instance, journals process IDs, and leaves no orphan Excel processes behind.

Sandboxing (opt-in)

Off by default: with nothing configured, the server reads and writes wherever you point it. Set KS4XL_ALLOWED_ROOTS to a list of directories separated by the OS path separator (; on Windows, : elsewhere) and every path the server touches must resolve inside one of them. Reads are gated as well as writes, since a read outside the sandbox exfiltrates workbook content just as surely as a write plants it. The containment check runs on canonicalized paths, so ..\ traversal, symlink and junction escapes, short names, case tricks, and lookalike sibling directories are all caught. A blocked call refuses with a typed error naming the offending path and the allowed roots before any file is opened.

Testing

1,294 tests in tests/unit, plus a separate local COM gate battery that drives a real Excel. On top of the suite, this release went through the family's gauntlet:

• Adversarial rounds through the raw MCP transport. Roughly 600 tool calls across six waves against scratch workbooks, every call journaled. Thirteen findings confirmed and fixed, each with a regression test.
• A live COM stress round. Nine phases, roughly 600 executor operations including a 200-operation endurance soak, more than 40 distinct Excel processes, all of it unattended with nobody available to dismiss a dialog. Zero critical, zero high, six lower findings, all fixed. It verified 87 properties that held, among them serialization under twelve concurrent submits and zero lost updates under eight concurrent writers on one workbook.
• A numbers-safety gate. 61 checks against a real Excel covering insert and delete of rows and columns, sort, move, merge, and an insert-plus-delete composition, each compared against Excel's own recalculation.
• A formula-fidelity gate. 80 formulas across every documented function family, written through all four writers, opened with zero repair prompts and 80 of 80 correct values.
• A discoverability gate. Six fresh agents starting from the 40-tool lite surface found the right pack six times out of six.
• The author's own field testing, running the tools in insane mode against the installed build from a separate window, which is where the last round of fixes came from.

Zero corruptions across the fixture corpus. That corpus is committed in-tree because the hazard, fidelity and adversarial tests need real Excel-authored workbooks (pivots, slicers, shapes, x14 conditional formatting) that openpyxl cannot author, and every fixture is scrubbed of authoring identity by a test-guarded pass. The corruption figure is the build record's, not a script's output; the script-derived numbers here are the test count and the surface measurements.

Young, and tested like it isn't

This is the newest sink in the family, and it shipped through the hardest gauntlet we have ever run: adversarial rounds through the raw protocol, ground-truth checks against Excel's own arithmetic, interrupted-save torture, and a final round built around one question: can any sequence of operations ruin the only copy of a workbook a business depends on. Every finding was fixed before this release. What it has not had yet is a long life in strangers' spreadsheets, and numbers earn trust in the field. The safety net is structural: a backup before every change, saves that verify before they replace your file, and honest labels on every number it hands you. If Excel is where you live, bring your ugliest workbook to the service counter and tell us what broke.

• 🔧 Found a dead circuit? Something glitched, refused, or came back with the wrong number. Never attach a private workbook; rebuild the structure with placeholder data. Before filing: ask your AI to run get_server_info and paste the output here; it is designed to be safe to share.
• 🔌 Missing an outlet? An Excel capability the sink should also have. Describe the real task behind it and it goes on the workbench.

Engineered not to corrupt

Atomic saves. Automatic backups before every change. And one extra wire, because this is a spreadsheet: before a release ships, every kind of edit is checked against Excel's own math. The wiring gets inspected before the power goes on.

Provided as-is, without warranty of any kind, per the license; the engineering above is simply how seriously your workbooks are taken. Keep backups. (It makes them for you.)

Known limits

• com_render_sheet needs a real desktop. Rendering a sheet to an image goes through Excel's clipboard, and a hidden Excel started in a non-interactive session has no window station to copy through, so the call fails there. It works from a normal signed-in desktop session, which is where anyone actually uses it. Every other COM tool runs fine unattended.
• One user, one machine. This is a desktop tool, not a shared service. The server runs on your computer, over stdio, under your own account, and the COM tier drives an Excel that belongs to your Windows session. Two people cannot point at one installation, and it is not something to stand up on a server for a team.
• Deleting or renaming a sheet does not rewrite references. That is Excel's own behavior, and it is documented on the tool: formulas and defined names pointing at a deleted sheet break to #REF! when Excel opens the file. Audit references first when in doubt.
• Cached values go stale after a file-tier write. Writing through the file tier cannot recompute anything, so a cell downstream of your edit keeps its old cached number until Excel recalculates. That is exactly what the labels are for, and recalculate in the COM pack is the fix.
• Some MCP clients drop a tool's schema when a pack is disabled and do not pick it back up on re-enable, even though the server announces the change both ways. If a re-enabled tool comes back as "no such tool", refresh the tool list on the client side.

Install options and settings

The install routes are at the top of this file. This section covers what you can change once the server is connected.

First launch through uvx downloads and builds the environment and can take 20 to 30 seconds before the server answers; every launch after that starts in about two. If a client reports a timeout on first install, launch once from a terminal and try again.

Claude Desktop extension settings

The extension's settings page offers two switches. Verify every save with Excel turns on the deep check, where a real hidden Excel has to open the saved file cleanly or the backup is restored; it is off by default because it costs a round trip through Excel on every write. Limit the server to one folder confines every path the server touches, reads included, to a directory you pick.

pip install kitchensink4xl

{"mcpServers": {"xl": {"command": "kitchensink4xl", "env": {"KS4XL_MODE": "lite"}}}}

git clone https://github.com/KitchenSink4AI/KitchenSink4XL
cd KitchenSink4XL
python -m venv .venv
.venv\Scripts\pip install -e ".[com]"
claude mcp add xl -s user -- <absolute-path>\.venv\Scripts\xl-mcp.exe

git clone https://github.com/KitchenSink4AI/KitchenSink4XL
cd KitchenSink4XL
python3 -m venv .venv
.venv/bin/pip install -e ".[com]"
claude mcp add xl -s user -- <absolute-path>/.venv/bin/xl-mcp

License

KitchenSink4XL is dual-licensed:

AGPL-3.0 (open source). Free for anyone (individuals, academics, and businesses) for any use that complies with the AGPL's terms. Those terms include sharing source, including your modifications, when you distribute the software or make it available over a network.

Commercial license. For organizations that want to build KitchenSink4XL into their own products or services without the AGPL's source-sharing obligations. Contact licensing@kitchensink4.ai.

Copyright (c) 2026 Alvut Consulting, LLC. KitchenSink4AI is a product line of Alvut Consulting, LLC.

Not affiliated with, endorsed by, or sponsored by Microsoft Corporation. Microsoft and Excel are trademarks of the Microsoft group of companies.