slim-sheet

slim-sheet

A spreadsheet MCP server that small local models can actually use. It reads xlsx xlsm xltx xltm xlsb xls ods numbers csv tsv parquet json locally (no Office, no network), keeps every response under a token budget, and gives agents SQL, cleaning, anomaly detection, a formula copilot, Excel-correct editing and one-call branded dashboards — through 8, 18 or 30 tools depending on how capable the model is.

Works with Claude Desktop / Claude Code (Free or Pro), OpenCode, Codex, Cursor, Cline, Continue, LM Studio, Ollama (via mcpo or Open WebUI) — and as a plain CLI for agents without MCP.

Install

uvx slim-sheet mcp                          # MCP server over stdio

uvx slim-sheet overview sales.xlsx # or use any tool from the shell uvx --from "slim-sheet[all]" slim-sheet mcp # + offline formula engine, encrypted files, Apple Numbers

Host configuration for every client (Claude, Cursor, OpenCode, Codex, LM Studio, Ollama/mcpo, Open WebUI, Cline, Continue) is in docs/hosts.md.

{

"mcpServers": { "slim-sheet": { "command": "uvx", "args": ["slim-sheet", "mcp"], "env": { "SLIM_SHEET_TIER": "lite", "SLIM_SHEET_ROOTS": "/path/to/your/data" } } } }

Tiers

Pick exact tools with SLIM_SHEET_TOOLS=overview,read,query.

Design: professional output with almost no description

The model names *what* to show; slim-sheet decides *how* it looks — layout, colours, fonts, number formats, chart type, sort order and an insight title — from a brand. One call:

build_report(path="sales.xlsx", data_sheet="Sales", title="Sales performance 2025",

kpis=["Revenue=sum:Revenue", "Margin=ratio:Profit/Revenue", "Orders=count"], charts=["trend:Month:Revenue", "rank:Region:Revenue", "share:Product:Revenue"])

gives a one-page dashboard: KPI tiles with live formulas (7.1m · 20.4% · 144), a line titled "Revenue rose 58% from 2025-01 to 2025-12", a sorted bar with the leader highlighted, a labelled doughnut, a source footer. style_sheet(path, sheet, preset="report") does the same for any table and reports the number formats it inferred.

Your company's look, set once. Built-in brands: slim (default), office, modern (Aptos), mono (IBCS-style), gov (UK government statistics style). Or point SLIM_SHEET_BRAND at a JSON file — only name is required:

{ "name": "Acme", "primary": "#C8102E", "body_font": "Arial", "currency": "$", "profile": "finance" }

Every workbook then carries the brand as a real Office theme, so Excel's own table and chart styles follow it too. Rules behind the defaults (IBCS, finance colour coding, UK spreadsheet accessibility guidance, chart-choice tables): PLAN.md §16.

What makes it different

• Reads everything with one interface, including .xls, .xlsb, .ods, Apple Numbers and password-protected xlsx.
• Budgeted output: Markdown tables with A1 coordinates, ~1,500 tokens per response, next cursors to continue.
• Errors that teach: every error has a code, a message and a corrected example call.
• SQL over sheets with DuckDB; messy sheets are split into regions with detected headers ("Report.2").
• Never destroys data: edits go to .slim.xlsx unless overwrite=true, writes are atomic, every replaced file is journaled, and undo steps back.
• Excel-correct structure edits: inserting/deleting rows or renaming a sheet rewrites formulas, names, merges, validations, conditional formats, tables and charts the way Excel does.
• Model checking: anomalies scope='formulas' finds hardcoded numbers, broken patterns, double counting and where an error starts; describe_cell trace= follows formulas across sheets; what_if runs delta, goal seek and sensitivity grids without touching the file.
• Formula copilot: 514-function catalog, version checks (XLOOKUP in Excel 2016 → suggests INDEX/MATCH), offline evaluation with dependency chasing.
• Macros, read-only: read sheet='#vba' shows VBA source, anomalies scope='macros' rates risk, and agents redo a macro's job with tools (prompt replace_macro). slim-sheet never runs macros.
• Excel-exact formulas: newer functions are stored with Excel's _xlfn. prefixes and spilling formulas as real dynamic arrays, with a warning when something blocks the spill.
• Flexible paths, safe by default: give any folder in chat (no config change); system folders, AppData, dot-folders such as .ssh and non-spreadsheet files are refused. Read-only SQL with file access locked, no model-written code is executed, no network.

Configuration

Extras: formula (offline evaluate and what_if), crypto (encrypted xlsx), numbers (Apple Numbers), images (insert pictures with annotate), all.

Prompt library

Ready-made, data-agnostic prompts — a shared CORE method, thirteen ROLES (nine analyst roles: sales, finance, marketing, customer, operations, HR, reconciliation, data quality, general; plus model auditor, workbook editor, macro migrator and file reviewer) and eight OUTPUT patterns (dashboard, dashboard pack, report workbook, clean dataset, reconciliation report, quick answer, audit report, edited workbook), with filled-in example sets, about 50 one-paste task recipes in 8 packs (finance, sales, HR, operations, data prep, Excel tasks, reporting, Vietnam) and compact blocks for small local models: docs/prompts.

Skill packs

skills/building-client-reports turns the tools into client-ready reports: a report brief, role presets (CFO, board, sales, marketing, operations, HR, market research), a standards checklist and evals. Copy it into ~/.claude/skills/ or attach it to any agent (skills/README.md).

Non-MCP hosts

slim-sheet schema --format openai --tier lite      # function-calling schemas for any OpenAI-compatible API

slim-sheet schema --format anthropic --tier full # Messages API tool definitions SLIM_SHEET_TOKEN=... slim-sheet mcp --transport streamable-http --host 0.0.0.0 --port 8765

Development

uv sync --all-extras

uv run pre-commit install # ruff, mypy, manifests, commit-message rule, tests on push uv run python scripts/make_demo.py # sample workbooks in ~/Documents/slim-sheet-demo uv run ruff check . && uv run ruff format --check . && uv run mypy src && uv run pytest -q

Issues and pull requests use the templates in .github/; security reports go through SECURITY.md.

Design and roadmap: PLAN.md · agent rules: AGENTS.md · MIT licensed.