csvfab

Fabulously fast CSV editing.

The CSV editor for files spreadsheets can't open. csvfab opens, filters and sorts 5 million rows in under 3 seconds each, and edits the file in place — on Linux, macOS and Windows, with no build step and no dependency beyond Python and a Chromium-based browser.

▶ Watch the 1-minute demo on YouTube — filmed in real time, nothing sped up.

Speed

5 million rows × 15 columns, 771 MB — a customer export generated by the benchmark:

	csvfab	Modern CSV 2.4.4	VisiData 3.4	Excel, LibreOffice Calc
Open	2.7 s	4.6 s	13 s	stop at 1,048,576 rows
Filter a column	0.9 s	N/A	5.6 s
Sort numbers	2.5 s	53 s	25 s
Sort text	1.5 s	24 s	11 s
Peak memory	2.4 GB	5.6 GB	6.0 GB

At 1 million rows (153 MB), csvfab opens the file in 0.6 s, filters it in 0.2 s and sorts it in 0.3–0.4 s:

• 4 to 10 times faster than VisiData
• 14 to 23 times faster than Modern CSV to sort

How it was measured: median of 3 runs on an Intel Core Ultra 7 laptop, Linux. Modern CSV is the free version, in edit mode (its read-only mode opens the file in 3.1 s but cannot sort).

Measure it on your machine: python3 benchmark/run.py generates the files, runs every tool and checks every result before it reports a time. The method, every size, and published figures for other tools →

Features

• Open anything: tabs, drag & drop, "Open with" from your file manager; delimiter (; , tab |), header line and encoding (UTF-8, Windows-1252, ISO-8859-1, ISO-8859-15, Mac Roman, UTF-16) detected, and changeable from the status bar. An Excel workbook or a JSON file is first written as a CSV beside it, then opened.
• Find what matters: global and per-column filters (accents ignored, regex, invert), or a filter written as a formula (num({Amount}) > 1000 && contains({City}, "lyon")); a column panel with a profile (type, distinct values, min / max / sum / average) and a filter by value; a profile of the whole file, every column on one line; sort by clicking a title (Shift+click to sort on several columns); group by one or more columns with counts, sums and averages into a new tab; go to a row by number (Ctrl+G); find without filtering (Ctrl+F, F3): every match highlighted in the rows shown, stepped through one by one; a scroll strip over the scrollbar with the initials, years or values of the sort column along it, marks for irregular, duplicate and changed rows, and the row under the pointer as you drag.
• See every column at a glance: under each title, a chart of what the column holds — a smooth curve of how numbers and dates spread, with their range and a gauge of how full the column is, and for text a ribbon of the most frequent values ending with the empty cells. Filter, and every chart shows where the rows left sit against the whole file; click a slice of a curve, a value or the empty part to filter on it, or drag across a curve to keep a range (from March to June, from 100 to 500 €).
• Let it tell you what it noticed: Insights reads the file and lists, as cards, what deserves a look — the key column, or the few values that come back in one meant to identify every row; a column that decides another (postal code → city) except on a few rows; values written otherwise than the rest of their column (7501 among 75011, an ISO date among 15/03/2024, DUPONT among Dupont, 12.50 among 10,50); numbers or dates far from all the others; rows that are exact copies; hidden arithmetic (TTC = HT × 1,2, TTC = HT + TVA) on every row or all but a few; empty and constant columns, stray spaces. A click shows the rows concerned, the faulty cell marked with what was expected.
• Fill by example: type the result you want in a cell or two of a column — Dupont beside M. Jean DUPONT, jean.dupont@acme.fr beside Jean and Dupont, 15/03/1984 beside 1984-03-15 — and the rest of the column appears in grey, read the same way from each row, with the rule spelled out; Tab fills it in (one undo), Escape forgets it.
• Totals under every column: a row held at the bottom of the grid sums the amounts of the rows shown, and a click on any column gives its average, median, smallest or largest value, earliest or latest date, filled, empty or distinct count — recomputed as you filter, on millions of rows, without slowing anything down. The filter row can be hidden, its filters still applying.
• See every filter at work: each filter in force — the search, a column's, the values hidden from its panel or a chart, duplicates or marks shown alone — is a chip in the toolbar, removed by its ×; the search options are symbols that light up.
• Rich values: a column of links or e-mail addresses (Ctrl+click opens or writes), of yes / no (checks), of colours (a swatch) or of JSON is recognised and drawn as such; rest the pointer on a value cut short to read it whole, JSON indented. Colour a numeric column by value, and freeze the columns up to any one of them.
• Read the grid at a glance: a column of few values (a status, a source, a city) shows them as pills in the colours of its chart; a sorted view draws a line between groups and keeps the current group's name under the header; the active row is tinted. Hovering a part of a column chart brightens the matching values and marks where they are along the scrollbar. Filtering, the rows slide to close the gaps.
• Right-click everywhere, the browser's menu never shows: a column title offers to split, merge, insert or delete the column, a row number its menu, a cell a menu to copy, paste, copy the row, fill by example, keep or hide the rows holding that value, open the row card or delete the row; a text field its own cut, copy, paste and, for a filter, clear.
• Read one record: the row card (Ctrl+I) shows the selected row as a form, every column with its editable value, and its raw line as it sits in the file — delimiters, quotes and invisible characters marked.
• Debug a broken file: Why does this row look wrong? examines the row's own bytes — expected and observed field counts, the probable cause (an unquoted delimiter, located by the column types it would realign; a line cut by an unquoted line break; a quote never closed; missing fields), the bytes around each anomaly in hex, UTF-8 lines in a Windows-1252 file, garbled accents, NUL and invisible characters, cells a spreadsheet would run as formulas — and proposes a repair, as an ordinary undoable edit.
• Edit like a spreadsheet: in-place and multi-line editing, range selection, copy & paste with Excel or Google Sheets, series fill (drag the corner or Ctrl+D), typing into many cells at once, find & replace with regex, bulk edits on filtered rows, mark or remove duplicates (exact, case-insensitive or slugified), delete hidden rows, split and merge columns, computed columns (formulas), look up values from another open file (like VLOOKUP), convert dates, numbers and phone numbers to one format, clean up the whole file (garbled accents like é, invisible characters, odd spaces, empty rows and columns), remove noise words (a list of words such as SARL, Monsieur or et Fils taken out of one column or of every column, whole words, case and accents ignored), fill empty cells from the value above, compare two versions of a file on a key column, anonymise a file for a demo (names swapped for others with the same initial, e-mails and phone numbers reshaped, postal codes keeping their department, dates shifted — the same value always giving the same fake), insert / move / rename / delete rows and columns — and undo (Ctrl+Z) all the way back to the last save, redo (Ctrl+Y) what you undid, with a review of every pending change against the file on disk before you save.
• Validate a file against its rules: create a schema from a file that is right — column types, required and unique columns, bounds, lengths, lists of allowed values, code formats suggested from the values' shapes, text checks (no invisible characters, no double spaces, no garbled accents, upper case or capitalised) — written beside it as name.schema.json (Table Schema, readable by other tools), then check any later file with it: missing or unexpected columns, and each kind of error counted with examples, a click showing its rows with the faulty cells marked. customers_2026-10-04.csv finds customers.schema.json by itself.
• Pivot table: rows × columns of counts, sums, averages, minimums, maximums or distinct values, dates grouped by month, quarter or year, with totals and a colour scale — computed live on millions of rows (a million in under a second), and a click on any number shows the rows behind it. Copy it into a spreadsheet or open it as a new tab.
• Split and combine files: one file per value of a column (per agency, per month…) in a new folder, every row copied byte for byte; or several files with different delimiters, encodings and column orders combined into one, columns matched by their titles.
• Large files: the file stays as its own bytes in memory, rows are decoded only when shown or searched — a 2 GB CSV of 20 million rows opens in seconds without freezing the window, and a save copies the untouched lines as they are.
• Write safely: opening and saving a file without changing it gives back the same bytes, valid or not — and an edit changes only the bytes of its own line (tested on a corpus of broken files: quoted newlines, mixed line endings, BOMs, UTF-16 with or without BOM and with broken surrogates, mixed encodings, unclosed quotes, NUL bytes, random bytes). Save rewrites the file in its own delimiter, line endings and encoding, keeps a timestamped .bak before the first overwrite, notices when another program changed the file meanwhile, and flags irregular rows. Save as another name, delimiter or encoding, or as a formatted Excel workbook (typed numbers and dates, bold frozen header, filters).

Install

Arch Linux	the script below — the AUR package csvfab will follow once the AUR reopens registrations
macOS (Homebrew)	brew install --cask fabiochelly/csvfab/csvfab — csvfab.app in /Applications, offered in the Finder's "Open with" for CSV files (brew install fabiochelly/csvfab/csvfab for the command line alone)
Windows (winget)	winget install FabioChelly.csvfab
Linux / macOS (script)	curl -fsSL https://raw.githubusercontent.com/fabiochelly/csvfab/main/install.sh | sh
Windows (script)	irm https://raw.githubusercontent.com/fabiochelly/csvfab/main/install.ps1 | iex

Requirements: Python 3.8+ and a Chromium-based browser (Chrome, Chromium, Brave or Edge) — csvfab opens in its own app window with a dedicated profile, so your usual browser is untouched. The scripts install per user, without admin rights; install.sh --uninstall / install.ps1 -Uninstall remove it. On macOS, install.sh also builds csvfab.app in ~/Applications and offers to make it the default app for CSV files (with duti; otherwise: select a .csv, File › Get Info › Open with › csvfab › Change All…).

Use

csvfab                  # open the window
csvfab data.csv more.tsv
csvfab export.xlsx      # converted to export.csv beside it, which is what opens
csvfab server.log app.py  # not a table: opened as raw text, one line per row, coloured
csvfab data.sqlite        # a table of a SQLite database, as a CSV beside it

A file that is not a table (a log, code, configuration) opens in Raw text mode: one line per row, every space kept, syntax colours from the theme's VS Code palette, Ln / Col in the status bar, saved back byte for byte. The delimiter menu of the status bar switches between raw text and a table. A JSON file asks whether to flatten its records into a CSV or open it as text. On a raw text file, Follow in the status bar shows new lines as they are written, as tail -f.

Files opened while the window is already there land in it as new tabs. A double-click on a column's resize handle fits it to its content.

Shortcut
Ctrl+P	command palette — every action, setting, tab, recent file and column, searched as you type
Ctrl+O / Ctrl+S	open / save
Ctrl+W / Ctrl+Q	close the tab / quit
F1	every keyboard shortcut
Ctrl+Z / Ctrl+Y	undo / redo
Ctrl+F, F3 / Shift+F3	find in the rows shown without filtering them — every match highlighted, next / previous match
Ctrl+C / Ctrl+V	copy / paste a range (Excel format)
Ctrl+A, arrows, Shift+arrows	select
Home / End, Ctrl+Home / Ctrl+End	first / last cell of the row, of the file
Enter, F2, double-click, or just type	edit — Enter selects the text, F2 and double-click put the caret at its end; Enter fills every selected cell, Ctrl+Enter a series, Shift+Enter a new line
Ctrl+D	fill the series down
Ctrl+E	fill by example — after a result or two typed in a column; Tab takes the values offered, Escape drops them
Ctrl+I	row card — the selected row as a form, and its raw line
Ctrl+G	go to a row by its number
Alt+← / Alt+→	switch tab

How it works

csvfab starts a small local server (server.py and its bridge/ package, Python standard library only, bound to 127.0.0.1 and protected by a per-session token) and opens viewer.htm (with its ui/ styles and scripts) in a Chromium app window. The page reads the CSV — its records found by a background worker, each row decoded on demand — and renders it; the server reads and writes the files you open, atomically, and builds the Excel workbooks. The server stops by itself a few seconds after the last window closes.

Settings and the browser profile live in ~/.local/state/csvfab (Linux), ~/Library/Application Support/csvfab (macOS) or %LOCALAPPDATA%\csvfab (Windows).

License

MIT © Fabio Chelly