CSV to JSON Converter

A zero-dependency, browser-local CSV (and TSV) to JSON converter with a real RFC 4180 parser at its core. Paste a spreadsheet export, a database dump, an API CSV response, or anything with delimiters — get back clean, typed JSON without the round-trip to a server. Auto-delimiter detection, header-row resolution, type coercion, and error reporting all run locally, and the output updates on every keystroke.

Why a dedicated CSV→JSON converter

Almost every online "CSV to JSON" tool is built on input.split("\n").map(l => l.split(",")). That breaks on the first quoted field containing a comma, the first embedded newline, the first escaped quote, the first CRLF line ending. Failures are usually silent — you get JSON that looks plausible but has cells shifted across columns and quotes leaked into values.

The second problem is privacy. CSV is one of the most common ways to move sensitive tabular data: customer exports, billing rows, invoice lines, HR spreadsheets. Server-side converters log every paste, often through a CDN and an error tracker too. A browser-local parser eliminates that exposure.

The third problem is type handling. Naïve coercion (parseFloat(cell)) is how databases end up with 7 in a column that used to hold "007". CSV is a string-only format, so the safe default is strings; anything else has to be opt-in with rules that recognise identifier patterns.

This tool runs a character-by-character state-machine parser locally. The tokeniser is ~200 lines of TypeScript, zero runtime dependencies, and handles the full RFC 4180 surface. Type coercion runs as a separate pass, so the parse itself can never lie about the input. Errors carry line and column so you can fix the source.

What it does

• Parse CSV, TSV, or any delimited text — comma, semicolon, tab, pipe, or a custom single-character delimiter.
• Auto-detect the delimiter by scoring the first ten logical lines for column consistency, with the pick shown as a badge on the input panel.
• Handle real-world edge cases: quoted fields with commas, escaped quotes (""), embedded newlines, CR / LF / CRLF endings, leading UTF-8 BOM, ragged rows, trailing blank lines.
• Optional type coercion: numeric-looking cells to numbers (preserving leading-zero identifiers), "true" / "false" to booleans, empty cells to JSON null.
• Header-row toggle: emit an array of objects keyed by the first row, or an array of arrays when the data has no column names. Empty header cells get column_N placeholders.
• Trim whitespace on unquoted cells without touching quoted content, and promote empty strings to null with one click.
• Output control: pretty-print with 2-space indent or minify to a single line. A row-count badge shows exactly how many records you got.
• Typed parse errors: unclosed quotes return an explicit "line X, column Y" message instead of silently emitting a shifted matrix.
• URL-hash share: input and every option encoded into #s=… for private, peer-to-peer sharing.

Under the hood

The parser is a single-pass state machine with two states — inside quotes and outside — handling one character per iteration. That structure makes RFC 4180 tractable: every edge case is a small transition rule rather than a regex with backtracking.

Outside quotes, a leading double-quote flips the state to "inside" and marks the field as quoted. A delimiter ends the field; LF, CR, or CRLF ends the row. A stray mid-field quote like ab"cd stays literal, matching csv-parse with relax_quotes on.

Inside quotes, every character is literal including commas and newlines. A double-quote closes the field, or when paired as "" is an escaped literal quote. Embedded CRLF pairs are normalised to \n.

Delimiter auto-detection runs before the main parse. It samples the first ten non-blank logical lines — honouring quoted newlines so a quoted \n doesn't split a row — and scores each candidate by average column count and variance. Ties fall back to , ; \t |.

Type coercion is conservative. The number check requires optional sign, integer, optional fraction, optional exponent, with no leading zeros on multi-digit integers, so 007, 01234, and 00.5 stay strings. Boolean coercion only matches exact "true" or "false" (case-insensitive); it won't misread "True story". Empty strings become null only when you opt in.

Header mode fills missing column names with column_N placeholders. Ragged rows shorter than the header pad with empty strings; longer rows drop extras, keeping the output matrix rectangular.

RFC 4180 compliance

RFC 4180 is a 2005 informational memo that codified CSV as the industry had settled on it. It is the de-facto reference every CSV tool targets. This parser implements the parts that matter:

• Fields separated by the delimiter; rows separated by CRLF, bare LF, bare CR, or any mix.
• Double-quoted fields that may contain the delimiter, CR, LF, and CRLF.
• Double-quote escaping via doubling ("" represents a literal " inside a quoted field).
• Optional header row, toggleable in the UI.
• Leading UTF-8 BOM tolerated and stripped silently — Excel on Windows emits it.

Behaviours outside the memo that this tool handles anyway:

• Delimiters other than comma — semicolon (European Excel), tab (TSV), pipe (legacy ETL).
• Mixed line endings inside a single document.
• Trailing blank rows, dropped automatically.
• Ragged rows, padded or truncated to match the header.
• Stray unmatched quotes mid-field, kept as literal characters rather than raising an error.

Privacy

No analytics on input content, no server round-trip, no third-party script in the transform path. The parser is a pure function of your CSV and your options — no network calls, no logging, no server round-trip. We do load Google Analytics 4 for aggregate page-view counts (cookieless by default, honours DNT/GPC — details), but it never receives textarea content, file contents, or URL-hash state. The Share button encodes a base64url snapshot into the URL fragment (#s=…), which HTTP spec guarantees is never transmitted to any server. Sharing a link is a peer-to-peer transfer of state, not an upload.

For CSV specifically, on-device parsing is the difference between "I debugged my customer export" and "I uploaded my customer export." Spreadsheets carry more PII per kilobyte than almost any other format, and a parser that only runs locally removes the question.

What's not supported (yet)

• JSON to CSV (the reverse direction): planned as a separate tool. In the meantime, JSON to YAML covers the common "serialise JSON into a human-readable row format" case.
• Multi-character delimiters: the parser operates on single characters by design, covering CSV, TSV, SSV, and pipe-delimited formats. For || or |~| delimited exports, pre-process with find-and-replace down to a single character.
• Streaming for very large files: documents that fit in browser memory parse fine (tens of megabytes are comfortable). Larger files should go through a CLI tool like csvkit or papaparse in a Node script — partial output in the UI would be worse than a clean "too big" signal.
• Custom column types: the coercion layer has three knobs applied uniformly across every cell. Per-column schema control is on the v1.1 roadmap.
• Dialect auto-detect beyond delimiter: quote character is fixed at " and escape convention is doubled-quote. Single-quoted CSVs and backslash-escape dialects are rare enough we have not hit them in user traces.

Related tools

• JSON Formatter and Validator — the natural next step after conversion. Pretty-print the JSON at a custom indent, minify it, or validate the structure against a live error banner.
• JSON to YAML Converter — if the final destination is a config file or Kubernetes manifest, convert the CSV here then pipe the JSON through YAML.
• JSONPath Evaluator — filter, project, or extract specific fields from the array of records. $[?(@.active==true)].email is a one-liner on the converter's output.
• JSON Diff — convert two CSV snapshots to JSON and diff them structurally, which is almost always more useful than diffing the raw CSV text.
• Sort Lines — pre-process a messy CSV (dedupe, sort alphabetically) before feeding it to the converter when you want canonical ordering.