ascii2svg

Your text diagrams, drawn properly. Every character stays exactly where you put it.

Try it in your browser · Live demo · Scroll demo · Download the Claude skill

_{The playground: fix a model's misaligned diagram, type one live, then a tour of combined and Mermaid-style diagrams.}

You sketch a diagram in plain text, in a README, a code comment, a chat with an AI, or a notes app. It looks right in a monospace font and falls apart everywhere else: Slack, email, Confluence, slides, a phone.

ascii2svg turns it into a crisp SVG that draws itself, follows your reader's light or dark mode, and shows data moving along the arrows. It is still 1:1: every run reads its own SVG back and checks it against your text, cell by cell. If anything moved, it tells you.

Type this…

+------------+     +------------+     +------------+     +-------------+
|    Idea    |---->|   Draft    |---->|   Review   |---->|   Publish   |
+------------+     +------------+     +-----+------+     +-------------+
                         ^                  |
                         |                  |
                         +------------------+
                           needs changes

…get this

python3 scripts/ascii2svg.py workflow.txt -o workflow.svg --color --theme auto --animate

Plain + - | became real lines, v ^ < > became arrowheads, and the words stayed words. It works just as well with Unicode box characters (┌─┐ │ └─┘ ╭╮ ═║ ▶).

Who it's for

If you don't write code: open the playground, paste your diagram, pick where it's going, and download the SVG, PNG or a web page. It runs entirely in your browser (Python compiled to WebAssembly via Pyodide); nothing is uploaded. Or ask Claude. With this repo installed as a skill, say "turn this diagram into an image for my slides" or "make this look nice for our wiki", and it picks the options for where the diagram is going, checks the result, and hands you the file. Nothing to learn.

If you do: it's one Python file with no dependencies. It reads a file, stdin, or --text, and writes an SVG plus a JSON report that an agent or a CI job can check (exit_code, roundtrip, warnings). It never guesses: a character it isn't sure about stays text.

LLM diagrams, fixed

Ask any chat model for an architecture diagram and the boxes rarely line up: an emoji counted as one column instead of two, a wall one space short, a connector that drifts sideways between rows. --repair puts them back:

As the model wrote it	--repair
2 of 5 boxes recognised, 10 warnings	5 boxes, 0 warnings, 5 fixes

ascii2svg diagram.txt -o diagram.svg --repair --json

{"status": "ok", "summary": "Repaired 5 misalignments. Rendered 5 boxes and 3 arrows (22x51); 1:1 self-check exact.",
 "repair": {"edits": [{"line": 2, "col": 23, "fix": "moved the right wall '│' 1 col left"},
                      {"line": 6, "col": 12, "fix": "moved '│' 1 col left to line it up"}, "..."],
            "text": "┌─────────────────────┐\n│  📱 Mobile app      │\n..."}}

• What it fixes: ragged right walls (the top corner, every row and the bottom corner vote on where the wall belongs), drifted left walls, edges one character too long or short, connector pieces and arrowheads one or two columns off, arrowheads that stop a cell short, and gaps of one or two cells before a box.
• What it never does: change your words. Only line characters and spaces move or appear, and only into empty cells. A test strips every line character from each row, before and after, and checks that what remains is identical.
• What you get back: every edit with its line and column, and the corrected text in repair.text, ready to paste back where the diagram came from. The 1:1 guarantee then holds against that text. Diagrams that were already right come back untouched.

In the playground, a diagram with warnings shows a Fix alignment button. It rewrites the editor text, so you see exactly what moved.

Beyond boxes

Flows that branch and merge, tables with callouts, swimlanes: all plain text, all 1:1.

Diamonds and fan-outs drawn with / \, and charts drawn with block characters, come out as real shapes: diagonals run corner to corner, and █▓▒░ become solid rectangles with no seams between cells.

The sources are in docs/examples/, and all of them are in the playground's example picker.

Pick a look

Same diagram, same guarantee, five ways:

Default	Colour	Dark	Print	Alive
(no options)	--color	--theme dark --color	--style flat --square	--animate --color --theme auto

The animation is careful about where it runs:

• It never changes what's drawn. An animated file is the static file plus timing. When the animation ends, you're looking at exactly what the self-check verified. Tools that don't play animation (PNG export, most editors) show the finished drawing.
• It respects the reader. With reduce motion turned on in the OS, nothing moves.
• draw and flow need no JavaScript. They use CSS and SVG <animate> only, so they play inside a plain <img> tag, which is how GitHub, Notion and most docs sites show images.
• Hover a box in a browser and it lights up. That helps when you're presenting.

Tall diagrams: reveal as you scroll

A 70-row architecture diagram drawn on a timer finishes before anyone scrolls down to it. With --animate scroll, the diagram unfolds with the reader instead:

python3 scripts/ascii2svg.py architecture.txt -o architecture.html --color --theme auto --animate scroll

• Boxes and labels fade in as they reach the lower edge of the screen.
• Long vertical connectors are drawn by the scroll: they grow downward in accent blue as you read, then settle to ink.
• Each flow pulse starts once its whole route is on screen.
• With reduce motion on, the page shows the finished diagram.

Why a web page, not just an SVG? An SVG shown as an image (which is how GitHub READMEs, Notion and most docs sites show them) can't see the page it sits in, so it can't know how far you've scrolled. CSS scroll timelines don't drive SVG shapes either: we tested it, and Chromium leaves them inactive. So scroll writes an .html page with the SVG inline and about 2 KB of plain JavaScript (IntersectionObserver and CSS animations, no libraries). It uses only standard browser features and works offline, though so far it has been tested in Chromium only. --animate scroll with a plain .svg output is refused with an explanation. The SVG inside the page is the same self-checked drawing; the script only decides when each part appears.

Try it: open the scroll demo and scroll. It's docs/architecture.html, served by GitHub Pages.

Where is it going?

A bigger one

The flows follow the real wiring: through junctions, down to both branches of a fan-out, and across the edge of a container.

Install

Python 3.9+ and nothing else:

pip install ascii2svg                 # adds the `ascii2svg` command and the `ascii2svg` module
pip install "ascii2svg[width,png]"    # optional: wcwidth (character widths) + cairosvg (--png)

It is a single file with no dependencies, so you can also just copy scripts/ascii2svg.py into your project.

As a Python library:

import ascii2svg

svg, report = ascii2svg.render(open("diagram.txt").read(), color=True, theme="auto", animate="flow")
assert report["roundtrip"] == "exact"          # the 1:1 self-check passed
open("diagram.svg", "w", encoding="utf-8").write(svg)

page, report = ascii2svg.render(text, color=True, animate="scroll", html=True)   # a scroll-reveal page

render() takes the same options as the CLI (style, square, theme, color, animate, html, title, tab_size) and returns the markup plus the same report as --json. It raises ValueError for empty or oversized input and for unknown options.

As a Claude skill: download ascii2svg.skill from the latest release (or build it: python3 tools/package_skill.py → dist/ascii2svg.skill), then:

• claude.ai / Claude desktop: Customize → Skills → upload ascii2svg.skill.
• Claude Code: unzip it into ~/.claude/skills/ (you get ~/.claude/skills/ascii2svg/).

SKILL.md tells Claude when to use it, which look fits which destination (including scroll pages for tall diagrams), and how to read the report before handing you the file.

Usage

ascii2svg [INPUT ...] [-o OUT.svg|OUT.html|DIR/] [--preset NAME] [--json] [--check] [--describe] [--brief]
          [--color] [--theme light|dark|auto] [--animate [draw|flow|scroll]] [--html]
          [--style glow|shadow|flat] [--square] [--accent #HEX] [--font NAME] [--width PX]
          [--all-blocks] [--block N] [--unescape] [--strict] [--schema] [--mcp]
          [--png [PATH]] [--strict] [--text TEXT] [--tab-size N] [--title TEXT]

stdout is always exactly one thing: the SVG (or the page, with --html), or (with --json) the report. A one-line summary goes to stderr.

What gets drawn

When unsure, a character stays text, so the worst case is the character itself in its own cell, never a wrong shape. a->b, --dry-run, user_id, C:\temp, yes/no, TCP/IP, markdown tables and |-- file trees all stay exactly as written.

How 1:1 is guaranteed

Every run parses the SVG it just wrote (lines, curves, diagonals, arrowheads, block rectangles, text positions) back into a character grid and compares it with the input, cell by cell. ASCII characters may only appear as the line they stand for (-→─, |→│, +→corner or junction, v→▼, /→╱). Any mismatch is exit code 2. This covers every look, including animated ones and the still frame used for PNGs.

Built for agents

Every outcome is machine-readable, nothing hangs, and every problem comes with a fix.

The loop: draft the diagram, check it, fix what the hints say, then render.

ascii2svg draft.txt --check --describe      # validate + structure; writes nothing
ascii2svg draft.txt -o out.svg --preset readme --json --brief

{"status": "warnings",
 "summary": "Rendered 2 boxes and 0 arrows (8x8); 1:1 self-check exact. 1 warning, first: row 4 col 4 (dangling_line): the arrowhead 'v' at row 5 col 5 is one column right; move one of them so they line up",
 "exit_code": 0, "...": "..."}

• One JSON object on stdout for every outcome with --json (or --check), including usage errors: {"status": "usage_error", "hint": "'--colour': did you mean --color?"}. status and summary come first, so a truncated report still says what happened. The JSON is ASCII-safe, so any harness on any OS decodes it the same way.
• Warnings you can act on: each has a stable code, a 1-based row/col, the source line, and a hint. When a line misses its partner by one cell, the hint says where it is.
• Silently wrong input is caught: one-line input with literal \n (escaped_newlines; fix it with --unescape), and ASCII boxes that never close (unclosed_box: the hint names the broken wall, e.g. "its left wall at line 5 col 23 is '<'").
• No false alarms on well-formed diagrams: a line may end at a label, meet another line side-on (sequence messages │───▶│, drawn touching), or stop in open space (axis ticks, lifeline ends). A warning means there is evidence of a mistake: a partner one cell off, a line that stops just short of a box, a junction with a missing arm, or a box that never closes.
• --describe adds what the diagram says: boxes (name, title, text, position, parent) and edges (Gateway → Orders), so you can check that the picture means what you intended.
• --preset readme|slides|chat|print|dark|page picks the look from the destination. Explicit flags still win (--preset readme --no-color).
• --brief never embeds the markup; without it, --json without -o returns the markup inline, which can run to 150 KB for a large diagram.
• No hangs: if you forget the input and stdin stays silent, it fails after 5 seconds with bad_input. Pass - to wait for stdin on purpose.
• --schema prints every option (kind, choices, default), preset, exit code, warning code and report field as JSON: enough to build a correct tool definition without reading docs.

Report fields

Exit codes

Many diagrams at once

ascii2svg README.md --all-blocks -o diagrams/ --preset readme   # every diagram in the file
ascii2svg docs/*.txt -o out/ --check                            # several files, one report each
ascii2svg README.md --block 3 -o flow.svg                       # just the third code block

Code blocks without lines or boxes (your bash and python examples) are skipped and listed under skipped. Output files are named after the input (README-2.svg for its second code block). The report holds one entry per diagram under diagrams, and its top-level status, summary and exit_code cover the whole batch. Every warning also carries source_line/source_col: its position in the file you passed, not in the extracted block, and the hint uses the same numbers.

As an MCP server

claude mcp add ascii2svg -- ascii2svg --mcp          # Claude Code

{"mcpServers": {"ascii2svg": {"command": "ascii2svg", "args": ["--mcp"]}}}

The second form works for Claude Desktop, Cursor, and any other MCP client. Two tools, no dependencies:

• check_diagram validates a diagram and returns the report, with boxes and edges; it writes nothing.
• render_diagram writes output_path (.svg, or .html for a page) with any preset or style option.

The diagram travels as a JSON string, so the escaped-newline problem can't happen. Tested against the official MCP Python SDK.

As a tool in any agent framework

Pass the diagram on stdin (or as a file), never through --text: escaped newlines are the most common way agents break diagrams.

import json, subprocess

def render_ascii_diagram(diagram, output_path, preset="readme", describe=False):
    args = ["ascii2svg", "-", "-o", output_path, "--preset", preset, "--json", "--brief"]
    p = subprocess.run(args + (["--describe"] if describe else []), input=diagram.encode(), capture_output=True)
    return json.loads(p.stdout)          # always one JSON object, including on errors

Or skip the subprocess and call the library: svg, report = ascii2svg.render(diagram, preset="readme").

Tests

python3 tests/test_ascii2svg.py        # or: python3 -m pytest tests
python3 docs/build.py                  # regenerate every image in this README

54 tests over 23 test diagrams:

• round-trip in every style and every look
• animation that ends on the static drawing and respects reduced motion
• flow routes that start at the right box
• scroll reveal driven by the page, with no script inside the SVG
• colour groups
• the ASCII traps
• glow never behind a label
• planted faults that must be caught
• byte-identical repeat runs
• width fallback vs wcwidth
• CLI behaviour, including UTF-8 output on Windows consoles, and the render() library API
• the agent contract: JSON usage errors (never exit 2), status and summary, near-miss hints, escaped-newline and broken-box detection, --describe edges, presets, --schema, and no hang on a forgotten input
• every diagram in a markdown file, several inputs, positions mapped back to the file, style options, and the MCP server's protocol
• the browser playground runs exactly this module, and every playground example round-trips
• no false alarms on sequence diagrams, timelines, charts with ticks and dashed boundaries, while real mistakes (a gap before a box, a broken wall) still warn; touching lines are drawn touching
• --repair on LLM-style misalignment: every fixture repaired to ok, text provably unchanged, well-formed diagrams untouched
• block elements drawn as exact rectangles (no glyphs, runs merged), and / \ drawn only as runs, never inside words like yes/no or C:\Users

They pass with and without the optional packages.

License

MIT. Use it, change it and ship it, commercially too; keep the copyright notice.

Not yet

• Repairing bigger misalignments. --repair fixes pieces up to two cells off and walls up to three; anything further is left alone and warned, since guessing would risk a wrong picture.
• ASCII rounded corners (.-'). These stay text.
• Free-floating ASCII connectors that touch no box (e.g. A ---> B between plain words). These stay text.
• A JS/npm port.
• Every renderer checked. The original static look was verified in cairo and resvg. The new looks (colour, themes, animation) have been checked in Chromium only so far, not yet in Firefox, Safari, or through cairo for --png.