libmmd

A fast Mermaid renderer written in C (-std=gnu2y). It turns Mermaid source into SVG. It works as a command-line tool (mmd) and as a small C library (libmmd). It needs no browser, no Node.js, no JavaScript engine and no dependencies beyond libc and libm.

• About 0.6–1.6 ms per diagram from the CLI. mermaid-cli takes about 0.9–1.6 s for the same diagrams (see Benchmarks).
• Library renders take about 15 µs for a small flowchart and about 0.5 ms for a 4 KB one.
• Flowchart layout follows dagre's algorithm, so node positions match mermaid's closely (mean node-centre error 1.6 px on the fixture set).
• All 396 syntax examples from mermaid's documentation that mermaid itself can render also render here.

Status: this is a young project. Output is meant to look like mermaid-cli 11.17, not to match it pixel for pixel. See Status and limitations.

Supported diagram types

Family	Keywords
Flowchart	flowchart, graph (all directions, subgraphs, per-subgraph direction, every node shape, markdown strings, fa: icons, edge ids and animation, classDef/style/linkStyle)
Sequence	sequenceDiagram (actors and participant types, boxes, activations, notes, loop/alt/opt/par/critical/break blocks, autonumber)
Class	classDiagram, classDiagram-v2
State	stateDiagram, stateDiagram-v2
Entity relationship	erDiagram
Gantt	gantt
Pie	pie
User journey	journey
Timeline	timeline
Mindmap	mindmap
Quadrant	quadrantChart
XY chart	xychart, xychart-beta
Git graph	gitGraph
C4	C4Context, C4Container, C4Component, C4Dynamic, C4Deployment
Requirement	requirementDiagram
Architecture	architecture, architecture-beta
Block	block, block-beta
Sankey	sankey, sankey-beta
Packet	packet, packet-beta
Kanban	kanban
Radar	radar, radar-beta
Treemap	treemap, treemap-beta
ZenUML	zenuml (native renderer in ZenUML's style)
Ishikawa (fishbone)	ishikawa, ishikawa-beta
Cynefin	cynefin, cynefin-beta
Venn	venn, venn-beta
Tree view	treeView, treeView-beta
Swimlane	swimlane-beta
Event modeling	eventmodeling
Railroad	railroad-beta, railroad-ebnf-beta, railroad-abnf-beta, railroad-peg-beta
Wardley map	wardley-beta

Every type also supports:

• the default, dark, forest and neutral themes;
• %%{init}%% and YAML front-matter config, with themeVariables values whitelisted;
• accTitle/accDescr, which become <title>/<desc> plus ARIA attributes.

Per-type feature notes are in docs/dev/INTERNALS.md.

Build

You need a C compiler with C23 support:

• clang (the default; tested with clang 23);
• or gcc 16.

The Makefile builds with -std=gnu2y -fms-extensions. You also need GNU make. Python 3 is needed only for the test and benchmark scripts.

make                      # -> build/mmd, build/libmmd.{a,so}, build/mmd.pc, build/bench
make CC=gcc-16            # build with gcc instead
make MARCH=native LTO=1   # optional: tuned build
make pgo                  # optional: profile-guided build in build-pgo/
make install PREFIX=/usr/local   # header, static + shared lib, pkg-config file, mmd

The build has zero warnings under -Wall -Wextra with both gcc 16 and clang.

CLI

build/mmd in.mmd -o out.svg
build/mmd -i in.mmd -o out.svg -t dark      # themes: default | dark | forest | neutral
cat in.mmd | build/mmd > out.svg            # stdin -> stdout
build/mmd in.mmd -o out.svg --no-avoid      # flowchart: skip obstacle re-routing (fastest)

On a syntax error the tool exits with a non-zero status and prints line:col: message.

C library

The API is in include/mmd.h:

#include <stdio.h>
#include <string.h>
#include "mmd.h"

int main(void) {
    const char *src = "flowchart LR\n  A[Start] --> B{Ok?}\n  B -- yes --> C[Done]\n";
    mmd_buffer svg = {};
    mmd_error err = {};
    if (mmd_render_svg(strlen(src), src, nullptr, &svg, &err) != MMD_OK) {
        fprintf(stderr, "%u:%u: %s\n", err.line, err.col, err.message);
        return 1;
    }
    fwrite(svg.data, 1, svg.len, stdout);
    mmd_buffer_free(&svg);
}

cc -std=gnu23 -Iinclude example.c build/libmmd.a -lm      # or: $(pkg-config --cflags --libs mmd)

• mmd_options (which may be nullptr) sets node and rank spacing, font size, theme, and the MMD_F_* layout flags.
• For batch rendering, mmd_context_new() / mmd_context_render_svg() / mmd_context_free() reuse the arena and output buffer between calls. Use one context per thread.
• Errors come back as mmd_status together with mmd_error (code, 1-based line and column, message).
• The shared library exports only mmd_* symbols.

Benchmarks

All numbers below were measured on the same 8-vCPU Intel Xeon VM. Times are wall-clock for a whole process run, including process start-up. Each value is the best of N runs: N = 20 for libmmd, 5 for mmdr and 3 for mmdc.

The tools compared are:

• mmd: this project;
• mermaid-rs-renderer (mmdr 0.3.1, Rust);
• mermaid-cli (mmdc 11.17.0, headless Chrome).

The run was made on 2026-10-23.

In brackets: how many times slower than libmmd. Data: bench/cli_times.csv; python3 tools/bench_chart.py regenerates the chart and this table.

input	libmmd	mmdr 0.3.1	mermaid-cli 11.17
flowchart_small (mmdr fixture)	0.59 ms	29.1 ms (49×)	917 ms (1554×)
flowchart_medium (mmdr fixture)	0.78 ms	137 ms (175×)	1064 ms (1364×)
flowchart_large (mmdr fixture)	1.56 ms	437 ms (280×)	1583 ms (1015×)
flowchart_mega_event_mesh (mmdr fixture)	0.96 ms	1777 ms (1851×)	1017 ms (1059×)
seq_full (tests/cases)	0.59 ms	3.15 ms (5.3×)	860 ms (1458×)
class_full (tests/cases)	0.72 ms	9.64 ms (13×)	1124 ms (1561×)

Rendering in-process through the library, timed with build/bench (minimum over repeated renders, clang build, from bench/results_v14.txt and bench/results_libmmd_inproc.txt):

input	size in → SVG out	time
bench/inputs/small.mmd	376 B → 9 KB	14.9 µs
bench/inputs/medium.mmd	4.4 KB → 80 KB	527 µs
bench/mmdr_fixtures/flowchart_large.mmd	6.4 KB → 75 KB	214 µs
bench/inputs/large.mmd	44 KB → 1.2 MB	9.3 ms
bench/inputs/xlarge.mmd	143 KB → 4.7 MB	56 ms

bench/results_fixtures.txt has library and CLI times against mmdr for all 31 mmdr flowchart fixtures. To reproduce:

make && build/bench bench/inputs/medium.mmd 30
MMDR=/path/to/mmdr python3 bench/compare.py bench/mmdr_fixtures/*.mmd

mermaid-cli spends most of its time starting Chrome. Its per-diagram cost is lower when it renders many diagrams in one process. The comparison is still useful for the common case of one CLI call per diagram.

Gallery: libmmd vs mermaid-cli

Each image puts libmmd on the left and mermaid-cli 11.17 on the right, rendered from the same source in tests/cases/. To regenerate them, run python3 tools/gallery.py tests/cases/<name>.mmd.

Flowchart with subgraphs	Flowchart
Sequence	Class
State	ER
Gantt	Git graph
Mindmap	Architecture
C4	Sankey
Pie	Timeline
XY chart	Quadrant
Journey

More before/after pairs from development are in samples/sbs/.

Testing and quality tooling

what	how
Golden + sanity tests (151 cases; XML well-formedness, error positions, viewBox bounds, mermaid docs corpus)	make test (UPDATE_GOLDEN=1 make test to refresh)
Mermaid docs corpus: 432 examples from mermaid's syntax docs, 396/396 that mermaid itself renders	python3 tools/docs_audit.py [-v]
ASan + UBSan build	make asan → build-asan/mmd
Fuzzing (mutational, under ASan)	python3 tests/fuzz.py build-asan/mmd 1000
Allocation-failure injection	sh tests/poison.sh
Stress generator (every type, density 1×–5×)	python3 tools/stress_gen.py OUT 2 SEED SCALE
Visual quality scanner (edges through nodes, label overlaps, labels on edges, out-of-bounds…)	python3 tools/quality_scan.py files… [-v]
Layout agreement with mermaid (rank/cross-axis order, node-centre error)	python3 bench/posagree.py
Side-by-side ranking against mermaid-cli renders	python3 tools/sbs_rank.py, tools/sbs_pair.sh
Byte-identity check for perf work	sh bench/eq/cmp.sh

Status and limitations

• Similar, not identical. Output follows mermaid-cli 11.17 in style and, for flowcharts, in layout (mermaid/dagre ordering and ranking). Mermaid-generated CSS selectors and pixel-exact SVGs are not goals.
• Mindmap does not use mermaid's cose-bilkent layout. It uses its own radial and spring layout, so mindmaps look different from mermaid's.
• Architecture also has its own grid layout, not fcose.
• Text is measured with built-in Helvetica/Arial-compatible advance tables, not a real font engine. Unusual fonts can be sized imperfectly. CJK ideographs, kana and fullwidth forms are 1em (as in Noto Sans CJK, YaHei and PingFang); halfwidth forms are half an em; combining marks have no advance of their own; other letters use a Latin average.
• Dense charts. Sankey, quadrant and Wardley move labels so they do not sit on each other. Wardley draws a short leader when a label has to sit farther from its node, and a white halo when it still crosses a link. Labels with user-given coordinates stay where the author put them, so those can still overlap.
• Known differences from mermaid include gantt tick density, state concurrent-region styling and journey legend order. See docs/dev/INTERNALS.md.
• fa: icons in flowcharts need the Font Awesome font where the SVG is viewed. Architecture and mindmap icons are embedded as paths.
• Not supported: interactivity (click callbacks, tooltips) and PNG/PDF output. To get a raster image, convert the SVG with any SVG rasteriser.

License

MIT. See LICENSE.

Embedded icon data comes from Font Awesome Free (CC BY 4.0) and simple-icons (CC0 1.0). See THIRD_PARTY_NOTICES.

---

中文简介

libmmd 是一个用 C 语言（gnu2y）编写的高速 Mermaid 渲染器，可作为命令行工具 mmd 或 C 库 libmmd 使用。它把 Mermaid 源码直接渲染成 SVG，不需要浏览器、Node.js 或任何第三方依赖。

• 支持上表列出的全部 30 多种图表类型。
• 单个图表的命令行渲染约 1 毫秒，mermaid-cli 约 1 秒。
• 流程图采用与 mermaid（dagre）一致的布局算法。

构建只需 make，用法为 build/mmd in.mmd -o out.svg。效果对比见上方 Gallery。

许可证：MIT（见 LICENSE）。内嵌图标数据来自 Font Awesome Free（CC BY 4.0）和 simple-icons（CC0 1.0），详见 THIRD_PARTY_NOTICES。