crcglot 0.14.0

Verified CRC source-code for C, C#, Go, Python, Rust, TypeScript, Verilog, and VHDL — catalogue-driven, typed introspection API, self-test embedded.

crcglot

Verified CRC source code for C / C++ ⚙️, Rust 🦀, Go 🚦, C# 💠, Java ☕, Python 🐍, TypeScript 🔷, Verilog 🔧, and VHDL 🔌. Catalogue-driven, execution-verified, self-test embedded, multi-language by design. Pure-stdlib package — zero runtime dependencies.

LLMs will gladly write you CRC code. It might even be right. crcglot doesn't ask you to trust the generator — it proves the output by running it: every algorithm, in every variant, in every language, is generated, compiled, and executed against the hardcoded canonical reveng catalogue vector (crc("123456789") == <check value>). More than 70 algorithms across nine languages, verified by execution rather than inspection — and every generated file ships that same self-test so you can re-prove it on your own toolchain.

Quick start

uv tool install crcglot         # or: pip install crcglot
crcglot c crc32 file=mycrc

That's it. You now have mycrc.h and mycrc.c — drop-in CRC-32 with a built-in _self_test() you can call to verify it matches the canonical reveng check value.

The whole model is three choices: which algorithm (crc32, crc16-modbus, … — crcglot list shows the more than 70), which language (c / python / rust / vhdl / verilog / go / csharp / java / typescript), and whether you want it --small (smallest code, the default) or --fast (fastest the target supports). crcglot figures out the implementation details — you never have to know what "slice-by-8" is.

crcglot rust crc32 --fast file=mycrc     # fastest Rust crc32 to file mycrc.rs
crcglot c crc8 --small                   # smallest C crc8, to stdout

Installation

Tool	Command	Use when
uv (recommended)	uv tool install crcglot	You just want the crcglot CLI on PATH. Isolated install, no global pollution.
uv (as a library)	uv add crcglot	You're calling the generators from Python (e.g. a build script that emits CRC code into your repo).
pip	pip install crcglot	You don't have uv. Identical package, slower install.
pipx	pipx install crcglot	Same isolation story as uv tool, if pipx is what you have.

Python 3.11+, no other runtime dependencies — crcglot itself is pure stdlib. Per-target toolchains (gcc, rustc, tsx, iverilog, etc.) only matter if you want to run the generated code; the generator produces source either way.

Or use it from Python code:

from crcglot import LANGUAGES
header, source = LANGUAGES["c"].generator("crc32")

Both surfaces are documented in detail below.

What you get per language

Function	Purpose
<fname>_init / _update / _finalize	Streaming triple — feed data chunk by chunk
<fname>	One-shot wrapper that calls the streaming triple
<fname>_self_test	Verify against the reveng check value on your toolchain

Every target ships a runtime-callable _self_test(): C returns 0/1; Rust / Go / C# / Java / TypeScript / Python / Verilog / VHDL return bool / boolean / bit. No #[cfg(test)] gating — call it from your release build, a boot self-check, or a startup assertion.

How it's verified

The guarantee is behavioral, not structural — crcglot doesn't lint the generated code, it runs it. Three axes, fully crossed: every one of the more than 70 algorithms, in every variant the target supports (bit-by-bit, table-driven, slice-by-8), in every one of the nine languages, is executed and its output checked against the hardcoded canonical vector. Nothing ships on "the generator looks correct."

CI runs the Python-level suite on every push: every algorithm in the reveng catalogue is checked against its hardcoded canonical check value — not the catalogue's own check field, so a silent regression in the engine can't hide — and the Python generator is run end-to-end (generated, exec'd, and called on b"123456789") against the same hardcoded vectors. The slow tier on top of that compiles and executes the generated source for every algorithm in C, Rust, Go, C#, Java, TypeScript, Verilog, and VHDL via gcc / rustc / go / dotnet / javac+java / tsx (Node) / iverilog / ghdl and re-checks the runtime result — same algorithm coverage, exercised through each real toolchain.

Every generated file also ships its own _self_test() carrying that same canonical vector. For every target except Python, you should call _self_test() once in your build environment — wire it into a unit test, a startup assertion, or your boot self-check. Our CI proves the generator emits correct code on our reference toolchain; only running _self_test() on yours proves your compiler version, optimization flags, target endianness, and integer widths haven't introduced a subtle disagreement. Python is the exception: the interpreter that ran the CI suite is the one running your code, so the in-environment check would be redundant.

Documentation comments

The generated code isn't just correct — it's documented. Every file gets a header (algorithm parameters, a copy-paste streaming example, the self-test contract) and a doc comment above each of the five functions, so a reader learns the init → update* → finalize streaming contract from the source, not from the tests. Pick the convention with --comment=<style>; plain (clean human-readable comments in each language's native syntax) is the default, and every language also has its idiomatic doc-tool style:

Language	--comment styles
C / C++ ⚙️	plain, doxygen
C# 💠	plain, doxygen, docfx (XML /// <summary>)
Java ☕	plain, doxygen, javadoc
Python 🐍	plain, google, numpy, rest (Sphinx :param:)
Rust 🦀	plain, rustdoc (/// + # Arguments)
Go 🚦	plain, godoc
TypeScript 🔷	plain, jsdoc (TSDoc)
Verilog 🔧 / VHDL 🔌	plain

crcglot c crc32 --comment=doxygen        # /** @brief @param @return */
crcglot python crc32 --comment=numpy     # numpydoc underlined Parameters / Returns
crcglot rust crc32 --comment=rustdoc     # /// with # Arguments markdown

crcglot offers each language only the styles its doc-tool actually understands — crcglot rust --comment=doxygen is rejected, because Doxygen doesn't read Rust. The matrix is derived from the styles themselves; nothing hardcodes it.

Building a UI? The matrix is queryable, so a front end can populate a language → style dropdown with no hardcoding. Each record carries a machine name (the dropdown value, handed back to the generator), a human label, and a description:

from crcglot.comments import comment_styles_for_language
for s in comment_styles_for_language("python"):
    print(s.name, "|", s.label, "|", s.description)
# plain  | Plain            | Human-readable comments in the language's native syntax
# google | Google           | Google-style docstrings (Args / Returns / Note)
# numpy  | NumPy            | NumPy (numpydoc) docstrings, underlined Parameters / Returns
# rest   | reStructuredText | Sphinx field-list docstrings (:param: / :returns:)

The generators take the chosen name directly — LANGUAGES["python"].generator("crc32", comment_style="numpy") — and the same {name, label, description} records are served over MCP in the crcglot://languages.json resource (each language's comment_styles).

Why generate the docs instead of asking an LLM?

You still can — point an LLM at the output and let it write whatever prose you like; nothing here stops you. But the generated code is fully known: the parameters, the API contract, and the streaming semantics are deterministic facts, so the documentation can be deterministic too. That buys three things an LLM pass can't:

• Reproducible. The same request produces the same comment, byte for byte. Everyone who generates crc32 gets the identical documentation — no drift, no "it phrased it differently this time," no diff churn between two runs.
• Correct by construction — or wrong in exactly one place. The comment is rendered from the same source of truth as the code, so it can't hallucinate a parameter or misdescribe the API. And if a description is wrong, it's wrong uniformly — caught once, fixed once in the generator, and the fix reaches every output everywhere. Per-invocation LLM wrongness is the opposite: subtly different each time, and far harder to audit.
• Free, offline, auditable. No API call, no token cost, no network — it runs in CI and on an air-gapped build. A reviewer (the class-III-medical-device kind) audits the comment generator once and can then trust every file it emits.

Layer an LLM on top when you want richer prose. The point is that the baseline everyone ships by default is deterministic, uniform, and reviewable.

CLI reference

crcglot <command> [options...]

crcglot list [GLOB] [--json]

Browse the catalogue. Optional GLOB filters by shell-style pattern (e.g. crc16-*). Exit code 1 if nothing matches.

crcglot list                # more than 70 algorithms
crcglot list 'crc32-*'      # just the CRC-32 family
crcglot list --json         # machine-readable list with full parameters

crcglot info <name>

Print parameters (width, poly, init, refin, refout, xorout, check, desc) for one algorithm. Exit 1 on unknown name.

crcglot info crc64-xz

crcglot detect [INPUTS...]

Brute-force identify which catalogue CRC matches a packet whose tail is the CRC. Useful for reverse-engineering unfamiliar protocols, debugging captured frames, or confirming a sample really uses the CRC you think it does.

crcglot detect packet.bin                            # binary file (or '-' for stdin)
crcglot detect a.bin b.bin c.bin                     # multi-packet (intersected)
crcglot detect --text "123456789 cbf43926"           # text mode, inline
crcglot detect --text -                              # text mode, one packet per line on stdin
crcglot detect --hex "313233343536373839cbf43926"    # hex-encoded bytes
crcglot detect --algorithms 'crc16-*' packet.bin     # narrow the scan to a family
crcglot detect --match all packet.bin                # forensic: every candidate
crcglot detect --match set a.bin b.bin               # strict: succeed only on a single algorithm

--match selects the strategy: first (default — early-stop on the first hit, priority order is crc32, crc32-jamcrc, crc32-iscsi, then the rest of the catalogue), all (exhaustive forensic view), set (strict singleton: succeed only if exactly one algorithm survives across all packets). Exit 0 on match, 1 otherwise. For text packets the inferred separator + hex leader + case are reported so you can reproduce the same format via crcglot encode.

crcglot encode <algorithm> [<data>]

Build a packet by appending the CRC. Round-trip partner to detect — feed detect's (algorithm, endianness, padding) shape back to encode to rebuild a packet in the same format.

crcglot encode crc32 "123456789"                                # → "123456789 cbf43926"
crcglot encode crc32 "123456789" --sep $'\t' --leader 0x --upper # tab + "0x" + uppercase
crcglot encode crc32 --binary < data.bin > packet.bin           # binary, big-endian
crcglot encode crc32-iscsi --binary --little < data.bin         # binary, little-endian

Option	Default	Effect
--binary	off	Read stdin as bytes; write packet bytes to stdout.
--little	off	Little-endian CRC byte order (default: big).
--sep STR	" "	Text separator between data and hex.
--leader STR	""	Text hex leader: "", "0x", or "0X".
--upper	off	Uppercase hex digits.
--fmt STR	"{data}{sep}{leader}{crc}"	str.format template; the four tokens may be reordered.

crcglot credits

Print acknowledgments for the upstream work crcglot stands on (also exported as crcglot.ATTRIBUTION / crcglot.ACKNOWLEDGMENTS). See Acknowledgments.

crcglot {c | csharp | go | java | python | rust | typescript | verilog | vhdl} <algorithm> [<algorithm>...] [options...] [tokens...]

Generate source code for the chosen target language. Pick your intent — crcglot picks the implementation:

Option / token	Effect
--small	Smallest code, zero RAM table (bit-by-bit). The default — works for any width.
--fast	Fastest the target supports: slice-by-8 for width 32/64 on compiled targets, table-driven otherwise.
--custom	Use raw Rocksoft/Williams params instead of a catalogue lookup (see below).
--comment=STYLE	Documentation style for the generated comments (default plain). See Documentation comments.
file=STEM	Write to disk (extension picked per language; see below). Omit for stdout.
symbol=NAME	Override the emitted function name. Default: derived from algorithm, or from file=STEM if given.

File extensions per language: C emits STEM.h + STEM.c; Python .py; Rust .rs; VHDL .vhd; Verilog .sv (SystemVerilog 2012); Go .go; C# .cs; Java .java; TypeScript .ts. (For Java, every algorithm shares one container class named after STEM, so the stem must be a valid Java identifier.)

Bundle several algorithms into one file by naming more than one — crcglot c crc32 crc16-modbus crc8 file=mycrcs writes a single mycrcs.h / mycrcs.c containing all three (one .go / .rs / .cs / … for the other languages). Each algorithm keeps its own catalogue-derived function names (crc32, crc16_modbus, …) and the tables are namespaced per symbol, so they never collide. symbol= is rejected with more than one algorithm (it names a single function); duplicates are de-duplicated; an unknown name aborts the whole bundle.

Expert overrides (you usually don't need these — --fast chooses for you): --table forces the 256-entry single-table form, and --slice8 forces the 8-table form. They exist for the rare case where you want the middle of the size/speed curve explicitly — e.g. a RAM-constrained target where the 1 KiB table is fine but slice-by-8's 8 KiB isn't. --slice8 is CRC-32/64 + compiled targets only.

Rules:

• The variant selectors --small / --fast / --table / --slice8 are mutually exclusive — pick at most one (exit 2 otherwise). No selector = --small.
• --slice8 python silently falls back to --table (CPython's per-int overhead eats the slice-by-8 speedup; stderr warns). --fast never needs this fallback — it only picks slice-by-8 where it actually applies.
• Without file=, output goes to stdout. For C, header is emitted first, then source.
• C / Rust / VHDL files embed <symbol>_self_test() returning 0 on success. In constrained embedded targets, standard toolchain flags (-Wl,--gc-sections for C, LTO for Rust) strip whatever you don't call.

--custom (raw Rocksoft/Williams parameters)

For algorithms not in the catalogue:

crcglot c --custom width=16 poly=0x1234 init=0xFFFF \
         refin=true refout=true xorout=0x0000 file=mycustom

Param	Required	Notes
width=N	yes	8, 16, 32, or 64 only
poly=X	yes	Hex (0x...) or decimal
init=X	no	Default 0. Hex or decimal.
refin=B	no	Default false. Accepts true/false/1/0/yes/no/on/off.
refout=B	no	Default false. Same boolean syntax.
xorout=X	no	Default 0.
name=NAME	no	Default crc_custom. Used in generated comments.
desc=TEXT	no	Free-form description in comments.

The check value for the custom parameters is computed automatically (generic_crc(b"123456789", ...)) and embedded into the generated _self_test().

Catalogue

More than 70 algorithms covering everything from CRC-8 (ATM, AUTOSAR, Bluetooth, Maxim 1-Wire) through CRC-16 (Modbus, XMODEM, CCITT, IBM SDLC) through CRC-32 (Ethernet, bzip2, iSCSI, AUTOSAR) to CRC-64 (XZ, ECMA-182, NVMe, Redis). Browse with crcglot list.

Programmatic API

Two registries, both keyed by short code:

LANGUAGES — supported target languages

from crcglot import LANGUAGES

for code, info in LANGUAGES.items():
    print(f"{info.emoji} {info.display_name:<10}  {info.extensions}  "
          f"{sorted(info.variants)}")
    # → ⚙️ C / C++       ('.h', '.c')  ['bitwise', 'slice8', 'table']
    # → 💠 C#            ('.cs',)      ['bitwise', 'slice8', 'table']
    # → 🚦 Go            ('.go',)      ['bitwise', 'slice8', 'table']
    # → ☕ Java          ('.java',)    ['bitwise', 'slice8', 'table']
    # → 🐍 Python        ('.py',)      ['bitwise', 'table']
    # → 🦀 Rust          ('.rs',)      ['bitwise', 'slice8', 'table']
    # → 🔷 TypeScript    ('.ts',)      ['bitwise', 'slice8', 'table']
    # → 🔧 Verilog       ('.sv',)      ['bitwise']
    # → 🔌 VHDL          ('.vhd',)     ['bitwise']

Each entry is a frozen LanguageInfo dataclass with:

• code — dispatch key ("c", "csharp", ..., "typescript", "verilog")
• extensions — file extension tuple ((".h", ".c") for C; single-element for the rest)
• variants — subset of {"bitwise", "table", "slice8"} that the generator accepts
• generator(name, ...) — name-lookup callable (returns source string, or (header, source) tuple for C)
• generator_from_entry(name, algo, ...) — bypass the catalogue with a custom AlgorithmInfo
• combiner(outputs, stem) — merge several generator outputs into one file (powers multi-algorithm bundling); per-symbol tables keep the merge collision-free
• emoji — single-grapheme pictographic identifier for terminals / docs
• display_name — human-readable name (e.g. "C / C++", "TypeScript") — distinct from code

ALGORITHMS — the reveng CRC catalogue

from crcglot import ALGORITHMS

modbus = ALGORITHMS["crc16-modbus"]
print(modbus.width, hex(modbus.check), modbus.desc)
# → 16 0x4b37 Modbus RTU serial protocol

# Filter to CRC-32 only.
crc32_family = [a for a in ALGORITHMS.values() if a.width == 32]

Each entry is a frozen AlgorithmInfo dataclass with the full Rocksoft / Williams parameter set: name, width, poly, init, refin, refout, xorout, check, desc.

Custom polynomials

from crcglot import AlgorithmInfo, LANGUAGES, generic_crc

# Compute the canonical check value for a custom poly.
check = generic_crc(b"123456789", 16, 0x1234, 0xFFFF, True, True, 0x0000)

# Build an AlgorithmInfo and feed it to any generator.
algo = AlgorithmInfo(
    name="my_crc16", width=16, poly=0x1234, init=0xFFFF,
    refin=True, refout=True, xorout=0x0000, check=check,
    desc="My custom CRC-16",
)
code = LANGUAGES["rust"].generator_from_entry("my_crc16", algo, table=True)

Use with an MCP client (optional)

crcglot[mcp] exposes the CLI surface as a Model Context Protocol server so LLM clients (Claude Desktop, Cursor, mcp-cli, …) can call crc_detect / crc_compute / crc_generate etc. as named tools. The LLM never has to remember a polynomial, slice bytes off a packet to find the CRC, or write a reflection loop — it asks crcglot.

pip install 'crcglot[mcp]'        # the extra ships the MCP SDK
# or:  uv tool install 'crcglot[mcp]'

Then wire it into your MCP client. Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "crcglot": {
      "command": "uvx",
      "args": ["--from", "crcglot[mcp]", "crcglot-mcp"]
    }
  }
}

Tools: crc_list · crc_info · crc_detect · crc_encode · crc_compute · crc_generate · crc_credits. Resources: crcglot://catalogue.json · crcglot://languages.json · crcglot://variants.json. Full reference and Claude Desktop walkthrough live in docs/MCP.md.

Fast runtime CRC (optional C extension)

Beyond generating code, crcglot can compute CRCs at runtime — and it's fast.

Performance, stated honestly: with the C extension, crcglot computes any of the more than 70 CRCs from Python at compiled-C-class throughput on bulk data (~1.7 GB/s on a 1 MiB buffer — on par with generated C and ahead of generated Rust), and for IEEE CRC-32 / JAMCRC it delegates to the stdlib's hardware path (~tens of GB/s), faster than the generated code. The pure-Python fallback always works but is ~1000× slower. Two caveats: the "compiled-class" numbers need the extension installed (the wheel / crcglot[fast]), and they hold for bulk/streaming data — many tiny one-shot calls pay Python↔C overhead per call (use the batch API for those). All figures are platform-specific; see BENCHMARKS.md.

At runtime there's no variant choice to make — the same philosophy as --small/--fast on the generator, taken all the way: you just call crcglot.generic_crc(data, width, poly, init, refin, refout, xorout) and it picks the fastest path available on your machine. There's no table=/slice8= knob here; the speed you get depends only on whether the C extension is installed.

Under the hood it dispatches three ways (you never select among them):

1. IEEE CRC-32 / JAMCRC → stdlib zlib.crc32 (hardware CRC folding — PCLMULQDQ on x86, PMULL / crc32 instructions on ARM): tens of GB/s. No software CRC out-runs silicon, so crcglot borrows the stdlib's path for the algorithms it covers.
2. Everything else → the optional C extension (crcglot._c, slice-by-8 / table-driven): ~1-2 GB/s, ~2,000× over pure Python.
3. No extension built → pure Python: always works, just slow.

The extension ships in the prebuilt wheels (pip install crcglot gets it on common platforms). To force it / pull the build deps explicitly:

uv tool install "crcglot[fast]"     # or: pip install "crcglot[fast]"

It's a single abi3 wheel per platform (CPython 3.11+), and crcglot stays fully functional in pure Python if no wheel matches your platform.

from crcglot import generic_crc

# One-shot.  crc32 here rides the zlib hardware path automatically.
crc = generic_crc(b"123456789", 32, 0x04C11DB7, 0xFFFFFFFF, True, True, 0xFFFFFFFF)

Streaming and batch (C extension)

For chunked data and high-volume small-buffer workloads, the extension exposes two more shapes:

from crcglot import _c   # present iff the extension is installed

# Streaming -- bind the algorithm once, feed chunks, digest on demand
# (hashlib idiom: update / digest / reset / copy).
s = _c.CrcStream(width=32, poly=0x04C11DB7, init=0xFFFFFFFF,
                 refin=True, refout=True, xorout=0xFFFFFFFF)
for chunk in stream:
    s.update(chunk)
result = s.digest()

# Batch -- CRC many buffers, paying the Python↔C transition once
# (the win for framed protocols / packet streams).
results = _c.c_crc_many(list_of_packets, 32, 0x04C11DB7, 0xFFFFFFFF,
                        True, True, 0xFFFFFFFF)

See BENCHMARKS.md for measured throughput of each runtime path against the generated-code gallery.

Example output

See EXAMPLES.md for the actual generated source for crc32 across every language × implementation combination (C / Rust / Python / VHDL / Verilog / Go / C# / Java / TypeScript crossed with bit-by-bit, table-driven, and slice-by-8 where supported). Every block is reproducible with one CLI command.

Benchmarks

See BENCHMARKS.md for measured crc32 throughput across every (language × variant) cell at 1 KiB and 1 MiB. Within each language the trend is monotonic (bit-by-bit < table < slice-by-8) but the absolute speedup at each step depends heavily on how well the compiler optimizes the baseline — Rust's LLVM-vectorized bit-by-bit nearly ties its table-driven, while C# / Python see a 10×+ jump just from table-driven because their bitwise loops aren't vectorized. VHDL and Verilog are excluded: they're simulator references for hardware datapaths, not software runtime.

Acknowledgments

crcglot stands on:

• The reveng CRC catalogue by Greg Cook — the canonical source of CRC algorithm parameters since 1999, and the source of the more than 70 parameter sets, descriptions, and check values every catalogue entry in crcglot is derived from.
• zlib by Mark Adler, Jean-loup Gailly et al. — the runtime fast path for CRC-32/ISO-HDLC and JAMCRC, which take the PCLMULQDQ folding path on x86 and the PMULL / crc32 instructions on ARM.
• The Rocksoft Model CRC parameterization by Ross N. Williams — the (width, poly, init, refin, refout, xorout, check) vocabulary every catalogue entry is expressed in.

crcglot credits prints this same content in the terminal, and crcglot.ATTRIBUTION / crcglot.ACKNOWLEDGMENTS expose it programmatically.

License

MIT