You deserve better compression.
A local-first archiver that routes files to different codecs by measuring what they actually contain, packs them into a custom container with authenticated encryption, and never opens a socket.
Status: 1.0.0, first release. The format specification is frozen. The detector, compressor, extractor, cryptography, container, GUI and command line are all implemented, with 280 tests passing. Benchmarks against 7-Zip, zstd, xz and tar are published below.
It is a first release and has not yet been run by many people on many machines. Keep your originals until you have verified an archive yourself —
pakt verifychecks every file against its recorded SHA-256.
| Windows 10 / 11 | Compakt_Setup.exe — desktop app + pakt command, no admin prompt |
| Debian / Ubuntu | compakt_1.0.0_amd64.deb — sudo apt install ./compakt_1.0.0_amd64.deb |
| Any Linux (glibc 2.28+) | pakt-1.0.0-linux-x86_64.tar.gz — command line only |
Checksums: SHA256SUMS.
The Windows installer is not code-signed, so Windows shows "Windows
protected your PC" — why, and how to check the file yourself.
pakt c photos/ # -> photos.pakt
pakt c docs/ cv.pdf -o private.pakt -p # encrypted, file names included
pakt x photos.pakt # also .zip .7z .tar.gz .rar and ~40 more
pakt verify photos.pakt # check every file's SHA-256
Most archivers decide how to compress a file by looking at its name.
Compakt reads the bytes. A SQLite database renamed to .txt is routed
as a database, not as prose. An already-compressed blob with a
misleading extension is stored raw instead of being pointlessly
recompressed, because its measured entropy gives it away.
That routing feeds a container format, .pakt, designed around six
commitments:
- Truncation is survivable. The index is written at both ends.
- Unknown archives fail loudly. Feature flags are explicit and readers reject what they do not understand.
- The file listing is not public. With a password set, the index itself is encrypted — unlike ZIP, which leaks every filename.
- Authentication precedes disclosure. No plaintext reaches disk before its GCM tag verifies.
- The container is relocatable. A reader locates it from the footer, so an arbitrary prefix cannot break an archive.
- There is no version 2 to wait for. Everything structural is defined in 1.0.
Packing reads each file once and holds a quarter of your RAM at most (never less than 256 MB), plus the few blocks being compressed. Past that it works from a temporary file beside the archive, so a folder larger than memory packs just the same.
Compakt does not open network sockets, send telemetry, use cloud APIs, or check for updates. Every library in the stack is a pure codec — maths on bytes, nothing more.
That claim is checkable rather than merely asserted. The socket guard lives in this repository, and the behaviour is observable from outside the process:
- block Compakt in your OS firewall — it works identically
- run it in an air-gapped VM, or with the network cable out
- watch it with Wireshark, Resource Monitor, or Little Snitch
Being precise about this up front, because the distinction matters.
| Component | Licence |
|---|---|
.pakt format spec + reference decoder |
Apache-2.0 |
| This repository — detector, extractor, crypto, socket guard, GUI, CLI, reference encoder | MPL-2.0 |
| The compression routing engine — the per-block codec decision | proprietary, not published |
Everything that could hurt you is open and auditable. What is closed is
one file that decides which codec to call, which has no bearing on
whether Compakt is trustworthy or on your ability to read your own
archives — core/reference_encoder.py in this repository is a real,
working encoder that produces valid .pakt files.
No part of reading your archive is withheld. Where the engine trains a compression dictionary, it trains it from your own files at pack time and embeds it in the archive. Compakt never writes an archive that refers to a dictionary it does not carry, and there is no dictionary corpus you need and cannot have. (The format can name an external dictionary by id — spec §12.4 — and a reader must then refuse with a message naming it rather than guess. Compakt does not write those.)
The format is free. Implement .pakt in any language, for any
purpose, commercially or not, without asking. Every archive ever
written stays readable by anyone, forever, whatever happens to this
project. See LICENSING.md and
docs/pakt-format-spec.md.
Packs: .pakt only.
Unpacks: .pakt, .zip, .7z, .tar and its variants, .gz,
.bz2, .xz, .zst, .br, .iso, plus roughly forty more via
libarchive including .rar, .cab, .lha, .cpio, .deb, .rpm
and .arj.
Writing .rar is never offered. RARLAB licenses no compressor.
python -m venv .venv
.venv/Scripts/pip install -r requirements-dev.txt # Linux/macOS: .venv/bin/pip
.venv/Scripts/python -m pytest # Linux/macOS: .venv/bin/python
Python 3.14 or newer. Every dependency ships a prebuilt wheel, so no C compiler is needed to run or test. One is needed only to build a release binary.
External code contributions are not being accepted at this time — see
CONTRIBUTING.md for why. Bug reports, benchmark
results (especially ones where Compakt loses), security findings and
independent .pakt implementations are all welcome.
Published, with the losses beside the wins:
benchmarks/RESULTS.md. Every corpus is
fetched from a public source, rebuilt from pinned dependencies, or
generated from a fixed seed, and benchmarks/run.py reproduces the
table on your own machine.
Against 7-Zip -mx9. Ratio is the archive's size as a fraction of
the original, so lower is better:
| Corpus | Compakt | 7-Zip | |
|---|---|---|---|
| silesia (202 MB) | 0.2288 in 45.7s | 0.2297 in 69.5s | 0.4% smaller, and faster |
| enwik8 (95 MB) | 0.2545 in 78.8s | 0.2480 in 84.5s | 2.6% behind |
| sitepackages (64 MB, 4,027 files) | 0.2126 in 24.4s | 0.2111 in 18.2s | 0.7% behind |
| jsonlogs (1.6 MB, 8,000 files) | 0.3724 in 4.4s | 0.1596 in 1.4s | 2.3× larger |
| sourcelike (1.3 MB, generated) | 0.0546 in 0.8s | 0.0123 in 0.2s | 4.4× larger |
silesia is the corpus chosen because it switches off every structural advantage Compakt has, so being ahead there is the result that matters. enwik8 is a single 95 MB text file, where splitting into 64 MiB blocks costs more than the routing wins back — that trade was made deliberately and is not free.
The big losses are small folders of many tiny files. There the index — a record of every file's path, size and digest, so each can be verified and extracted on its own, stored twice so a damaged archive can still be recovered — is most of the archive: 66% of it for jsonlogs, 92% for sourcelike. On folders like those, 7-Zip makes the smaller archive.
Unpacking is where Compakt is quickest: the 4,027 files of sitepackages come out in 2.7s, against 7.7s for 7-Zip.
Ratios are deterministic and reproducible. Timings come from one machine and will vary with yours.
Copyright (c) 2026 Rounak Miskin (Founder: Kasanki Labs)