Format-aware compression without the compression expertise. Point Claude Code at your data - it identifies the structure, writes the data description, trains an OpenZL compressor, verifies the round-trip byte-for-byte, and reports what you'd actually save versus zstd and gzip.
OpenZL (Meta, open-sourced Oct 2025) compresses structured data far better than generic compressors by decomposing it into typed streams - if someone describes the data's structure in its SDDL language or a C++ parser. That description step is the adoption barrier. This project makes an LLM do it.
Real datasets, zli 0.2.4, every result decompressed and cmp-verified byte-identical
(full logs in results/):
| Dataset | Path | vs zstd -3 |
vs zstd -19 |
|---|---|---|---|
| NYC 311 CSV (87.5 MB) | csv profile, trained |
−50.0% | −11.7% |
| Binary tick feed (65.6 MB) | LLM-written SDDL | −48.7% | −34.5% |
| PyTorch checkpoint (46.8 MB) | pytorch profile |
−10.0% | −10.0% |
| GH Archive JSONL (80 MB) | serial, trained |
−6.0% | loses |
The second row is the interesting one: a binary format no OpenZL profile covers, with the
data description written zero-shot by the LLM from a hexdump - and it beats zstd -19 by
a third while compressing 22x faster. The last two rows are why the agent benchmarks
honestly instead of overselling: checkpoints win on throughput and bytes removed, not
ratio, and JSON/JSONL is currently weak (OpenZL has no JSON profile).
Decompression needs no trained compressor and no description - the .zl frame is
self-describing, so there's no lock-in.
git clone https://github.com/avionicharshit-byte/openzl-agent.git
cp -r openzl-agent/skill/openzl ~/.claude/skills/You'll need zli built from facebook/openzl
(v0.2.4+ recommended) - the skill locates it, or offers to build it for you. Then just
ask Claude Code for compression work, e.g.:
- "Compress this directory of CSVs and show me what it saves."
- "We store 40 TB/month of these event logs. Is there anything better than zstd?"
- "Here's a binary file from our trading feed - write an SDDL description for it."
See skill/openzl/README.md for full docs, including the list
of execution-verified gotchas the skill knows that the public docs don't (two mutually
incompatible SDDL dialects, silent fallback without --strict, training-budget overruns,
and more).
| Path | What |
|---|---|
skill/openzl/ |
The Claude Code skill: playbook, verified zli cheatsheet, SDDL1/SDDL2 guides, detect + benchmark scripts |
results/ |
Benchmark logs from the validation spike, plus the LLM-written SDDL descriptions |
- Claude Code skill - shipped (this repo)
- Standalone CLI (no Claude Code dependency, calls the Claude API directly)
- GitHub Action, MCP server
- Hosted storage-savings audit - only if users pull for it