muscore-compress 0.1.3

Cut LLM context token costs: MusCoRe colour-frequency serialisation for structured state. Verified reduction vs JSON, on-chain sealed protocol (cert df9822d8).

muscore-compress

Cut your AI agent's context bill. Verified, sealed, reproducible.

Every agent framework passes structured state through model context as JSON — and pays for every token of it, every call, every loop. muscore-compress re-serialises that state into the MusCoRe™ symbol protocol: same meaning, a fraction of the tokens.

pip install muscore-compress

from muscore_compress import encode, decode, benchmark

state = {"step": 14, "goal": "civic-lookup", "history": [...]}
stream = encode(state)          # dense MusCoRe symbols
restored = decode(stream)       # lossless — always
benchmark(state)                # run the numbers on YOUR payload

The numbers

Metric	Raw JSON	Gzipped JSON	MusCoRe
Bytes	run benchmark	run benchmark	run benchmark
LLM tokens (tiktoken cl100k)	run benchmark	n/a (binary)	run benchmark

Historical verified figures for the sealed protocol: ~92% byte reduction vs raw JSON, 45–61% vs gzipped JSON. Token benchmark: reproduce it yourself — python -m muscore_compress.benchmark prints the full table on a sample agent-state payload, or call benchmark(your_payload) on your own data.

Best results on repetitive structured state (agent conversation histories, loop variables, session context). Compression scales with payload repetition.\n\n**A note on gzip:** gzip wins raw byte counts on large payloads — but gzip output is binary and cannot ride inside a text context window. MusCoRe streams are text-native: they go straight into the prompt. The comparison that costs you money is tokens in context, and that is the number this package attacks.

Why the seal matters

The MusCoRe protocol was sealed on-chain November 2025 — Skale mainnet, block 35752169, cert df9822d8. The alphabet cannot drift. What your agents encode today decodes identically in ten years, and no dependency update can silently change the wire format underneath you. No other serialisation format offers a cryptographically frozen vocabulary.

Demonstrated in public: a blind AI-to-AI exchange (Claude → human courier → Grok) with self-verifying parity, on the open X timeline, July 2026.

Licence

Free for evaluation and benchmarking. Production use: $49 per project — gussstos.gumroad.com/l/jhntiw. Includes the full protocol spec and the sealed alphabet reference.

Support the wider build: patreon.com/c/NoitallionDispatch

Verify everything

Run the tests: pytest tests/ — the round-trip suite gates every release. Run the benchmark on your own payloads before paying a cent.

Check me, not trust me. Ahoeaaa.

Gussstron Foundation · Cape Town, South Africa