SuperCompress v2

---

Why it exists

Every LLM call ships a pile of context: RAG chunks, chat history, tool dumps, logs, JSON. Most of it is irrelevant to the current question — but the model still reads it, and you still pay for it.

Usual “fix”	What actually happens
Truncate	Deletes the middle. The answer is often in the middle.
Summarize	Rewrites evidence. IDs, stack traces, and exact errors get soft.
Hope	Ship the full dump. Watch the bill climb.

SuperCompress v2 compresses context against the query. It keeps answer-critical lines in their original wording and drops the rest — on our coding-agent benchmark (B5), 64.1% mean context reduction with 24/24 evidence-retention passes.

---

Product

SuperCompress is a compression layer in front of inference:

1. Takes a long context + the current query
2. Segments and scores blocks by relevance to that query
3. Keeps entities, errors, definitions, nearby dependencies
4. Returns a smaller prompt + token stats

The query is never compressed — only the surrounding context.

Two paths

	Neural v2 (recommended)	Compiler (fast/local)
Engine	~400M query-aware cross-encoder	Lightweight local policy
Runtime	Hosted Fly CPU (Neural Keep)	Local CPU · millisecond-class
Best for	Highest-quality keep on agent dumps	Local preprocessing / speed
Benchmarks	Launch / B5	Legacy section

Two products

	Coding-agent plugin	API / Python
For	Cursor, Claude Code, Codex, and 40+ agent harnesses	Apps, RAG, agents, backends
Install	npm i -g supercompress-proxy && npx supercompress setup	pip install supercompress
What you get	MCP compress_context on big dumps	Compress before every model call
Login	Keep your normal agent login	API key from the dashboard

Docs: coding agents (first 5 minutes checklist) · API quickstart

Repo map

Path	What
packages/proxy	Coding-agent plugin (npm)
api/	Hosted API + billing
web/	Site + docs HTML
supercompress/	Python package
docs/REPO_LAYOUT.md	What belongs in OSS vs private

Private marketing, outreach, and model training stay out of this repo (see .gitignore + docs/REPO_LAYOUT.md).

---

Benchmarks & stats

We measure whether required evidence survives (containment), not downstream LLM completion.

Neural v2 launch (hosted ~400M) — B5 coding-agent suite

Metric	Result
Mean context cut	64.1%
Evidence passes	24 / 24
Tokens	16,647 → 5,148
Max cut @ ≥99% retention	96.6%
B5 latency p50	~5.8 s (hosted Neural Keep on Fly CPU)
Public cases (full suite)	390
Downstream LLM eval	Not yet run

Aggregate mean cut across all 390 cases is ~3.6% — the engine often refuses to over-cut dense needle/QA slices. The 64.1% figure is the coding-agent suite where dumps are noisy.

Raw JSON: launch-benchmark.json · writeup: benchmarks

Legacy compiler (separate product)

CPU / millisecond-class local path. Older held-out compiler numbers (≈58–66% cut, 99.4% gold containment) live under Legacy/compiler on /benchmarks. Do not mix with Neural v2.

---

Try it

Coding agents (recommended)

npm install -g supercompress-proxy
npx supercompress setup

Links your account, detects agents, installs MCP. Docs: coding agents

Python / HTTP

pip install supercompress
export SUPERCOMPRESS_API_KEY=sc_live_YOUR_KEY

from supercompress.client import SuperCompress

sc = SuperCompress()
result = sc.compress(
    context=long_context,
    query="What failed and how do we fix it?",
)
print(f"{result.original_tokens} → {result.kept_tokens} tokens")
print(result.compressed_text)

curl -X POST https://www.supercompress.dev/api/v1/compress \
  -H "X-API-Key: $SUPERCOMPRESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"context":"...","query":"What failed?"}'

Or paste a dump into the Arena — no integration required.

---

How it compares

	Truncate	Summarize	SuperCompress
Cuts tokens	✓	✓	✓
Uses the query	✗	weak	✓
Keeps original evidence	sometimes	✗	✓
Auditable kept lines	partial	✗	✓

More: vs truncation · vs summarization · vs alternatives

---