Product launch · September 29, 2026 · updated Sep 30
SuperCompress
v2 is live
Human progress has always been an efficiency story. Bigger context windows answered how much an AI can read. SuperCompress is trying to answer how much it should have to.
The API's default model is Neural Keep. It scores lines against the current query, keeps the original lines that answer it, and drops the rest.
Why v2 is different from v1
v1 was a tiny / lightweight compiler: a local CPU path, good at aggressive drop with answer-keep gates. v2 is the hosted query-aware Neural Keep model. It scores lines against this query and keeps original wording.
Critically, it knows when not to compress. Across the full 390-case public suite many dense needle and LongBench slices keep nearly everything. Query-aware keep shines on coding-agent dumps, where most lines are irrelevant log or tool output. Mixing those results with compiler-era local-CPU claims is wrong, they are different engines.
The problem
Coding agents drown models in tool dumps: logs, diffs, stack traces, RFID heartbeats, the same boot line twelve times. Somewhere in that pile is the one error, ID, or path that answers the question.
Truncation often deletes that line. Summarization rewrites it. Either way you pay for tokens that do not help, or you lose the evidence that does.
What v2 does
Neural Keep is the default on POST https://api.supercompress.dev/compress. Hosted requests use the query-aware engine: it scores lines against this query, keeps original wording, and drops the noise. It does not paraphrase the stack trace. The question itself is never compressed.
That is strongest on coding-agent turns. It also helps on RAG chunks and long support threads when most of the text is not about the ask in front of you.
Illustrative · keep vs drop
INFO boot ok INFO heartbeat INFO heartbeat INFO heartbeat ERROR db timeout at 12:01, pool exhausted (conn_id=9f2a) INFO heartbeat DEBUG retry scheduled
Query
what failed?
Kept
ERROR db timeout at 12:01, pool exhausted (conn_id=9f2a)
Coding-agent benchmark
Held-out coding-agent suite. Same scorer for every case. We measure whether the required evidence survives, not whether a downstream chat model finishes a task. Downstream LLM eval has not been run yet. Headline cut figures from this suite are on hold pending re-verification.
| What we publish | SuperCompress v2 |
|---|---|
| Engine | Hosted Neural Keep (query-aware) |
| Method | Extractive keep, original lines, not a summary |
| Quality metric | Evidence containment, required lines still present |
| Public suite size | 390 cases |
Raw methodology JSON: launch-benchmark.json · full writeup on SuperCompress benchmarks.
How we measured
Evidence containment in the compressed text (tiktoken cl100k_base). Pass means required lines are still present in original wording. This is stricter than vibes and weaker than a full downstream completion eval, we say so on purpose.
Limitations
- Downstream eval: we have not yet published LLM task-success after compression.
- Workload fit: strongest on coding-agent / tool-dump noise; conservative when every line could be the needle.
How to start
The product path is the hosted API. Current plans are in the dashboard. For coding agents:
npm install -g supercompress-proxy supercompress setup
How to use it
Available now as a hosted API, and as an agent plugin for Cursor, Claude Code, Codex and other coding agents.
- Get an API key, compress over HTTPS before the model call.
- Install the agent plugin, hooks compress tool dumps in your harness.
- See the benchmarks for the published v2 suites.
Ship with less context
Keep what answers the query. Cut the rest. Open source.