Meaning-First Compression
Retains critical information, intent, and structure—so your models stay effective.
Cut your LLM API Costs by 65%.
For teams running chat, RAG, support, or coding agents. SuperCompress removes low-value context before inference—so your models do more with the tokens you already pay for.
1M free tokens/mo · then $0.30/1M · no credit card
Prompt compression · RAG context · chat memory · CPU-only policy
Consumer AI apps send chat history, retrieved docs, tool traces, and user context on every request. That context drives LLM API cost, latency, and context-window pressure before the answer is generated. SuperCompress runs before inference, selecting the context most relevant to the current user request. Use it for chatbots, AI search, support agents, copilots, RAG, and any feature where context grows with users.
EXAMPLE: LONG CONVERSATION20,814 TOKENS
const result = await compress({
context: conversation,
query: "What failed in the last deploy?",
mode: "compiler"
})
// → kept answer-critical lines; cut waste
COMPILER MODE
0%+
Mean Token Cut
Drop as much as possible on held-out LongBench / OOD dumps — typically ~58–66% mean cut.
GATE
0%+
Critical Kept
Always target >98% answer-critical retention. Pooled held-out gold containment: 99.4% (180/181).
CPU
no GPU
Query-aware
Semantic keep/drop vs the current ask. No extra LLM summarizer call.
BUILT FOR DEVELOPERS
Control AI Feature Cost
Before the Model Call
Drop-In Simple
Use our API or SDK in minutes. Works with OpenAI, Anthropic, Mistral, and more.
Private by Design
Your data stays yours. Local inference option and enterprise-ready deployments.
HOW IT WORKS
Compress before
the model call.
- 01 Ingest Chat, RAG hits, tool traces
- 02 Score Relevance vs current ask
- 03 Keep Meaning-first retention
- 04 Ship Fewer tokens to inference
Context arrives oversized — history, docs, and traces piled into one prompt.
COMPILER MODE · SEMANTIC KEEP/DROP
Cut hard.
Keep what answers.
No fixed keep ratio. Score context against the query, remove everything you can, and fail the run if critical retention drops below 98%.
Full methodology →
METHODMEAN TOKEN CUT
(HIGHER IS BETTER)CRITICAL / ANSWER KEPT
(GATE >98%)HOW IT WORKS
(HIGHER IS BETTER)CRITICAL / ANSWER KEPT
(GATE >98%)HOW IT WORKS
SuperCompress
~58–66%
≥98%
Query-aware compiler · maximize drop
LLM summarization
variable
rewrites
Extra model call · can invent / omit
Truncation
forced
mid-context risk
Head/tail only · blind to the ask
Headroom (public claim)
~15–20%
unpublished
Type / tool-output heuristics
SuperCompress numbers: held-out compiler suites (real / fresh4 / fresh5) — pooled gold answer containment 99.4% (180/181), mean cut typically ~58–66%, gates ≥98% answer keep and ≥55% cut. Not a fixed keep budget. Details on /benchmarks.
curl -X POST https://api.supercompress.dev/compress \
-H "Authorization: Bearer $SC_LIVE_..." \
-H "Content-Type: application/json" \
-d '{
"context": "$(cat conversation.txt)",
"query": "What failed in the last deploy?",
"mode": "compiler"
}'// Response
{
"compressed": "...",
"original_tokens": 23014,
"compressed_tokens": 8055,
"tokens_saved_pct": 62,
"important_kept_pct": 99,
"policy_name": "SuperCompress"
}
Cursor
Claude Code
Codex
Gemini CLI
Windsurf
VS Code
Copilot
OpenCodeEvery coding agent. One compression layer.
CODING AGENT PLUGIN · MCP-FIRST
One Install. Every Agent.
Fewer Tokens.
Auto-detect Cursor, Claude Code, Codex, OpenCode, Windsurf, and more. The MCP plugin compresses huge dumps before they burn tokens — keep your normal login.
Or without a global install: npm install supercompress-proxy then npx supercompress setup.
$
npm install -g supercompress-proxy$
npx supercompress setup- ✓ Detects 48 integrations
- ✓ MCP plugin installed
- ✓ Works with login — no API-key mode
USE CASES
Use Cases That
Compound
From agentic coding to document-heavy work, meaning-first compression helps teams do more with less—without losing what matters.
Coding Agents
Compress task history, repo context, tool traces, and diffs so agents reason farther inside the same context window—and spend less per turn.
- Shrink multi-file diffs before the model call
- Keep plans + decisions, drop stale tool noise
- Works as an MCP plugin across Cursor, Claude Code, Codex
Long Conversations
Keep long chats useful and coherent. SuperCompress retains the decisions, constraints, and facts that drive better answers—not every filler turn.
- Lower cost on multi-hour support or sales threads
- Preserve user preferences and prior commitments
- Reduce latency as history grows
RAG & Search
Retrieved chunks often drown the query. Compress retrieved context so the model sees the evidence that matters for the current ask.
- Cut redundant passages across top-k results
- Keep citations and key claims intact
- Fit more evidence into the same budget
Support Copilots
Ticket history, macros, and knowledge-base hits add up fast. Compress before generation to keep replies accurate without burning tokens.
- Prioritize current issue + account state
- Drop repeated boilerplate from prior tickets
- Ship faster replies at lower per-ticket cost
Document Workflows
Specs, tickets, research notes, and PRDs are dense. Compress for reviews and synthesis while retaining requirements and decisions.
- Faster design and compliance reviews
- Keep must-have constraints and open questions
- Pair with human-in-the-loop editing
Production Inference
Reduce token usage and latency in live apps while preserving output quality. Run locally on CPU or deploy beside your API gateway.
- 5–20ms local compression path
- Drop-in before OpenAI, Anthropic, and more
- Meaning retention built for product workloads
Live polling…
—
tokens processed
across SuperCompress · updates as people compress
SuperCompress
Ship AI features with lower per-request cost.
1M free tokens/mo · then $0.30/1M · no credit card.