Query-aware context compression
Scores prompt segments against the current user question and removes context that is less relevant to answering that query.
SuperCompress
SuperCompress is a query-aware context compressor for LLM apps that cuts input tokens before inference while preserving key evidence. Open source, CPU, API and Python package.
Overview
SuperCompress is a query-aware context compressor for LLM applications. It sits between your app and the model, reducing input tokens before inference by selecting the parts of a prompt that are most relevant to the user’s current question.
The product is designed for long-context workflows such as chatbots, RAG pipelines, support copilots, AI search, and agent loops. The source describes it as open source under MIT, CPU-only, and around 5K parameters, with a hosted API and a local Python package available for use.
Features
Pre-inference prompt preprocessing
Supports chat history, RAG chunks, support transcripts, tool traces, logs, and other long-context inputs before the model call.
Compression reporting and diagnostics
Returns token counts, tokens saved, kept percentage, and a compression risk level so teams can review the impact before rollout or in logs.
Hosted and local workflows
Provides a hosted API and a local Python package, with curl, JSON, and form-encoded request options shown in the docs.
Dashboard and test workflow
Includes an interactive playground and dashboard for trying contexts, creating keys, and viewing metered usage.
Integration guidance
Documents integration paths for OpenAI, LangChain, Express.js, Vercel AI SDK, and local model calls.
Use Cases
Chatbots and conversational AI
Shorten chat history and repeated conversation state before sending each request to a model, helping consumer AI apps control token growth over time.
Retrieval-augmented generation
Compress retrieved passages and document chunks in RAG pipelines so the model sees the most relevant evidence without overshooting the context window.
Support copilots
Reduce support transcripts, ticket history, and knowledge base excerpts before copilots and agent tools call the LLM.
Agent loops and tool traces
Trim tool traces, intermediate outputs, and logs that accumulate across turns in agent workflows.
Prompt compression evaluation
Test the product against your own prompt, compare original and compressed context, and inspect what was kept or removed before rollout.
Pros and Cons
Pros
- Targets token cost and latency before the model call rather than after generation.
- Preserves answer-critical evidence instead of blindly truncating the tail of the context.
- Works on CPU and does not require a GPU.
- Offers both hosted and local usage paths.
- Provides measurable output fields such as tokens saved and important-kept percentage.
Cons
- The pricing page was not available in the collected source, so public plan details are limited to the dashboard.
- The documentation shows both compiler mode and fixed-ratio mode, which may require users to choose the right workflow for their use case.
- The source does not provide extensive deployment, security, or team-administration detail beyond the dashboard and docs.
FAQ
How do you use SuperCompress?
SuperCompress can be used either through the hosted HTTP API or through the Python package. The docs show a quick start with the API key in the dashboard and examples for Python and curl.
Does SuperCompress run locally or only as a hosted API?
The documentation says SuperCompress works in compiler mode on the hosted API, while the local Python package supports compression functions such as compress_context(), compress_for_turn(), and compress_detailed().
What tools and model stacks does it fit with?
The source lists OpenAI, Claude, Gemini, and local model calls as supported targets, and the docs mention integrations for OpenAI, LangChain, Express.js, and the Vercel AI SDK.
What does the output include?
The documentation describes a query-aware compiler that compresses context before inference, with output fields such as compressed text, tokens saved, important-kept percentage, and compression risk.
Is there pricing information available?
The docs say the hosted API is metered through the dashboard and that the plan limits are monthly token limits. A pricing page was not available, so only the dashboard plan labels are visible in the source.
Quick Facts
- Category
- Developer Tool
- Product type
- Query-aware context compressor for LLMs
- Platform
- Hosted API and Python package
- Runtime
- CPU-only; no GPU required
- License
- MIT
- Source domain
- supercompress.dev
SuperCompress Alternatives
AakarDev AI
AakarDev AI helps teams manage AI provider access, project setup, logs, and analytics in one dashboard. BYOK support included.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
Arduino VENTUNO Q
Arduino VENTUNO Q is an edge AI computer for AI and robotics, combining AI inference and deterministic control on one board with Arduino App Lab support.
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
Codex Plugins
Codex Plugins bundle reusable skills, app integrations, and MCP servers into workflows you can install in the Codex app or use from Codex CLI.
hob
hob is an independent workspace for coding agents, with local control over sessions, terminals, history, routing, and follow-up work.