Skip to main content
Tgtoolglade.
SuperCompress logo

SuperCompress

Query-aware prompt compression that cuts LLM input tokens by roughly 60% before inference.

Hand-reviewed by the Toolglade team·Updated Aug 2026
coding#llm#developer-tools#cost-optimization
Visit Website
Free plan Claimed API Self-hosted
Toolglade’s take

SuperCompress is a promising, genuinely open-source (MIT) way to cut LLM input tokens before inference, with a real free tier and a drop-in API, SDK, and coding-agent proxy. It is still an early-stage project and its headline compression (~60%) and answer-retention (>98%) figures are vendor-reported, so test it on your own prompts in the free playground before relying on the savings.

About SuperCompress

SuperCompress is an open-source, query-aware prompt compression tool that removes low-value context before an LLM call to cut input token costs and latency while preserving answer-critical evidence. It runs on CPU (no GPU), ships as a hosted API, a Python package, a browser playground, and an MCP proxy for coding agents, and works with OpenAI, Anthropic, Mistral, and more. Pricing is freemium: 1M free tokens/month, then $0.30 per 1M tokens, plus free self-hosting under MIT.

SuperCompress is a developer tool that sits between your application and a large language model, preprocessing the prompt to strip out tokens that are not relevant to the current user question. Instead of blindly truncating the tail of a prompt or paying for an extra summarization call, it scores each segment of the context against the query and keeps only what matters, which the project reports typically cuts input tokens by around 58-66% while retaining answer-critical evidence. It is open source under the MIT license and runs on CPU using a small learned policy (~5,200 parameters, ~60ms per compression, no GPU or model download). You can self-host the Python package (pip install supercompress) or call the hosted API, which is freemium: 1,000,000 free tokens per month with no credit card, then $0.30 per 1,000,000 tokens. A drop-in API and SDK work with OpenAI, Anthropic, Mistral, LangChain, and the Vercel AI SDK, and an MCP-based proxy compresses context for coding agents such as Cursor, Claude Code, Codex, Windsurf, and GitHub Copilot. SuperCompress is aimed at teams shipping production LLM features where context grows over time: chatbots with long histories, RAG pipelines with oversized retrieved chunks, support copilots, and agent loops accumulating tool traces. It is an early-stage, independently developed project, and its headline compression and retention numbers are vendor-reported, so real-world savings are best confirmed by testing your own prompts in the free playground.

TL;DR

SuperCompress is an open-source (MIT), query-aware prompt compression tool that removes low-value context before an LLM call, reportedly cutting input tokens by around 58-66% while retaining answer-critical evidence. It runs on CPU with no GPU, ships as a hosted API, Python package, browser playground, and coding-agent MCP proxy, and works with OpenAI, Anthropic, Mistral, and more. Pricing is freemium (1M free tokens/month, then $0.30 per 1M tokens) plus free self-hosting. It is an early-stage, independently built project, and its benchmark figures are vendor-reported.

Company overview

SuperCompress is an open-source developer tool for prompt and context compression, distributed under the MIT license via a Python package (pip install supercompress), a hosted API, a browser playground, and an MCP-based coding-agent proxy. Founded by Arjun Shah, it launched publicly in 2026 (PyPI and Product Hunt) and is developed in the open on GitHub. The project positions itself against alternatives such as blind truncation and LLM summarization by scoring context against the user's query. As an early-stage, independently run open-source project, it has no publicly disclosed venture funding.

Product features

In its default compiler mode, SuperCompress segments the input context, scores each block against the current query using a small learned policy (~5,200 parameters), removes duplicate and low-value blocks, and preserves nearby dependencies (headings, imports, trace context), reporting tokens saved, percentage of important context kept, and a risk level. It runs on a single CPU core in roughly 60ms with no GPU or model download, since weights ship inside the pip package.

Access is via a REST API, Python and TypeScript SDKs, or self-hosting. Integrations cover OpenAI, Anthropic, Mistral, LangChain, and the Vercel AI SDK, and a one-command MCP proxy (npx supercompress setup) compresses context for coding agents including Cursor, Claude Code, Codex, Windsurf, and GitHub Copilot. A private/local deployment option keeps data in your own environment.

Target market

SuperCompress targets software developers and teams shipping production LLM features where context grows with usage: chatbots and assistants with long histories, RAG pipelines, AI search, support copilots, and agent loops that accumulate tool traces. Its usage-based, low-cost pricing and free self-hosting make it accessible to individual developers as well as cost-conscious engineering teams; it is less suited to non-technical users who want a finished app.

Buyer personas

End users

Software developers and ML engineers who integrate the API, SDK, or coding-agent proxy to compress context before LLM calls.

Buyers

Engineering leads, founders, and CTOs at startups and product teams looking to cut LLM API spend without switching models.

Key influencers

AI/platform engineers, DevOps and cost-optimization advocates, and developers comparing compression approaches (truncation vs summarization vs query-aware).

Ideal customer profile

A team spending meaningfully on LLM API calls with large or repetitive context (RAG, long chats, agent traces) that wants a drop-in, model-agnostic way to reduce tokens, and is comfortable integrating an API or self-hosting an open-source package.

Funding & performance

SuperCompress appears to be an early-stage, independent open-source project (MIT-licensed, launched on PyPI and Product Hunt in 2026) founded by Arjun Shah. No venture funding has been publicly disclosed; treat any funding figures as unavailable/undisclosed.

Pros & cons

Pros

  • Open source under the MIT license and free to self-host
  • Genuine free tier: 1M tokens/month with no credit card
  • Cheap
  • transparent usage pricing at $0.30 per 1M tokens
  • Runs on CPU with no GPU or model download (~60ms per compression)
  • Drop-in API and SDK for OpenAI
  • Anthropic
  • Mistral
  • LangChain
  • and more
  • MCP proxy compresses context for coding agents like Cursor and Claude Code
  • Query-aware: keeps context relevant to the current question instead of blind truncation

Cons

  • Early-stage project with a small team and limited independent track record
  • Headline compression (~58-82%) and >98% retention figures are vendor-reported and benchmark-dependent
  • Compression is lossy
  • so aggressive settings can drop context that later turns out to matter
  • Text-only: it does not compress image or audio context
  • Adds a preprocessing step and a small latency overhead before each model call
  • No public team
  • collaboration
  • or enterprise-console features documented yet
  • Savings depend heavily on how bloated your prompts already are

Pricing plans

Open Source (MIT)
Free
  • Self-host the Python package (pip install supercompress)
  • Runs on CPU, no GPU or model download
  • Full compiler mode and query-aware scoring
  • MIT license, free for commercial use
Free (hosted API)
$0 / month
  • 1,000,000 tokens per month included
  • No credit card required
  • Hosted API and SDKs
  • Free interactive playground (no signup)
Pay-as-you-go
$0.30 per 1M tokens
  • Usage-based after the free tier
  • Same API, SDKs, and coding-agent proxy
  • Scales with token volume
  • No minimum commitment
Enterprise / Local
Custom
  • Higher and custom volumes
  • Local / private deployment option
  • Keep data in your own environment
  • Contact SuperCompress for terms

Key features

API
Self-hosted
Multi-language
Integrations
OpenAI, Anthropic, Mistral, LangChain, Vercel AI SDK, MCP, Cursor, Claude Code, GitHub Copilot, Windsurf
Input types
text
Output types
text
Best For
Reducing LLM API costs, RAG context compression, Coding agent context, Long chat histories

Compare key features

View all alternatives →
Feature
SuperCompress
Windsurf
Tabnine
Pricing
Freemium
Freemium
Freemium
Free plan
Yes
Yes
Yes
Free trial
No
No
No
API
Yes
Yes
No
Self-hosted
Yes
Yes
Yes
Team support
No
Yes
Yes

Frequently asked questions

What is SuperCompress?+

SuperCompress is an open-source, query-aware prompt compression tool for LLM apps. It sits before your model call and removes context that is not relevant to the current question, reducing input tokens (and therefore cost and latency) while trying to keep answer-critical information intact.

Is SuperCompress free?+

It is open source under the MIT license, so you can self-host the Python package for free. The hosted API is freemium: 1,000,000 tokens per month free with no credit card, then $0.30 per 1,000,000 tokens. There is also a free playground with no signup.

How much can it actually save?+

The project reports a mean token cut of roughly 58-66% on its held-out benchmark suites (with over 98% answer-critical retention), and higher figures on bundled presets. Real savings depend on how much redundant context your prompts carry, so the best test is to run your own data through the free playground.

How is it different from truncation or summarization?+

Truncation blindly drops the head or tail of a prompt and can cut the part that answers the question. Summarization rewrites the context with an extra LLM call and can invent or omit details. SuperCompress instead scores each segment against the query and removes low-value tokens, with no extra model call.

Does it work with my LLM and coding agent?+

Yes. It is model-agnostic and works via API or SDK with OpenAI, Anthropic, Mistral, LangChain, and the Vercel AI SDK. For coding agents, an MCP proxy (npx supercompress setup) auto-detects Cursor, Claude Code, Codex, Windsurf, GitHub Copilot, and more.

Reviews (0)

Write a review

Pick a rating
Loading reviews…
Compare

Compare SuperCompress with other AI tools

Side-by-side pages for pricing, features, and best-fit use cases.

All comparisons →
Compare side by side

SuperCompress vs Windsurf

Compare →
Compare side by side

SuperCompress vs Tabnine

Compare →
Compare side by side

SuperCompress vs Lovable

Compare →
Compare side by side

SuperCompress vs Continue

Compare →
Compare side by side

SuperCompress vs GitHub Copilot

Compare →
Compare side by side

SuperCompress vs Replit AI

Compare →

Similar tools you may like