Compresr

Compresr is a context compression API for LLM pipelines that reduces token usage by up to 200x without quality loss. It offers multiple compression models — from query-agnostic pre-compression to aggressive query-specific filtering — enabling developers to cut costs, reduce latency, and improve accuracy in RAG, search, Q&A, and agent workflows. Backed by Y Combinator (W26), Compresr also provides an open-source Context Gateway for agent frameworks like Claude Code and OpenClaw.

• Espresso V1 — Query-agnostic token-level compression; pre-compress long documents, system prompts, or agent histories once and reuse across multiple queries.
• Latte V1 — Query-specific token-level compression that retains only tokens relevant to a given question, enabling up to 200x compression; ideal for RAG, search, and Q&A pipelines.
• Coldbrew V1 — Query-specific chunk-level filtering that drops entire irrelevant chunks before they reach the model; ideal for structured data like transcripts or logs.
• Context Gateway — Open-source proxy for agents that compresses conversation history, tool outputs, and tool lists; compatible with Claude Code, OpenClaw, Codex, and more.
• Coarse-grained filtering — Pass a query and a list of chunks to retrieve only the relevant ones, reducing context before fine-grained compression.
• Fine-grained compression — Token-level compression given a query and context, stripping irrelevant tokens for maximum efficiency.
• Usage-based pricing — Pay per million tokens processed, with no upfront commitment; get started by signing up and calling the API.
• Demo environment — Try compression live on sample datasets like SEC filings directly from the dashboard without writing any code.