EXAMPLES

5-minute wow, once per framework.

Each example ships with the same four things: the problem it solves, exact before/after token counts, a copy-paste install, and a 2-minute demo. Runnable code lives in the repo. Numbers come from our 15-turn product-support benchmark.

LANGCHAIN

LangChain

ConversationBufferMemory replays every turn — cost explodes after 10 turns.

1,259 → 221 tokens/turn · 82% reduction

OPENAI AGENTS SDK

OpenAI Agents SDK

Every tool call re-sends the prior conversation into the agent's context.

1,259 → 221 tokens/turn · 82% reduction

CREWAI

CrewAI

Each crew member receives all prior crew outputs as context — token cost scales with crew size.

1,259 → 221 tokens/turn · 82% reduction

MCP

Claude Desktop, Cursor, and Continue accumulate tool outputs across the session — context window fills with repetition.

1,259 → 221 tokens/turn · 82% reduction

VERCEL AI SDK

Vercel AI SDK

streamText() receives the full message history on every request — token cost climbs with every turn.

1,259 → 221 tokens/turn · 82% reduction

Your framework isn't here yet?

The Nocturnus compression layer is framework-agnostic — you can call POST /context from any language. We're adding integrations based on what developers ask for. Open a request issue or send a pull request with your own example.