CrewAI + NocturnusAI
The problem: Each crew member receives all prior crew outputs as context — token cost scales with crew size.
Measured on our 15-turn product-support benchmark against Claude Opus 4.
The compression happens at the Nocturnus layer — it's framework-agnostic.
Run bench.py
against your own workload.
Copy-paste install
pip install nocturnusai crewai
docker run -d -p 9300:9300 ghcr.io/auctalis/nocturnusai:latest
export OPENAI_API_KEY=sk-...
2-minute demo
3-agent research crew writing an investment brief.
What changes in your code
Every crew member gets get_nocturnusai_tools(client, scope=TOPIC). They all read/write the same scoped knowledge base — but each only retrieves the facts relevant to its role.
from nocturnusai import SyncNocturnusAIClient
from nocturnusai.crewai import get_nocturnusai_tools
from crewai import Agent, Task, Crew, Process
with SyncNocturnusAIClient("http://localhost:9300") as client:
TOPIC = "ai-agent-infrastructure-2026"
tools = get_nocturnusai_tools(client, scope=TOPIC)
researcher = Agent(
role="Market Researcher",
goal="Identify market size and key players.",
tools=tools,
)
analyst = Agent(
role="Strategy Analyst",
goal="Identify risks and tailwinds for the context layer.",
tools=tools,
)
writer = Agent(
role="Brief Writer",
goal="Produce a 3-sentence investment brief.",
tools=tools,
)
crew = Crew(
agents=[researcher, analyst, writer],
tasks=[...],
process=Process.sequential,
)
brief = crew.kickoff()Why it works
Without Nocturnus: the Writer receives the full transcript of the Researcher’s and Analyst’s runs. With Nocturnus: each agent queries the shared knowledge base for the facts its role needs — market_size(*), key_player(*), incumbent_risk(*) — and the Writer only pulls what it’s writing about.
What’s in the repo
research_crew.py— 3-agent crew seeded with 10 facts, produces an investment briefbench.py— measure token counts across the crew with and without compressionrequirements.txt