marl-middleware 1.0.0

MARL — Model-Agnostic Runtime Middleware for LLMs

MARL — Model-Agnostic Runtime Middleware for LLMs

5-stage multi-agent reasoning pipeline that reduces LLM hallucination by 70%+.

Quick Start

from marl import Marl

# OpenAI
ml = Marl.from_openai("sk-...")
result = ml.run("Your complex question")
print(result.answer)

# Any OpenAI-compatible API
ml = Marl.from_openai_compatible("https://api.x.ai/v1", "key", "grok-3")

9 Emergence Modes

from marl import Marl, MarlConfig

ml = Marl.from_openai("sk-...", config=MarlConfig(
    mode="emergence",
    emergence_type="create"  # invent|create|recipe|pharma|genomics|chemistry|ecology|law|document
))

As Proxy Server (1-line integration)

ml = Marl.from_openai("sk-...")
ml.serve(port=8080)
# Now use base_url="http://localhost:8080/v1" in any app

Pipeline

S1 Hypothesis → S2 Solver → S3 Auditor → S4 Verifier → S5 Synthesizer

Links

• Demo: HuggingFace Space
• FINAL Bench: MA=0.694 vs ER=0.302 (70%+ improvement)
• License: Apache 2.0
• Author: VIDRAFT.net