Algebraic middleware for AI agent swarms. Reduce LLM costs 40–70% by replacing "send everything" with symmetry-aware context filtering — using root systems from Lie algebra, interaction knowledge graphs, and agent topology routing. Every decision is logged on Celo mainnet for verifiable, auditable agent infrastructure.
RootRouter makes autonomous AI agent infrastructure economically viable at scale. By applying algebraic geometry and graph theory to the context filtering problem and anchoring every decision on Celo's low-cost, high-throughput chain it gives any developer running agents or swarms a principled way to cut costs in half without sacrificing quality.
The first real world use case production deployment is MotusDAO, a decentralized mental health platform connecting Spanish-speaking psychologists and patients globally, where graph-structured context retrieval is particularly critical for therapeutic agent workflows. The broader mission: open-source, mathematically grounded, on-chain-verifiable infrastructure that any agent builder can use today.
AI agents are burning money. Every autonomous agent sends its full conversation history on every LLM call — a typical session accumulates 50K+ tokens, most of it irrelevant. OpenRouter routes between models but doesn't optimize what gets sent.
Agent frameworks orchestrate tasks but ignore the mathematical structure of the interaction space. The result: agents waste 40–70% of their token budget on irrelevant context, route expensive models at tasks that don't need them, and have no on-chain record of their decision-making. At scale, this makes autonomous agent infrastructure economically unviable.
RootRouter is a drop-in SDK that sits between agents and their LLM providers. It operates on three mathematical layers simultaneously:
Algebraic layer every interaction generates a root vector (intent − execution). PCA over accumulated root pairs reveals Weyl chambers: algebraically motivated regions of the interaction space derived from Lie algebra root systems. Context from the same chamber is geometrically relevant; everything else is filtered out.
Graph layer interactions form a knowledge graph (nodes: query-response pairs; edges: semantic relatedness, chamber adjacency, temporal proximity). Graph traversal retrieves multi-hop relational context that vector similarity misses. For swarms, an Agent Topology Graph tracks which agent specializes in which chamber — routing becomes a weighted shortest-path problem on this self-organizing network.
Distributed systems layer all routing decisions are logged deterministically on Celo mainnet. Chamber classifications are recomputable by anyone. Agent capabilities are registered via ERC-8004. No black-box oracle, no central trust assumption.
Result: ~49% cost reduction, zero quality loss, fully verifiable on-chain.
npm install rootrouter — one API call change, immediate savings.
open source software and public good service.
Want to grow through just grant funding
global
