helioy

every token counts

Infrastructure for AI agents
that waste less and remember more.

Context windows are finite.

Most tokens are wasted on navigation — reading files to find files, scanning docs to find sections, re-learning what was known yesterday.

We build the tools to eliminate that waste.

Three problems. Three libraries. One orchestrator.

memory

attention-matters

Geometric recall on the S³ hypersphere. Quaternion drift, phasor interference, Kuramoto coupling. One brain per developer, not per project.

Rust · MCP npm i attention-matters
code

fmm

Structural metadata for source files. Auto-generated sidecars via tree-sitter. LLMs know what a file does without reading it.

80–90% fewer file reads
Rust · tree-sitter · MCP npm i frontmatter-matters
documents

mdcontext

Structural intelligence for markdown. Hybrid search with BM25 and semantic embeddings. Section-level indexing.

80%+ token compression
TypeScript · Effect · MCP npm i mdcontext
orchestration

nancyr

Multi-agent orchestrator wrapping Claude Code CLI. Hub-and-spoke coordination with token budgets. Compiles context. Learns from outcomes.

Rust · Claude Code · TUI

all open source. all composable.

srobinson
npx -y attention-matters serve# memory
npx -y frontmatter-matters serve# code
npx -y --package mdcontext mdcontext-mcp# docs