GitHub Repository

Token-optimized AI agent runtime with 12 providers, multi-agent DAG, TOON compression & encrypted vault .

5 starsTypeScript

Beta-Claw – I built an AI agent runtime that cuts token costs by 44%

Name: Beta-Claw – I built an AI agent runtime that cuts token costs by 44%
Availability: InStock
Author: Kyoiske

by Kyoiske·Mar 9, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●MidBold BetBig Brain

Custom TOON format saves tokens but LangChain and CrewAI already solve orchestration.

Strengths

•TOON serialization genuinely reduces token costs by 28-44% versus standard JSON.
•Four-tier smart model routing picks cheapest provider that handles each task.
•AES-256-GCM encrypted vault never stores secrets in plaintext on disk.

Weaknesses

•Zero GitHub stars indicates embryonic project competing against established frameworks.
•Multi-agent orchestration space is extremely crowded with LangChain, CrewAI, AutoGen.

Post Description

I built Beta-Claw during a competition and kept pushing it after because I genuinely think the token waste problem in AI agents is underrated.

The core idea: most agent runtimes serialize everything as JSON. JSON is great for humans but terrible for tokens. So I built TOON (Token-Oriented Object Notation) — same structure, 28–44% fewer tokens. At scale that's millions of tokens saved per day.

What else it does: → Routes across 12 providers (Anthropic, OpenAI, Groq, Ollama, DeepSeek, OpenRouter + more) → 4-tier smart model routing — picks the cheapest model that can handle the task → Multi-agent DAG: Planner → Research → Execution → Memory → Composer → Encrypted vault (AES-256-GCM), never stores secrets in plaintext → Prompt injection defense + PII redaction built in → 19 hot-swappable skills, < 60ms reload → Full benchmark suite included — 9ms dry-run pipeline latency

It's CLI-first, TypeScript, runs on Linux/Mac/WSL2.

Repo: https://github.com/Rawknee-69/Beta-Claw

Still rough in places but the core is solid. Brutal feedback welcome.