Beta-Claw – I built an AI agent runtime that cuts token costs by 44%
Custom TOON format saves tokens but LangChain and CrewAI already solve orchestration.
Convert Java to JSON/TOON and back. Supports complex object graphs, cyclic references, and TOON format for 40-50% LLM token savings
TOON format cuts LLM token costs 40-50%, but this is just Java support for an existing format.
Java developers building LLM applications
Jackson · Gson · JToon
Same data as JSON, ~40-50% fewer tokens. No braces, brackets, commas, or unnecessary quoting. Arrays of uniform objects become CSV-like tables automatically.
JSON: {"team":"Rockets","players":[{"name":"John","age":30},{"name":"Sue","age":27}]}
TOON: team: Rockets players: name, age John, 30 Sue, 27
json-io reads/writes JSON, JSON5, and TOON with 60+ built-in Java types, cycle support, and zero config. There's also a Spring Boot starter and Spring AI module.
GitHub: https://github.com/jdereg/json-io TOON spec: https://toonformat.dev Baeldung article: https://www.baeldung.com/java-json-toon-format-libraries
Custom TOON format saves tokens but LangChain and CrewAI already solve orchestration.
LLM token savings angle is interesting, but binary formats are crowded.
Token efficiency beats Stagehand — 2-5k vs 29-51k per action with cached selectors.
MAKO compresses what matters into a HEAD-friendly payload — frontmatter, declared actions and semantic links — so agents can find relevance without downloading 181KB of navigation, ads and scripts. The project ships a spec plus real tooling (typed SDK, Express middleware, an analyzer/score and edge-friendly /md conversion), which is a rare combo of protocol thinking and usable developer ergonomics. Whether it becomes a standard depends on buy-in from CMS/plugin authors and agent platforms, but technically it's a smart, practical swing at an obvious pain point.
Artifact-backed tool output cuts token usage 95%, accuracy jumps 33% to 99%.
Language purpose-built for token costs: 55 tokens vs 120 in JavaScript. Real compiler, 1291 tests.