Dev.to
7/5/2026

The original title is about Claude Code's token usage and how a developer cut their bill in half using a tool called Lynkr.
Original: Where Claude Code's Tokens Actually Go (and How I Cut My Bill in Half)
Short summary
Claude Code and similar AI coding tools waste tokens on unnecessary tool schemas, raw JSON payloads, and routing simple questions to expensive models. A developer built Lynkr, an open-source proxy that classifies requests, optimizes JSON via TOON format, caches repeated queries, and intelligently routes simple tasks to local models while reserving expensive backends for complex work. Benchmarks show ~50% token savings; setup is three commands, everything self-hosted.
- •Tool schemas and raw JSON payloads are the largest hidden token costs in agentic coding systems
- •Lynkr intelligently routes requests (local for simple, cloud for complex) and optimizes JSON, saving ~50% on typical workloads
- •Open-source proxy, self-hosted, setup in three commands
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



