mcptoon cuts MCP tool-discovery tokens ~97% with the TOON format
Builders are attacking MCP's token bloat with the TOON format — plus a Vercel AI SDK fix for reasoning-token counts and Liquid AI's phone-ready 2.6B agent.
Copy markdown
Tool discovery was quietly burning ~10,000 tokens
mcptoon, a Show HN front-runner today, routes MCP calls through TOON (a compact, token-efficient alternative to JSON) and claims tool discovery drops from ~10,000 tokens to ~350 — roughly 97% — with tool results 40–60% leaner. It's a zero-dependency ~50KB Python CLI that drops into Claude Code, Codex, Cursor, or any shell-capable agent.
TOON is turning into a standard, not a hack
The TOON format now has several MCP implementations — mcptoon, copyleftdev's toon-mcp, jellyjamin's TOON-context server — plus an open request to bake it into the MCP spec itself. If you run several MCP servers, converting their responses to TOON is the cheapest context win on the table right now.
Vercel AI SDK fixes a reasoning-token miscount
An Aug 9 patch to the SDK's openai-compatible provider clamps output text tokens at 0 when a model reports more reasoning tokens than completion tokens — so reasoning models that stop mid-thought no longer corrupt your usage and billing math. Bump your provider packages.
Liquid AI's LFM2.5-2.6B puts a real agent on a phone
Liquid's biggest on-device model yet — 2.6B open weights, 128K context, native tool calling — runs around 220 tokens/sec in under 2.5GB of RAM, down to a Raspberry Pi. Build this weekend: a fully local agent that actually calls tools, with no API bill.