TECH Signal 499
Show HN: Mcptoon – MCP CLI client that cuts tool discovery tokens by 97%
mcptoon is a lightweight CLI that replaces verbose JSON tool manifests with a compact TOON format, cutting MCP discovery tokens by about 97% and halving result tokens.
MCP-enabled agents spend a large fraction of their context window on syntax tokens, limiting the space for actual reasoning. By shrinking manifests and tool responses, mcptoon frees up 30-55% of a 128K context window, enabling longer or more complex interactions. The tool works across major agents and platforms without adding runtime dependencies.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
TOON notation reduces tool discovery tokens from ~2,000 to ~60, a 97% saving.
Zero-dependency, 50 KB Python package that runs on Windows, macOS, and Linux.
Compatible with any AI agent that can invoke shell commands, with optional JSON fallback.
THE READ
What the cluster adds up to.
MCP conversations currently waste tens of thousands of tokens just to describe available tools and to wrap each tool's output in a verbose JSON envelope. An agent that talks to several MCP servers can consume 40,000-70,000 tokens before any domain data is processed, which can consume half of a 128K context window. This overhead reduces the effective reasoning capacity of the model and inflates API costs.
mcptoon addresses the problem by providing a command-line client that talks to any MCP server and emits TOON, a token-optimized representation, instead of JSON. The manifest for 96 tools drops from roughly 2,000 tokens to 60, and structured results shrink from about 800 tokens to 350, delivering roughly a 56% reduction for full schemas. The compact format retains full semantics while dramatically lowering token consumption.
Adopting mcptoon requires installing a pure-Python package via pip, which is about 50 KB and has no external libraries. After a one-time configuration of server endpoints in a user-level config file, agents can invoke tools through simple shell commands such as `mcptoon manifest --toon` or `mcptoon call … --toon`. The tool also offers a `--json` flag for cases where standard JSON is required, preserving compatibility with existing scripts.
The client runs on Windows, macOS, and Linux and supports Python 3.10 or newer, making it broadly usable for most development environments. It works with any AI agent that can execute shell commands, including Claude Code, Codex, OpenCode, Cursor, and CatPaw, by embedding the appropriate mcptoon calls in their skill files. Agents that cannot run external commands or that rely exclusively on JSON parsing would need to fall back to the `--json` mode, which eliminates the token savings for those calls.
Compared with other MCP clients, mcptoon delivers the largest token reductions for both manifests and tool results, while keeping a single configuration shared across agents and offering multiple output formats (TOON, compact, JSON, raw). Other clients either lack token savings, require larger runtimes, or are limited to specific agents such as Claude only. The combination of minimal size, zero dependencies, and broad agent compatibility makes mcptoon a practical drop-in replacement for existing MCP tooling.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER