Show HN: Mcptoon – Token-efficient MCP CLI client
Mcptoon, a new CLI client for Model Context Protocol (MCP) servers, claims to slash AI agent token consumption by up to 97% by translating verbose JSON into a custom "TOON" (Token-Optimized Object Notation). While the project aims to drastically reduce operational costs for AI agent interactions, the Hacker News community largely scrutinizes its token saving claims, questioning the understanding of tokenization and the practical implications of the custom format.
The Lowdown
Mcptoon is introduced as a novel CLI client designed to drastically reduce token usage for AI agents interacting with Model Context Protocol (MCP) servers. The tool proposes to replace standard, token-heavy JSON communication with a proprietary "Token-Optimized Object Notation" (TOON), aiming for significant cost savings in AI agent operations.
- The Token Problem: The author highlights that existing MCP interactions waste 30-55% of an AI agent's context window on verbose JSON syntax, particularly during tool discovery and when handling tool results.
- The TOON Solution: Mcptoon converts this verbose JSON into TOON, a more compact format. It boasts token savings of up to 97% for tool discovery and 40-60% for structured data results.
- Key Features: The client is pure Python, zero-dependency, cross-platform, and compatible with any AI agent capable of running shell commands (e.g., Claude Code, OpenAI Codex). It includes safety features to block dangerous operations, local usage tracking, and a Python API.
- TOON Mechanics: TOON achieves compaction by replacing common JSON elements with shorthand: pipes for braces/quotes/colons, spaces for brackets/commas, single characters for booleans (
T/F), a symbol for null (∅), and special characters for escape sequences (↲for\n). - Superiority Claims: Mcptoon positions itself as superior to other MCP clients by offering greater token savings, universal agent compatibility, a single configuration for all agents, and a smaller installation footprint.
In essence, mcptoon offers a lean, efficient alternative for managing AI agent tools, seeking to cut down the substantial token costs associated with current generative AI operational models.
The Gossip
Tokenization Takedown
Many commenters expressed deep skepticism about `mcptoon`'s token saving claims. They argue the author misunderstands how tokenizers work, pointing out that some TOON substitutions (like `T`/`F` for `true`/`false` or `∅` for `null`) might actually increase token counts or are unnecessary as the originals are often single tokens. The core contention is that the comparison often appears to be based on character count rather than actual token output from LLM tokenizers.
TOON's Truncated Trade-offs
The discussion questioned the fundamental design of TOON, specifically whether its aggressive compaction leads to "zero information lost." Critics highlighted that stripping away descriptions and schema from tool manifests could force LLMs into more turns or hallucination, ultimately wasting tokens and effort. There's also a debate on whether a custom format like TOON is necessary when LLMs could be trained to handle structured JSON or even leverage existing CLIs more effectively.
Agent Apathy & AI Anecdotes
Commenters aired broader frustrations with the current state of AI agent tooling and the Model Context Protocol (MCP) itself. Many described MCP projects as "vibe-coded" or lacking deep thought, echoing the sentiment that the promised token efficiency often overlooks fundamental design flaws in tool creation. There was a general cynicism towards "AI hustlers" and the rapid, often unresearched, proliferation of tools in the AI space, with some users questioning the legitimacy of new projects from new accounts.