Why does MCP consume so many tokens just for tool definitions?

The underlying chat-completions API is stateless so MCP host clients must re-inject the full tool catalog (name, description, JSON schema) on every single request. With 10 MCP servers that is easily 10,000 to 60,000 tokens per turn before any work begins.

How much context does a typical MCP deployment consume?

The GitHub MCP server alone uses 55,000 tokens for its 93 tool definitions. One developer measured 66,000 tokens burned at conversation start - one third of Claude's 200k context window.

What is mcp2cli and how does it reduce token overhead?

mcp2cli converts MCP server tool schemas into CLI commands at runtime. Agents discover tools via --list (~16 tokens per tool) and --help as needed, instead of injecting all schemas upfront. Reduces overhead 96-99%.

Does lazy loading solve the MCP token tax?

Partially. Claude Code now supports on-demand tool loading reducing overhead 85%. Gateway proxies expose 4 lightweight tools instead of 200+ schemas. But there is no official MCP protocol standard for lazy loading yet.

What did Perplexity find about MCP context overhead?

Perplexity dropped MCP internally citing 72% context window waste from tool definitions before agents could do any real work.

← Back to dashboard

clawsmith.com/signal/mcp-tool-schema-token-tax-per-request

⚠ IssueUnderservedai_agent_mcpLive

MCP forces agents to reload all tool schemas on every request burning 10k to 60k tokens before any work

The MCP protocol is stateless at the API level so host clients must re-serialize all connected tool definitions on every turn. With 10+ MCP servers agents burn 10,000 to 60,000 tokens on tool catalogs. One team measured 143k of 200k context tokens consumed before a single message. mcp2cli (2,200+ stars) and multiple gateway projects show strong demand for a fix.

Product Idea from this Signal

A web app that proxies MCP servers and lazy-loads tool schemas on demand to eliminate per-request token overhead

2.6k ▲

Every MCP-connected agent today loads the full schema for every registered tool on every request, burning thousands of tokens before a single line of useful work runs. Teams operating agents at scale pay for this overhead on every call, and their context windows fill up with schema definitions that are rarely needed. A hosted proxy layer sits between the agent and its MCP servers, exposes a small discovery interface, and only injects a tool's full schema into context when the agent actually needs to call it.

mcpai-agentstoken-optimizationdeveloper-toolsinfrastructurecontext-windowproxy

Competitive73 leadsView Opportunity →

Score Breakdown

GitHub

2,382

177

Issues

Social Proof 3 sources

mcp2cli: Turn any MCP server or OpenAPI spec into a CLI with 96-99% fewer tokens

gh:knowsuchagency · 3/9/2026

2,382 HN

Show HN: Mcp2cli - One CLI for every API, 96-99% fewer tokens than native MCP

knowsuchagency · 3/9/2026

177 GH

Lazy-load MCP tool definitions to reduce context usage

gh:davidmoneil · 11/10/2025

Existing Solutions 3 competitors

mcp2cli2,227 stars

Converts MCP servers to CLI tools reducing token overhead 96-99%

mcp-gateway (RaiAnsar)1 star, very early

Lazy-loading proxy that loads MCP servers on demand instead of upfront

Claude Code MCP Tool SearchBuilt into Claude Code

Anthropic native on-demand tool loading reducing overhead 85%

Gap Assessment

UnderservedExisting solutions leave gaps

mcp2cli (2,227 stars) converts tools to CLI reducing overhead 96-99% but requires CLI wrappers. Multiple gateway solutions emerging but none is the official standard.

Frequently Asked Questions

Virality Score

2,563

across 0 platforms

Details

Signalissue

Ecosystemai_agent_mcp

Sources3

Platforms0

Updated1d ago

Trend→ stable

Top ideas

All ideas →

0A CLI tool that wraps stateful MCP servers and externalizes their session state so they run behind standard round-robin load balancers without sticky routing 0A CLI tool that tails and greps logs across multiple remote hosts in a single TUI without centralized infrastructure 0A browser extension that removes Gemini and all Google AI injections from Chrome across Search, Gmail, Docs, and Drive in one toggle

Related signals

All signals →

65.5KAI Agent Session Amnesia: Agents Forget Everything Between Sessions by Design 923MCP Tool Schema Context Bloat: Agents Burn 40-60% of Tokens Before Doing Any Work 613Pipelock: Open-Source AI Agent Firewall — 11-Layer Scanner Pipeline, MCP Security, Apache 2.0