Why does MCP eat so many tokens?

Every MCP tool schema is injected into the LLM context window on every request, including full JSON schemas, types, and constraints. Connecting 5 servers can burn 50k+ tokens before the agent processes anything.

How much context do MCP tools consume?

Perplexity CTO measured 40-50% of context window consumed by tool catalogs alone. One developer reported 67k tokens burned on startup with 4 MCP servers connected.

What tools fix MCP context bloat?

Context Mode (570pt HN), mcp2cli (146pt HN), MCPlexor, and Anthropic's Tool Search feature which reduced tokens 46.9% in Claude Code.

Is MCP token bloat a solved problem?

Partially. Tool Search (Anthropic) helps but is partial. No universal standard exists. Teams building more than 5 MCP connections still hit significant bloat.

How does mcp2cli reduce tokens by 99%?

It converts MCP servers into CLI tools at runtime. LLMs discover tools on demand via CLI rather than loading all schemas upfront, achieving 96-99% token reduction.

← Back to dashboard

clawsmith.com/signal/mcp-context-token-bloat

⚠ IssueUnderservedToolLive

MCP Tool Schema Context Bloat: Agents Burn 40-60% of Tokens Before Doing Any Work

Every MCP tool schema is injected into the LLM context window on every request. Connecting 5 MCP servers can load 50k+ tokens of tool definitions before the agent reasons about anything. Perplexity CTO cited 40-50% context window consumption from tool catalogs alone. Multiple tools emerged (Context Mode, mcp2cli, MCPlexor) hitting HN front page, and Claude Code shipped Tool Search to cut bloat 46.9%.

Product Idea from this Signal

A CLI tool that dynamically prunes and caches MCP tool schemas per request to cut agent context bloat

1.1k ▲

Every MCP connection injects its full tool schema into the context window before any agent reasoning happens. Connecting 5 servers burns 50k+ tokens on definitions alone, 40-60% of a typical context budget gone before the first tool call. Existing fixes are partial: Atlassian mcp-compressor does static compression but cannot do dynamic per-request selection; Claude Code Tool Search cuts only 46.9%; Context Mode cannot intercept third-party MCP tool responses. This CLI proxy sits between the agent and all connected MCP servers, dynamically selecting only the schema slice relevant to the current request using lightweight semantic matching, caching permission profiles so read-only tools auto-approve, and compressing the remaining schema at request time. The result is a single local proxy that cuts context overhead by 80-95% and eliminates approval friction for pre-cleared tool categories without requiring changes to individual MCP servers or the host client.

mcpdeveloper-toolscontext-windowtoken-optimizationcli-proxyai-agentspermission-management

Competitive148 leadsView Opportunity →