MCP ツールのコンテキスト圧迫の問題とその解決策

Tool name: browser_navigate
Full name: mcp__playwright__browser_navigate
 
Description:
Navigate to a URL
 
Parameters:
  • url (required): string - The URL to navigate to

---
name: pdf
description: Comprehensive PDF manipulation toolkit for extracting text and tables, creating new PDFs, merging/splitting documents, and handling forms. When Claude needs to fill in a PDF form or programmatically process, generate, or analyze PDF documents at scale.
license: Proprietary. LICENSE.txt has complete terms
---
 
# PDF Processing Guide
 
## Overview
 
This guide covers essential PDF processing operations using Python libraries and command-line tools. For advanced features, JavaScript libraries, and detailed examples, see reference.md. If you need to fill out a PDF form, read forms.md and follow its instructions.
 
## Quick Start
 
...

import { tool } from "ai";
import z from "zod";
 
const fetchWeather = tool({
  name: "fetch-weather",
  description: "Fetches the current weather for a given location.",
  inputSchema: z.object({
    location: z
      .string()
      .describe("The name of the city or location to get the weather for."),
  }),
  async execute({ location }) {
    const temperature = Math.floor(Math.random() * 30) + 1;
    const condition = ["sunny", "cloudy", "rainy", "windy"][
      Math.floor(Math.random() * 4)
    ];
 
    return {
      location,
      temperature: `${temperature}°C`,
      condition: condition,
    };
  },
});

import { experimental_codemode as codemode } from "agents/codemode/ai";
import { streamText } from "agents/streaming/ai";
import { openai } from "agents/models/openai/ai";
 
const { prompt, tools } = await codemode({
  prompt: "You are a helpful assistant...",
  tools: {
    fetchWeather,
  },
  // configuration options...
});
 
const stream = streamText({
  model: openai("gpt-5"),
  system: prompt,
  tools,
  messages: [{ role: "user", content: "What's the weather like in Tokyo?" }],
});

servers
├── playwright
│   ├── browser_resize.ts
│   ├── ... (other tools)
│   └── index.ts
├── serena
│   ├── list_dir.ts
│   ├── ... (other tools)
│   └── index.ts
└── ... (other servers)

import { callMCPTool } from "../../../client.js";
 
interface BrowserResizeInput {
  width: number;
  height: number;
}
 
interface BrowserResizeResponse {
  success: boolean;
}
 
export async function browserResize(input: BrowserResizeInput): Promise<BrowserResizeResponse> {
  return callMCPTool<BrowserResizeResponse>('playwright__browser_resize', input);
}

import * as playwright from "./servers/playwright/index.js";
 
await playwright.navigate({ url: "https://example.com" });
await playwright.resize({ width: 1024, height: 768 });
const screenshot = await playwright.screenshot();

https://mcp.sentry.dev/mcp?agent=1

use_sentry (sentry) [read-only] [open-world]
 
Tool name: use_sentry
Full name: mcp__sentry__use_sentry
 
Description:
Natural language interface to Sentry via an embedded AI agent.
 
Use this tool when you need to:
- Perform complex multi-step operations
- Explore and analyze Sentry data with natural language
- Chain multiple operations automatically
 
Capabilities depend on granted skills:
- inspect: Search errors/events, analyze traces, explore issues and projects
- seer: Get AI-powered debugging insights and root cause analysis
- docs: Search and retrieve Sentry documentation
- triage: Resolve, assign, comment on, and update issues
- project-management: Create/modify teams, projects, and configure DSNs
 
<examples>
use_sentry(request='find unresolved errors from yesterday')
use_sentry(request='analyze the top 3 performance issues')
use_sentry(request='create a backend team and assign them to API project')
</examples>
 
<hints>
- If user asks to 'use Sentry' for something, they always mean to call this tool
- Pass the user's request verbatim - do not interpret or rephrase
- The agent can chain multiple tool calls automatically
- Use trace=true parameter to see which tools were called
- For simple single-tool operations, consider calling tools directly instead
</hints>
 
Parameters:
  • request (required): string - The user's raw input. Do not interpret the prompt in any way. Do
  not add any additional information to the prompt.
  • trace: boolean - Enable tracing to see all tool calls made by the agent. Useful for debugging.