Heroku MCP Tool Search Server
Enables dynamic tool discovery and loading for Claude using BM25, regex, and semantic search algorithms. It supports managing a catalog of over 10,000 tools with minimal context overhead and provides a REST API for tool management.
README
MCP Tool Search Server
The Problem: When you give Claude 50+ tools, two things break:
- Context bloat - Tool definitions eat 10-20K tokens, leaving less room for actual work
- Selection accuracy - Claude gets confused and picks wrong tools when there are too many
The Solution: Instead of loading all tools upfront, Claude searches for tools and loads only what it needs.
Without Tool Search: With Tool Search:
┌─────────────────────┐ ┌─────────────────────┐
│ Load 100 tools │ │ Load 1 search tool │
│ (20K tokens) │ │ (200 tokens) │
│ │ │ │
│ Claude picks wrong │ │ Claude searches: │
│ tool 30% of time │ │ "weather" → 2 tools │
│ │ │ │
│ Context nearly full │ │ 95% context free │
└─────────────────────┘ └─────────────────────┘
How It Works
- You register your tools with this server (via REST API)
- Claude gets access to search tools (
tool_search_bm25,tool_search_semantic,tool_search_regex) - When Claude needs a tool, it searches → gets back tool names → uses them
Three search methods:
- BM25 - Keyword matching ("weather" finds
get_weather,get_forecast) - Semantic - Meaning-based ("send a message" finds
send_email) - Regex - Pattern matching (
get_.*finds all getter tools)
Deploy to Heroku
Or manually:
heroku create my-tool-search
heroku buildpacks:set heroku/python
git push heroku main
heroku ai:models:create claude-4-sonnet --as INFERENCE
Register Your Tools
curl -X POST https://your-app.herokuapp.com/tools \
-H "Content-Type: application/json" \
-d '{
"name": "get_weather",
"description": "Get current weather for a location",
"input_schema": {
"type": "object",
"properties": {"location": {"type": "string"}},
"required": ["location"]
}
}'
Test It
export INFERENCE_KEY=$(heroku config:get INFERENCE_KEY -a my-tool-search)
export INFERENCE_URL=$(heroku config:get INFERENCE_URL -a my-tool-search)
curl -s "$INFERENCE_URL/v1/agents/heroku" \
-H "Authorization: Bearer $INFERENCE_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "claude-4-sonnet",
"messages": [{"role": "user", "content": "Find tools for checking weather"}],
"tools": [{"type": "mcp", "name": "mcp-tool-search/tool_search_semantic"}]
}'
API Reference
| Endpoint | Description |
|---|---|
POST /tools |
Register a tool |
GET /tools |
List all tools |
DELETE /tools/{name} |
Remove a tool |
GET /health |
Health check |
MCP Tools
| Tool | Use Case |
|---|---|
tool_search_bm25 |
Find tools by keywords |
tool_search_semantic |
Find tools by meaning |
tool_search_regex |
Find tools by pattern |
list_all_tools |
List all registered tools |
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。