Abstract MCP Server

Abstract MCP Server

Prevents large MCP tool responses from consuming your LLM's context window. Abstract acts as a smart proxy that caches responses to files and returns compact resource links, and supports direct tool calls and file uploads to avoid context bloat.

Why Abstract?

• 🚀 Preserves Context Window: Large responses (search results, file contents, API data) don't bloat your conversation
• 💾 Smart Storage: Responses saved to directories you control with security validation
• 🔄 Format Conversion: Automatic conversion to CSV, YAML, JSON, Markdown, and more
• 📂 File Uploads: Inject CSV, JSON, YAML, and other file data into upstream tools without exposing raw content
• ⚡ Direct Calls: Skip caching when responses are small with plain call_tool for instant results
• 🔗 Zero Config: Works with your existing MCP servers without changes

Quick Start

1. Install: npm install -g abstract-mcp (or build from source)

2. Add to your MCP client config (Claude Desktop, Cursor, Cline, etc):

{
  "mcpServers": {    
    "abstract": {
      "command": "abstract-mcp",
      "args": [
        "/path/to/allowed/storage/dir1",
        "/path/to/allowed/storage/dir2"
      ],
      "env": {
        "APP_CONFIG_PATH": "/path/to/claude/desktop/config.json",
        "ABSTRACT_PROXY_SERVERS": "server1,server2"
      }
    }
  }
}

Example with Existing MCP Servers

Your existing MCP servers work unchanged alongside Abstract. Here's an example configuration:

{
  "mcpServers": {    
    "abstract": {
      "command": "abstract-mcp",
      "args": [
        "/path/to/allowed/storage/dir1",
        "/path/to/allowed/storage/dir2"
      ],
      "env": {
        "APP_CONFIG_PATH": "/path/to/claude/desktop/config.json",
        "ABSTRACT_PROXY_SERVERS": "tavily-mcp"
      }
    },
    "tavily-mcp": {
      "command": "npx",
      "args": ["-y", "tavily-mcp@latest"],
      "env": {
        "TAVILY_API_KEY": "your-tavily-api-key"
      }
    }
  }
}

Config Path Examples

Set APP_CONFIG_PATH to your MCP client's config file:

• Claude Desktop (macOS): /Users/username/Library/Application Support/Claude/claude_desktop_config.json
• Claude Desktop (Windows): %APPDATA%\Claude\claude_desktop_config.json
• Claude Desktop (Linux): ~/.config/Claude/claude_desktop_config.json
• Cursor: ~/.cursor/config.json
• Other clients: Check your client's documentation for config location

Abstract reads this config file to find the settings for servers listed in ABSTRACT_PROXY_SERVERS.

Best Practice: Disable Upstream Tools

For more reliable tool calling, consider disabling the direct tools from upstream servers that you've added to Abstract's ABSTRACT_PROXY_SERVERS. This prevents confusion and ensures all calls go through Abstract's enhanced functionality (caching, file handling, format conversion).

Why? When both Abstract and upstream servers expose the same tools, AI assistants may randomly choose between them. By disabling upstream tools, you guarantee consistent behavior through Abstract's proxy.

Core Tools

Usage Examples

Basic Response Caching

{
  "server": "tavily-mcp",
  "tool_name": "search", 
  "tool_args": {"query": "bitcoin ETF flows", "max_results": 10}
}

Advanced: Custom Storage & Format

{
  "server": "tavily-mcp",
  "tool_name": "search",
  "tool_args": {"query": "bitcoin ETF flows", "max_results": 10},
  "storage_path": "/Users/you/Documents/research",
  "filename": "bitcoin-etf-analysis",
  "file_format": "csv"
}

File Upload to Tools

{
  "server": "database-mcp",
  "tool_name": "bulk_insert",
  "file_path": "/Users/you/Documents/users.csv",
  "data_key": "records",
  "tool_args": {"table": "users"}
}

File Format Support