Skip to content

MCP_PROXY_USAGE.md

Latest commit

 

History

History
153 lines (113 loc) · 4.16 KB

MCP_PROXY_USAGE.md

File metadata and controls

153 lines (113 loc) · 4.16 KB

MCP Hot-Reload Proxy - Abstract Agent Team

94% Context Savings | Zero-Restart Hot-Reload | Programmatic Orchestration

Quick Start

After restarting Claude Code ONCE to load the proxy, you can now dynamically load and call tools!

The Proxy Tool

# Load any MCP server dynamically (add to .mcp.json first or use install_and_load)
load_mcp_server_dynamically("server-name")

# Install from git and load immediately
install_and_load_mcp_server("https://github.com/user/mcp-server")

# Call any tool programmatically
call_dynamic_server_tool("server-name", "tool_name", {"param": "value"})

# See what's loaded
get_loaded_servers()

Template Repository

This project IS the template repository. All other projects copy from here.

Template location: mcp_proxy_system/

Contains:

  • README.md - Complete documentation (1,000+ lines)
  • utils/dynamic_server_loader.py - Subprocess management with JSON-RPC
  • servers/proxy_server.py - 5 meta-tools for hot-reload
  • templates/ - Example servers and configurations

Creating Project-Specific Tools

Edit mcp_proxy_system/servers/proxy_server.py to expose project-specific operations:

# Import your operations
from your_module import operation1, operation2

# Register in TOOL_REGISTRY
TOOL_REGISTRY = {
    "operation1": operation1,
    "operation2": operation2,
}

# Now callable via:
call_mcp_tool("operation1", {"param": "value"})

Example Workflows

Multi-Project Server Management

# Load different project servers
load_mcp_server_dynamically("ai-tutor-proxy")
load_mcp_server_dynamically("cto-tycoon-proxy")
load_mcp_server_dynamically("simulation-proxy")

# Call tools across projects
ai_students = call_dynamic_server_tool("ai-tutor-proxy", "list_students", {})
cto_projects = call_dynamic_server_tool("cto-tycoon-proxy", "list_projects", {})
sim_agents = call_dynamic_server_tool("simulation-proxy", "list_agents", {})

# Cross-project analysis
for student in ai_students["data"]:
    # Coordinate across projects
    pass

Installing MCP Servers from Git

# Install and load in one step
install_and_load_mcp_server("https://github.com/modelcontextprotocol/servers.git", "filesystem")

# Now use the tools
files = call_dynamic_server_tool("filesystem", "list_directory", {"path": "/tmp"})

Template Development

# Test template changes
reload_mcp_server("abstract-agent-proxy")

# Verify tools available
servers = get_loaded_servers()
print(servers["servers"][0]["tools"])

Benefits

  • Context Savings: Load only what you need
  • Hot-Reload: No restart for new tools
  • Programmatic: Use tools in loops/conditions
  • Template Testing: Test proxy changes instantly
  • Cross-Project: Manage multiple project proxies

Configuration

  • .mcp.json - Proxy server configuration
  • mcp_proxy_system/ - Template implementation
  • Customize servers/proxy_server.py for project-specific tools

Deploying to Other Projects

Copy template to new project:

cp -r /Users/annhoward/src/abstract_agent_team/mcp_proxy_system /path/to/new/project/

Update new project's .mcp.json:

{
  "mcpServers": {
    "new-project-proxy": {
      "command": "python",
      "args": ["-m", "mcp_proxy_system.servers.proxy_server"],
      "cwd": "/path/to/new/project",
      "env": {
        "PYTHONPATH": "/path/to/new/project"
      }
    }
  }
}

Customize proxy_server.py for new project's operations.

Updating Template

When updating the template:

  1. Make changes to mcp_proxy_system/ in this repo
  2. Test locally: reload_mcp_server("abstract-agent-proxy")
  3. Deploy to other projects: 
    cp -r mcp_proxy_system/ /path/to/project/
  4. Reload in each project: reload_mcp_server("project-proxy")

See Also