mcp-proxy-daemoning

MCP Proxy Daemoning

Pattern for running a shared MCP proxy daemon that multiple Claude Code sessions connect to simultaneously, reducing resource overhead and enabling cross-session tool sharing.

The Problem

When multiple Claude Code sessions each spawn their own MCP server processes:

• Resource waste: N sessions × M servers = N×M processes
• No shared state: Each session has its own isolated tool instances
• Startup overhead: Every new session pays the full initialization cost
• Serena duplication: A project-level MCP server like Serena that should maintain knowledge across projects gets duplicated per-session, losing cross-project context

The Solution: Daemon Architecture

A single daemon process runs the MCP proxy, and multiple sessions connect to it via HTTP:

Session A ──stdio──→ run-mcp-proxy.sh ──HTTP──→ mcp-proxy daemon (port 12476)
Session B ──stdio──→ run-mcp-proxy.sh ──HTTP──→     (same daemon)
Session C ──stdio──→ run-mcp-proxy.sh ──HTTP──→     (same daemon)

Reference Implementation (nsheaps/aitkit)

The run-mcp-proxy.sh script in nsheaps/aitkit implements a full daemon lifecycle:

Configuration

MCPPROXY_TOP_K="10"                              # Tools to register per session
MCPPROXY_TOOLS_LIMIT="50"                        # Max tools in active pool
MCPPROXY_DATA_DIR="$AITKIT_ROOT/lib/mcpproxy"    # Database/index location
MCPPROXY_CONFIG_PATH="$MCPPROXY_DATA_DIR/mcp_proxy_config.json"
MCPPROXY_PORT="12476"                            # Listen port
MCPPROXY_TRANSPORT="streamable-http"              # Transport type

Directory Structure

~/.aitkit/
├── run/                          # PID files for daemon management
│   ├── mcp-proxy-<name>.pid     # Daemon PID
│   └── mcp-servers/             # Per-session registration
│       ├── 12345                 # Session A's PID
│       └── 12346                 # Session B's PID
└── lib/mcpproxy/
    ├── mcp_proxy_config.json    # MCP server definitions
    ├── mcp-proxy.log            # Daemon log
    └── (database/index files)   # Smart proxy state

Daemon Lifecycle

1. First session starts → checks if daemon running → starts daemon
2. Subsequent sessions → reuse existing daemon, register their PID
3. Session exits → unregisters PID, checks if last session
4. Last session exits → kills daemon

# Simplified lifecycle logic
start_or_reuse_daemon() {
    if daemon_is_running; then
        echo "Reusing existing daemon at PID $(daemon_pid)"
    else
        daemon --name "$DAEMON_NAME" --pidfile "$PID_FILE" \
            --command "uvx smart-mcp-proxy \
                --port $MCPPROXY_PORT \
                --transport $MCPPROXY_TRANSPORT \
                --config $MCPPROXY_CONFIG_PATH"
    fi
    # Register this session
    echo $$ > "$MCP_PID_DIR/$$"
}

on_exit() {
    rm -f "$MCP_PID_DIR/$$"
    if [ "$(ls "$MCP_PID_DIR" | wc -l)" -eq 0 ]; then
        kill_daemon
    fi
}
trap on_exit EXIT

MCP Proxy Config (mcp_proxy_config.json)

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
    },
    "serena": {
      "command": "uvx",
      "args": ["serena", "--workspace", "/path/to/project"]
    }
  }
}

Smart Proxy Features

The smart-mcp-proxy (via uvx) provides:

• BM25 tool indexing: Only registers top-K relevant tools per session
• Active tool pooling: Limits active tools to prevent context bloat
• Tool output truncation: Prevents oversized responses
• Dynamic routing: Routes tool calls to appropriate backend servers

The Serena De-Duplication Pattern

Problem

Serena is a code-aware AI agent tool that stores project knowledge in .serena/memories/. When configured at the project level:

• Each project gets its own Serena instance
• If Serena should track knowledge across projects (e.g., for a monorepo or related repos), duplicating it per-project defeats the purpose
• The hardcoded path pattern (/Users/nheaps/.aitkit/upstreams/gathertown/.ai/bin/mcp/serena.sh) in per-project configs is fragile and non-portable

Solution: User-Level Daemon with Project Registration

1. Run Serena at user level via the mcp-proxy daemon (not per-project)
2. Register projects with the daemon config, not in each project's .mcp.json
3. De-duplicate: If a project's .mcp.json declares Serena AND the user-level daemon already runs it, the daemon version takes precedence
4. Cross-project memory: Serena's memories persist in the daemon's data directory, shared across all sessions

# Instead of this (per-project, fragile):
# project/.mcp.json → serena.sh → standalone serena process

# Do this (user-level, shared):
# ~/.aitkit/lib/mcpproxy/mcp_proxy_config.json includes serena
# All sessions connect to the shared daemon
# Serena runs once, with access to all registered projects

Implementation Steps

1. Add Serena to the mcp-proxy config at ~/.aitkit/lib/mcpproxy/mcp_proxy_config.json
2. Remove Serena from individual project .mcp.json files
3. The daemon manages Serena's lifecycle
4. All sessions access Serena through the shared daemon's HTTP endpoint

Integrating with Claude Code

In your project's .mcp.json or user-level settings, point to the daemon wrapper:

{
  "mcpServers": {
    "shared-proxy": {
      "type": "stdio",
      "command": "/path/to/aitkit/bin/run-mcp-proxy.sh",
      "args": ["mcp-user"]
    }
  }
}

The wrapper script:

1. Starts the daemon if not running
2. Registers this session
3. Proxies stdio ↔ HTTP between Claude Code and the daemon
4. Cleans up on exit

Troubleshooting

Issue	Resolution
Daemon not starting	Check ~/.aitkit/lib/mcpproxy/mcp-proxy.log
Stale PID files	Remove ~/.aitkit/run/mcp-servers/* and ~/.aitkit/run/mcp-proxy-*.pid
Port conflict	Change MCPPROXY_PORT in the config
Tool not found	Verify mcp_proxy_config.json includes the server
Daemon crashes between sessions	Background health loop auto-recovers; check logs