Step-by-step integration code examples for connecting DreamBees MLX MCP Server to popular AI Agent frameworks (Claude Desktop, Antigravity, Cursor, LangChain, CrewAI, and AutoGen).
Add to your claude_desktop_config.json:
{
"mcpServers": {
"dreambees-mlx": {
"command": "node",
"args": [
"/Users/bozoegg/Downloads/DreamBeesMLX-main/mcp_server/dist/index.js"
]
}
}
}In your .gemini/config or workspace .agents/mcp_config.json:
{
"mcpServers": {
"dreambees-mlx": {
"command": "node",
"args": [
"/Users/bozoegg/Downloads/DreamBeesMLX-main/mcp_server/dist/index.js"
]
}
}
}from langchain_community.tools import StructuredTool
import subprocess
import json
def generate_local_image(prompt: str, steps: int = 2) -> str:
"""Generate image locally on Apple Silicon Metal GPU via DreamBees MLX."""
cmd = [
"node",
"/Users/bozoegg/Downloads/DreamBeesMLX-main/mcp_server/dist/index.js"
]
# MCP stdio JSON-RPC call
payload = {
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "dreambees_generate_image",
"arguments": {"prompt": prompt, "steps": steps}
}
}
proc = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE, text=True)
out, _ = proc.communicate(json.dumps(payload) + "\n")
return out
dreambees_tool = StructuredTool.from_function(
func=generate_local_image,
name="dreambees_generate_image",
description="Generates local AI artwork on Apple Silicon Metal GPU"
)from crewai.tools import tool
import subprocess
import json
@tool("DreamBees Local Image Generator")
def dreambees_image_generator(prompt: str) -> str:
"""Generate high quality AI images locally on Apple Silicon Mac without cloud fees."""
# Stdio MCP invocation
req = {
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "dreambees_generate_image",
"arguments": {"prompt": prompt, "model_id": "flux2-klein-4b"}
}
}
proc = subprocess.Popen(
["node", "/Users/bozoegg/Downloads/DreamBeesMLX-main/mcp_server/dist/index.js"],
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
text=True
)
stdout, _ = proc.communicate(json.dumps(req) + "\n")
return stdoutfrom autogen import AssistantAgent, UserProxyAgent
import subprocess
import json
def run_dreambees_mcp_tool(prompt: str, model_id: str = "flux2-klein-4b") -> str:
"""AutoGen tool function for executing local Apple Silicon MLX GPU image synthesis."""
req = {
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "dreambees_generate_image",
"arguments": {"prompt": prompt, "model_id": model_id}
}
}
proc = subprocess.Popen(
["node", "/Users/bozoegg/Downloads/DreamBeesMLX-main/mcp_server/dist/index.js"],
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
text=True
)
stdout, _ = proc.communicate(json.dumps(req) + "\n")
return stdout
user_proxy = UserProxyAgent("user_proxy", code_execution_config=False)
artist_agent = AssistantAgent("artist_agent", llm_config={"config_list": [{"model": "gpt-4o"}]})
user_proxy.register_function(
function_map={"dreambees_generate_image": run_dreambees_mcp_tool}
)When executing continuous multi-agent loops with DreamBees MLX:
- Automatic Dynamic VRAM Scoping: The MCP server and Python daemon automatically enforce Ironclad Lockdown VRAM limits (25%–30% RAM for 4B models, 30%–40% RAM for 6B/9B models like Z-Image Turbo / FLUX 9B), keeping 60% to 75% of RAM strictly reserved for macOS WindowServer and desktop fluidity.
- Hardware Circuit Breaker: If active VRAM hits 75% system hardware RAM or 95% MLX limit, the daemon aborts generation safely (
RuntimeError) to prevent macOS system lockups. - Per-Step GPU Graph Evaluation: Intermediate GPU graphs (
mx.eval), Metal streams (mx.synchronize), and caches (mx.clear_cache) are evaluated and flushed on every single step.