Feature Request: MCP Client Support for Multi-Agent Integration

[chongchonghaoman]

Feature Request: MCP Client Support for Multi-Agent Integration

Problem

OpenRoom is excellent as a unified desktop AI interface! However, many users (like myself) have multiple AI agents running in parallel (e.g., one for work, one for personal tasks). We want to access both agents through a single OpenRoom conversation.

Proposed Solution

Add MCP Client support so OpenRoom can connect to external AI agents via the MCP (Model Context Protocol) standard.

Why This Matters

1. Unified Experience - One chat interface for all AI agents
2. Leverage Existing Infrastructure - Users don't need to rebuild their agents
3. MCP is a Growing Standard - Many platforms (Claude, OpenAI, etc.) are adopting MCP
4. Win-Win - OpenRoom becomes the hub for multi-agent workflows

Technical Notes

OpenRoom already has a great tool system (llmClient.ts supports tools and tool_calls). Adding MCP Client support would allow dynamic tool discovery from external MCP servers.

Implementation Suggestion

Start simple - support stdio-based MCP servers:

const mcpServers = [
  { name: 'xiao-mei', command: 'npx', args: ['-y', '@openclaw/mcp-server'] },
];

---

Happy to discuss further! I'm a MiniMax API user and would love to see OpenRoom + MiniMax become the ultimate multi-agent desktop.

[chongchonghaoman]

Clarification: What I'm Actually Proposing

I want OpenRoom to be the frontend chat interface for my OpenClaw agents. Here's the specific use case:

┌─────────────────────────────────────────────────────────┐
│  OpenRoom Browser Desktop (chat panel)                  │
│                                                         │
│  User types: "Hey Xiao Mei, help me plan a trip"       │
│              "Ming, what's my Bitcoin portfolio?"       │
│                                                         │
│         ↓  Natural language  ↓                          │
│                                                         │
│  OpenRoom AI Agent (router)                            │
│         ↓  MCP tool call  ↓                             │
│  ┌──────────────────────────────────────────┐          │
│  │  MCP Agent Bridge (my custom server)      │          │
│  │  - call_xiaomei(message) → OpenClaw     │          │
│  │  - call_ming(message) → OpenClaw        │          │
│  └──────────────────────────────────────────┘          │
└─────────────────────────────────────────────────────────┘

My agents run in OpenClaw — they have long-term memory, access to files, can run code, search the web, etc. I don't want to re-implement all that inside OpenRoom.

I just want OpenRoom's chat panel to relay messages to my OpenClaw agents and display the responses.

The MCP bridge would:

1. Expose call_xiaomei(message: string) -> string as an MCP tool
2. Expose call_ming(message: string) -> string as an MCP tool
3. OpenClaw receives the message, processes it, returns the response

This way OpenRoom handles the UI/desktop, OpenClaw handles the agent reasoning — best of both worlds!

---

Why this specific architecture:

• OpenClaw agents already have memory, tools, and complex workflows
• Re-implementing them inside OpenRoom would be huge duplication of work
• MCP is the perfect glue protocol for this
• Other OpenClaw users would benefit too!

---

TL;DR: I want OpenRoom to be like a "unified agent chat hub" that can talk to any OpenClaw agent via MCP. Not asking OpenRoom to become an agent framework — just to be the friendly face in front of existing agents.

[DanteAlighieri13210914]

Thanks, we’ll consider it. PRs are welcome!