Defining MCP servers through IPC socket API

[m-ibm]

App Version

v3.26.1

API Provider

OpenAI Compatible

Model Used

llama 4 maverick

Roo Code Task Links (Optional)

No response

🔁 Steps to Reproduce

I am trying to create a headless VSCode with RooCode (used via docker).
I translated https://github.com/cte/roo-cli to Python and followed the same guidelines -- this works (mostly) with tasks that do not require MCP/tool calls.

I am using a configuration like this:

configuration = {
	"apiProvider": "openai",
	"openAiApiKey": os.getenv("OPENAI_API_KEY", ""),
	"openAiBaseUrl": os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
	"openAiModelId": os.getenv("OPENAI_MODEL", "gpt-4o-mini"),
	"autoApprovalEnabled": True,
	"alwaysAllowReadOnly": True,
	"alwaysAllowWrite": True,
	"alwaysAllowBrowser": True,
	"alwaysApproveResubmit": True,
	"alwaysAllowMcp": True,
	"alwaysAllowModeSwitch": True,
	"alwaysAllowSubtasks": True,
	"alwaysAllowExecute": True,
	"allowedCommands": ["*"],
	"browserToolEnabled": False,
	"enableCheckpoints": False,
	"mcpEnabled": True,
	"mode": "architect",
}

mcps = os.getenv('ROO_CODE_MCP_SERVERS')
if mcps:
	configuration['mcpServers'] = json.loads(mcps)

where MCPs and provider details are read from .env:
ROO_CODE_MCP_SERVERS={"addi":{"type":"sse","url":"http://localhost:8004/sse","headers":{},"alwaysAllow":[],"disabled":false}}

This is the same MCP server format I would use in mcp.json.
As far as I can tell, no MCPs are defined; certainly my tools aren't called.

💥 Outcome Summary

Expected tools to be called (same task works with predefined MCPs), but nothing happens

📄 Relevant Logs or Errors (Optional)

[m-ibm]

PR looks good to me

[daniel-lxs]

Closed the PR, it didn't seem like a sound implementation, it would be best if someone implemented this.

I think it makes sense passing a list of MCPs via environment variables