can subagents access MCP deferred tools?

[GaryIsHere]

Can sub-agents use deferred MCP tools?

Environment

• MCP server: 烯牛MCP (HTTP) configured in ~/.deepseek/mcp.json
• Tools provided: mcp_烯牛MCP_get_company_info, mcp_烯牛MCP_get_data, etc.
• These tools are deferred — schema is loaded on first invocation.

What works

Calling the MCP tool directly from the parent agent:

→ mcp_烯牛MCP_get_company_info(firm_name="小米科技", aspect="融资状况分析")
← (business result from xiniu API) ✅

What doesn't work

Experiment 1: Sub-agent without allowed_tools

agent_open(prompt="call mcp_烯牛MCP_get_company_info", type="general", fork_context=false)

Result: sub-agent reports "Tool mcp_烯牛MCP_get_company_info is not registered" ❌
(Tried 5 naming variants — same error for all.)

Experiment 2: Sub-agent with allowed_tools explicitly listing the MCP tool

agent_open(
  prompt="call mcp_烯牛MCP_get_company_info",
  type="general",
  fork_context=false,
  allowed_tools=["mcp_烯牛MCP_get_company_info"]
)

Result: sub-agent fails to start with:

Failed: Sub-agent requested unavailable tools: mcp_烯牛MCP_get_company_info

The framework rejects the tool name at spawn time — it never even reaches the sub-agent. ❌

The core problem

MCP deferred tools are available in the parent agent's tool registry, but they cannot be passed to sub-agents — not even via explicit allowed_tools. The sub-agent framework treats them as "unavailable."

Questions

1. Is this by design, or is it a bug/limitation?
2. If by design: what is the intended way for sub-agents to access external MCP servers? Do sub-agents have their own MCP configuration path, or must all MCP interactions go through the parent agent?
3. If this is a limitation that will be fixed: is there a timeline or workaround in the meantime?

Context

I'm building a find-reach skill where the parent agent dispatches sub-agents to search different data sources (MCP, REST APIs, etc.) in parallel. The current limitation means I can't delegate MCP queries to sub-agents — the parent agent must do them all sequentially, which defeats the purpose of the agent_open parallelism model.