[BUG] Subagents do not receive MCP server instructions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I use an MCP server (my-cool-proxy) that relies on instructions to preload usage guidelines generically across multiple agent harnesses. This works perfectly in the main agent, but if I try to use it in subagents, they can't see instructions at all, forcing me to include redundant (and easily outdated) information in agent prompts to make up for that.

What Should Happen?

If a subagent has access to MCP tools from a particular server, it should also receive the instructions field from that server.

Error Messages/Logs

{"parentUuid":null,"isSidechain":true,"userType":"external","cwd":"/home/karashiiro/my-cool-proxy","sessionId":"f1176ff3-cbfa-438c-a8e9-63d0623b106d","version":"2.1.63","gitBranch":"dev","agentId":"a1d6d78a247e4c703","slug":"jazzy-finding-koala","type":"user","message":{"role":"user","content":"IMPORTANT: Do NOT make any tool calls. Answer only from what's in your system context.\n\nCan you see any of the following in your context? Answer yes or no for each, and if yes, quote the relevant text:\n\n1. A section called \"# MCP Server Instructions\"\n2. A section called \"# Available Gateway Skills\" that lists skills like \"writing-gateway-skills\", \"commit-style\", \"creating-a-new-project\", or \"investigating-open-source-projects\"\n3. Any text that says \"Use `gw-skill://{skill-name}` to load skill instructions\"\n4. Any server-specific instructions from servers like \"github\", \"context7\", or \"my-cool-proxy\"\n5. The text \"This is an MCP gateway proxy that aggregates multiple MCP servers\""},"uuid":"08295c01-fd19-4b45-8fd7-4665cc71199b","timestamp":"2026-02-28T19:45:54.452Z"}
{"parentUuid":"08295c01-fd19-4b45-8fd7-4665cc71199b","isSidechain":true,"userType":"external","cwd":"/home/karashiiro/my-cool-proxy","sessionId":"f1176ff3-cbfa-438c-a8e9-63d0623b106d","version":"2.1.63","gitBranch":"dev","agentId":"a1d6d78a247e4c703","slug":"jazzy-finding-koala","message":{"model":"claude-sonnet-4-6","id":"msg_01HNn1RFgmo1ndDQCznNDtdz","type":"message","role":"assistant","content":[{"type":"text","text":"Here are my answers for each item:\n\n1. **No.** There is no section called \"# MCP Server Instructions\" in my context.\n\n2. **No.** There is no section called \"# Available Gateway Skills\" listing those skills in my context.\n\n3. **No.** There is no text saying \"Use `gw-skill://{skill-name}` to load skill instructions\" in my context.\n\n4. **No.** There are no server-specific instructions from servers like \"github\", \"context7\", or \"my-cool-proxy\" in my context.\n\n5. **No.** The text \"This is an MCP gateway proxy that aggregates multiple MCP servers\" does not appear in my context.\n\nThe only project-level context I have is the contents of `/home/karashiiro/my-cool-proxy/CLAUDE.md`, which describes the project as \"an MCP (Model Context Protocol) server gateway that proxies between agents and multiple MCP servers\" — but that is not the exact phrase you listed in item 5."}],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":3,"cache_creation_input_tokens":2246,"cache_read_input_tokens":13011,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":2246},"output_tokens":1,"service_tier":"standard","inference_geo":"not_available"}},"requestId":"req_011CYb3QN6LrrbGt271ebDmF","type":"assistant","uuid":"b27b6e5c-9954-4da5-b0e4-ae804193692c","timestamp":"2026-02-28T19:46:00.025Z"}

Steps to Reproduce

1. Connect to any MCP server with instructions - I use my own (my-cool-proxy), but the GitHub MCP Server will work as well.
2. Create a subagent with access to tools from that server.
3. Have Claude ask the subagent to quote any information from the instructions - include some sample quotes to verify.

This is the subagent I was using, but any will do: research-strategist.md

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.63

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Windows Terminal

Additional Information

I suspect this happens because subagents don't really have a concept of MCP servers in general, they only have a tool allowlist - but I would expect that if a subagent is given access to tools from an MCP server, it would also get the instructions from that server to guide it.