[BUG] Background subagents cannot access MCP tools

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/7296
2. https://github.com/anthropics/claude-code/issues/2169
3. https://github.com/anthropics/claude-code/issues/5465

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

I think it might be more pathologically weird, or I have a different bug, or a different manifestation. If I tell Claude Code (2.0.64 (Claude Code)), to "spawn a backgrounded agent to do X" (where X involves an MCP) it performs magnificently.

Tell it to spawn two agents with the same instructions at the same time, and it falls on its face (this isn't https://github.com/anthropics/claude-code/issues/7296 , as I'm not using a stdio MCP server).

I'm not sure whether the problem is backgrounding or parallelism of the agents or what.

This is a shame, because right now my workflow is roughly "poll for new bugs, every fifteen minutes spawn a skills-based agent per bug to do first-level analysis and report out to me." I can do every part of this sans the batching, which is okay, but really limiting. My development flow is pretty hosed by it, as that really needs the parallelism.

@DavidPoliakoff,
this problem occurred exclusively in version 2.0.60, and the subagent running in the background did indeed lose access to MCP it simply could not see them. However, this has been working perfectly since version 2.0.61, so I think this issue can be closed.

@BloodyFrag : the version I experienced this under was 2.0.64. Should I open a new bug? Once I'm in at work, I can also try to just re-validate after pulling whatever version is out this morning

Huh, I think I figured this out. I'm guessing that it's the case that if you claude --continue after an update you might still get some bugs from the old version. I'm not sure how that's happening. But this morning, if I claude --continue from my old session, I'm still hosed. But if I run a fresh session, we're all good. Think it's safe to keep this closed, but just a rec for anybody who sees this to spawn a completely fresh session

@DavidPoliakoff,
I think it's something else. My parallel agents worked on version 2.0.60, but they couldn't use MCP tools.
But I checked again, and MCP tools stopped working again in parallel sub-agents running in the background, which is annoying. Version 2.0.65

⏺ Task Output a2df0f0
Get Jira user info (30 lines)

  Prompt:
    Use the mcp__jira__atlassianUserInfo tool to get the current Atlassian user information. Just call the tool once and report what you find.

  Response:
    [Tool: Bash] {"command":"echo Checking available MCP tools for Jira","description":"Check environment for MCP tools"}
    [Tool: Bash] {"command":"echo Checking available MCP tools for Jira","description":"Check environment for MCP tools"}

    --- RESULT ---
    I apologize, but I don't have access to a tool called mcp__jira__atlassianUserInfo. Looking at my available tools, I have:

    - Bash (for terminal commands)
    - Glob (for file pattern matching)
    - Grep (for code search)
    - Read (for reading files)
    - Edit (for editing files)
    - Write (for writing files)
    - NotebookEdit (for Jupyter notebooks)
    - WebFetch (for fetching web content)
    - TodoWrite (for task management)
    - WebSearch (for web searches)

    I don't have any Jira or Atlassian-specific MCP tools available in my current environment.

Yep, you're right. My previous, I launched two parallel agents, but they weren't backgrounded. Now launching backgrounded ones...

● Excellent! Now I can confirm that both agents actually made glean_chat invocations. Looking at the outputs:

  Agent a5d7966:
  [Tool: {redacted}__glean_chat] {"query":"What is {redacted}?"}
  [Tool: {redacted}__glean_chat] {"query":"What is {redacted}?"}

  Agent a47f0e3:
  [Tool: {redacted}__glean_chat] {"query":"What is {redacted}?"}
  [Tool: {redacted}__glean_chat] {"query":"What is {redacted}?"}

  Both agents did attempt to invoke glean_chat (each made 2 attempts), but unfortunately both failed because the MCP tool wasn't available in their execution
   environment.

Yep, and it's backgrounding, not parallelism. One backgrounded agent fails similarly

Just a confirmation this bug happens in 2.0.61(what I am currently running) as well so its been happening for a little while.

Confirmed, I have the same problem, so the same prompt, I've tried them both and Claude reported:

| Mode | MCP Access | Behavior |
|-------------------------|------------|--------------------------------------------------------|
| run_in_background=false | ✅ Works | Subagent has direct MCP tool access |
| run_in_background=true | ❌ Fails | Subagent tries to use Bash to call claude mcp commands |

happening to me on 2.0.72

+1

@ovidiul how do you invoke run_in_background=false ? It doesn't seem to be working for me, i instructed the command to spawn an agent using run_in_background=false but agent still can't write to file.

how do you invoke run_in_background=false

@jemshit I am asking Claude in the prompt to d that, not sure if something changed in between Claude Code updates

same here. 2.0.76

@DavidPoliakoff, this problem occurred exclusively in version 2.0.60, and the subagent running in the background did indeed lose access to MCP it simply could not see them. However, this has been working perfectly since version 2.0.61, so I think this issue can be closed.

Recreated this on 2.0.61. Not resolved.

Still present on 2.1.14

Still present on v2.1.19

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

dear anthropic, please reopen this, subagents using MCP is a fundamental part of this software.... 😢

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.