[BUG] claude.ai MCP connectors not available in -p / --print mode

[PrincevilleCapital]

Preflight Checklist

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

What's Wrong?

MCP servers connected via claude.ai native connectors (HTTP-based, OAuth-authenticated) show as "Connected" in claude mcp list but are completely unavailable when Claude Code runs in any non-interactive mode. The init message in stream-json output shows "mcp_servers":[], and the model has zero access to MCP tools.

This affects all programmatic invocation paths: --print, --output-format stream-json, and piped stdin.

What Should Happen?

MCP servers that are connected and healthy (verified via claude mcp list) should also be available when using -p / --print mode or --input-format stream-json. The mcp_servers array in the stream-json init message should list the same servers that appear in interactive mode.

Error Messages/Logs

Steps to Reproduce

1. Confirm MCP servers are connected:
   claude mcp list
   Output:
   rube
   Type: project
   Status: connected
   Protocol: HTTP (Streamable)
   URL: https://rube.app/mcp
   Tools: 11

claude_ai_Rube
Type: claude_ai
Status: connected
Protocol: HTTP (Streamable)
Tools: 11

claude_ai_Granola
Type: claude_ai
Status: connected
Protocol: HTTP (Streamable)
Tools: 4

claude_ai_Slack
Type: claude_ai
Status: connected
Protocol: HTTP (Streamable)
Tools: 12

2. Run in print mode -- no MCP tools available:
   claude -p "List every tool you have access to that starts with mcp" --dangerously-skip-permissions
   Result: Model lists only 17 core tools (Read, Write, Edit, Bash, Glob, Grep, etc.). Zero MCP tools.

3. Verify via stream-json init message:
   claude -p "hello" --output-format stream-json --verbose --dangerously-skip-permissions
   Init message contains:
   {"type":"system","subtype":"init","mcp_servers":[], ...}

4. Also tested and confirmed broken:
   claude --input-format stream-json with JSON piped via stdin
   echo "test" | claude (piped stdin, non-TTY)
   claude -p --mcp-config '{"mcpServers":{"rube":{"type":"http","url":"https://rube.app/mcp"}}}' -- tools appear in --mcp-config listing but mcp_servers:[] persists in init

5. Confirm it works interactively: Launch claude in a TTY terminal (VSCode integrated terminal). Same machine, same config. MCP servers connect and all tools are available.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

No response

Claude Code Version

2.1.14

Platform

Other

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

Claude subscription (not API key)
MCP server type: claude_ai (HTTP, OAuth-authenticated via claude.ai)

Impact
This blocks programmatic use of Claude Code for any workflow that depends on MCP tools. Our use case: wrapping existing CLI slash commands with a Slack bot so a team can invoke them without terminal access. The commands require Rube MCP (Dropbox, Gmail, Calendar, Sheets) to function.

[github-actions]

Found 3 possible duplicate issues:

1. [BUG] Claude.ai MCPs not showing up on Claude code #21874
2. [BUG] Official connectors (e.g., Supabase) not available as tools in agentic/code mode #23223
3. Claude.AI MCP not available from terminal emulator #24062

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

[PrincevilleCapital]

The duplicate issues are about the feature gate being false (MCPs never load anywhere). Our problem is different -- gate is true interactively, but -p mode ignores MCP entirely.

[ChristopherMacGown]

Additional reproduction and root cause analysis

Independently reproduced on macOS (v2.1.47) with a precise control test isolating the bug to --mcp-config / scope: "dynamic" handling.

Reproduction

An HTTP MCP server at localhost:8788/mcp was configured both as a user-scoped global server (claude mcp add) and via --mcp-config with an identical JSON config file.

Works — globally configured (scope: user):

claude -p --dangerously-skip-permissions "List your mcp__ tools"
# → returns MCP tools instantly (~3s total)

Hangs — same server via --mcp-config (scope: dynamic):

claude -p --dangerously-skip-permissions \
  --mcp-config '{"mcpServers":{"metacog":{"type":"http","url":"http://localhost:8788/mcp"}}}' \
  "List your mcp__ tools"
# → hangs for exactly 120s, then completes with zero MCP tools

Same server, same URL, same transport type, same machine. The only difference is the config source.

Findings from tracing the binary

• --mcp-config entries are parsed and tagged with scope: "dynamic" via uRT(servers, s => ({...s, scope: "dynamic"})) before being merged into the server map
• Globally configured servers carry scope: "user" / "local" / "project"
• Both are merged into the same server map (w_ = {...L$, ...YT}) and passed to the connection function (CER())
• In non-interactive mode (-p), CER() is awaited synchronously — it hangs for 120s on scope: "dynamic" servers but connects instantly to identical scope: "user" servers
• The MCP server never receives a connection attempt during the 120s hang (confirmed via server logs)
• --strict-mcp-config does not help — it bypasses config file loading but --mcp-config entries still get scope: "dynamic"
• --debug-to-stderr shows no MCP-related debug output at all — only init and session end, with a 120s gap

Workaround

Add the MCP server to the user config instead of using --mcp-config:

claude mcp add-json myserver '{"type":"http","url":"http://..."}' -s user
claude -p --dangerously-skip-permissions "prompt"
claude mcp remove myserver

This is clunky for dynamic/per-session servers but avoids the scope: "dynamic" code path entirely.

[github-actions]

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