[BUG] Cowork macOS — MCP tools listed in session config but absent from active tool list (mcp__workspace__bash, cowork tools, all connectors)

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?

App version: 1.13576.4 (414f85) — 2026-06-18T02:11:10.000Z
Platform: macOS, Claude Desktop (Cowork)

VM health: confirmed healthy

• Startup complete in ~5s (normal sessions)
• guest_vsock_connect, guest_ready, sdk_install all complete
• API reachability: REACHABLE
• SDK version: 2.1.177

Problem:
Session config correctly lists all MCP servers as connected
(Workspace/bash, Cowork tools, Fathom, Supabase, Scheduled Tasks, etc.)
but none appear in the active tool list. Only file tools are callable
(Read, Write, Grep, Glob, Skill).

Confirmed absent from active tool list:

• mcp__workspace__bash
• mcp__workspace__web_fetch
• mcp__cowork__present_files
• mcp__cowork__create_artifact
• mcp__cowork__request_cowork_directory
• mcp__scheduled-tasks__*
• All desktop extension MCPs (Fathom, Supabase, Apollo, etc.)

Steps taken (none resolved):

• Full app uninstall/reinstall (x2)
• Manual deletion of vm_bundles/claudevm.bundle and re-download
• Sessions with and without folder selected
• Sign out and back in
• Reproduced across every session over multiple weeks

Related issues: #26259, #31647

What Should Happen?

14:17Claude responded: When a Cowork session starts correctly, this is what should happen:When a Cowork session starts correctly, this is what should happen:

The VM boots in the background (~5–10 seconds on a healthy session)
The session connects to the VM over vsock
The VM registers mcp__workspace__bash and mcp__workspace__web_fetch into the active tool list
The Cowork layer registers mcp__cowork__present_files, mcp__cowork__create_artifact, and mcp__cowork__request_cowork_directory
Any connected desktop extension MCPs (Fathom, Supabase, Apollo, etc.) get passed into the session's --allowedTools and --mcp-config arguments
All of the above are callable from the first message

What you actually get: Steps 1 and 2 complete correctly (the VM log confirms this). Steps 3–6 never happen. The session config knows the tools exist but the process that registers them into the active tool list silently fails.
The gap is between the VM being ready and the tools being handed to the session. The VM is healthy. The handoff is broken.

Error Messages/Logs

Steps to Reproduce

1. Open Claude Desktop app (macOS)
2. Start a new Cowork session (with or without a folder selected —

both reproduce)

1. Ask Claude: "what tools do you have in your active tool list

right now?"

1. Observe that session config lists MCP servers as connected

(Workspace, Cowork, Fathom, Supabase, Scheduled Tasks, etc.)
but the active tool list only contains Read, Write, Grep, Glob,
Skill

1. Attempt to call mcp__workspace__bash directly — returns

"No such tool available"

1. Repeat across fresh sessions, with and without folder, after

full reinstall, after manual vm_bundles deletion — issue
reproduces every time

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown — no version number recorded before the regression. Functionality was confirmed working briefly after a clean reinstall, then broke again on the subsequent app update.

Claude Code Version

1.13576.4 (414f85) — 2026-06-18T02:11:10.000Z

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

VM boot logs confirm the sandbox is healthy — this is not a VM
provisioning failure:

2026-06-18 12:12:39 [info] [VM:steps] vm_boot completed (83ms)
2026-06-18 12:12:43 [info] [VM:steps] guest_vsock_connect completed (3896ms)
2026-06-18 12:12:43 [info] [VM:steps] guest_ready completed (303ms)
2026-06-18 12:12:44 [info] [VM:steps] sdk_install completed (0ms)
2026-06-18 12:12:44 [info] [VM] API reachability: REACHABLE
2026-06-18 12:12:44 [info] [VM:start] Startup complete, total time: 5006ms

The failure is specifically in the handoff between VM startup and
tool registration into the session's active tool list. The session
config correctly identifies all MCP servers as connected, but none
appear as callable tools.

Troubleshooting steps completed before filing:

• Full app uninstall/reinstall x2
• Manual deletion of ~/Library/Application Support/Claude/

vm_bundles/claudevm.bundle/ and full re-download (~10GB)

• Tested with and without workspace folder selected
• Signed out and back in to the app
• Confirmed latest app version (1.13576.4)
• Issue reproduces across every session over multiple weeks

The split between what the session config reports and what is
actually callable suggests the --allowedTools and --mcp-config
arguments are not being constructed correctly when the Cowork
session spawns, consistent with #26259.

OAuth scope confirmed missing from Cowork token requests.

All Cowork sessions request:
  user:inference user:file_upload user:profile

user:mcp_servers is never requested. Claude Code CLI on the 
same machine correctly requests user:sessions:claude_code 
but Cowork does not request user:mcp_servers.

Without this scope the claude.ai MCP gateway rejects all 
connector tool registration, explaining why workspace bash, 
cowork tools, and all connectors show in session config 
but are absent from the active tool list.

Evidence from ~/Library/Logs/Claude/main.log:
[oauth] cacheKey=...:user:inference user:file_upload user:profile
scope: 'user:inference user:file_upload user:profile'
(repeated across all Cowork sessions — user:mcp_servers never present)

Related: #62556