[BUG] [2.1.133] MCP tools not exposed to model + model misidentifies itself as not-Claude-Code

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?

After upgrading to 2.1.133, two related regressions appear:

1. MCP tools missing from model context. MCP servers connect successfully

(UI shows ✔ connected · N tools), but none of those tools are exposed to
the model — the model behaves as if no MCP server is configured.

1. Model identity confusion. When asked about the missing MCP tools, the

model confidently claims it is not Claude Code and recommends writing
config to paths Claude Code does not read.

## Environment

• Claude Code: 2.1.133 (regression from 2.1.132)
• Install: npm install -g @anthropic-ai/claude-code
• OS: macOS (Darwin 25.4.0, arm64)
• Model: Opus 4.7 (1M context)

## Bug 1: MCP tools not injected

### Setup
~/.claude.json contains a working stdio MCP server (lark-mcp in my case,
but the bug is not server-specific):

```json
{
"mcpServers": {
"lark-mcp": {
"type": "stdio",
"command": "/opt/homebrew/bin/lark-mcp",
"args": ["mcp", "-a", "...", "-s", "...", "--oauth", "-t", "..."],
"env": { "NODE_OPTIONS": "--use-bundled-ca" }
}
}
}

Steps to reproduce

1. Run claude on 2.1.133. UI shows lark-mcp · ✔ connected · 22 tools.
2. Ask the model to call any mcp__lark* tool, or simply ask

"what MCP tools do you have?"

1. Model reports it has no MCP tools available; no mcp__* tools appear in

its tool list.

Expected

Tools from connected MCP servers are callable, as on 2.1.132.

Actual

• Zero MCP tools visible to the model.
• npm install -g @anthropic-ai/claude-code@2.1.132 → tools reappear with

the same untouched config.

• Reinstall 2.1.133 → tools disappear again. Deterministic.

Bug 2: Model identity confusion

In a 2.1.133 session, when asked why MCP tools weren't available, the model
responded (paraphrased, full transcript available on request):

▎ "You're looking at Claude Code's config (~/.claude.json). I am Claude
▎ running in Anthropic, a different product — my MCP config lives at
▎ ~/.anthropic/anthropic/mcp_config.json. Want me to create that file and
▎ copy the config over?"

That path is not read by Claude Code. Following the model's advice would
just create dead files. The model only conceded it was wrong after the user
pushed back with the version banner.

Possible causes:

• System prompt regression in 2.1.133 that no longer asserts the model is

running inside Claude Code.

• Interaction with Bug 1: the model sees no MCP tools, can't reconcile that

with the user's claim, and confabulates a plausible-sounding explanation.

Either way, this is high-impact: the model's confident wrong advice can lead
users to pollute their filesystem or chase phantom config problems.

Workaround

npm install -g @anthropic-ai/claude-code@2.1.132
export DISABLE_AUTOUPDATER=1 # or: claude config set -g autoUpdates false

Additional info

Happy to provide /mcp output, debug logs, or full session transcripts.

What Should Happen?

Tools from connected MCP servers are callable, as on 2.1.132.

Error Messages/Logs

▎ "You're looking at Claude Code's config (~/.claude.json). I am Claude                                               
  ▎  running in Anthropic, a different product — my MCP config lives at                                                 
  ▎  ~/.anthropic/anthropic/mcp_config.json. Want me to create that file and                                            
  ▎  copy the config over?"                                                                                             
                                                                                                                        
  That path is not read by Claude Code. Following the model's advice would                                              
  just create dead files. The model only conceded it was wrong after the user                                           
  pushed back with the version banner.

Steps to Reproduce

1. Run claude on 2.1.133. UI shows lark-mcp · ✔ connected · 22 tools.
2. Ask the model to call any mcp__lark* tool, or simply ask

"what MCP tools do you have?"

1. Model reports it has no MCP tools available; no mcp__* tools appear in

its tool list.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.132

Claude Code Version

2.1.133

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_