[BUG] Fresh session consumes 50k+ context tokens before any real work; disabling MCP connectors only reduces it ~9%

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?

Every fresh Claude Code session pays a large fixed token cost before any real work is done. A single "hello" message in a brand new session consumes 51,700–56,900 context tokens. This happens consistently, is not explained by conversation length (context window was 0 before the message), and is barely affected by disabling MCP connectors (~9% reduction only). This is causing Claude Pro quota to be exhausted much faster than with the same workload a few months ago.

What Should Happen?

A trivial first message in a fresh session should cost a small, roughly constant number of tokens. Repeated fresh sessions under identical conditions should benefit from prompt caching (lower cost on repeat), rather than paying the exact same full cost every time.

Error Messages/Logs

No error message — this is a resource/cost anomaly, not a crash. Observed via the "Context window" and "Account & Usage" panels in the desktop app.

Steps to Reproduce

1. Open a brand new session in the Claude desktop app (Cowork/Claude Code mode), no --resume/--continue.
2. Note the context window shown (0 / 967.0k).
3. Send a single message: "hello".
4. Immediately check the context window shown after the reply.
5. Repeat steps 1–4 in a second, completely separate fresh session under identical conditions.
6. Repeat steps 1–4 a third time with all MCP connectors disabled (Settings → Connectors: Meta Ads, n8n, Kling, Google Calendar, Google Drive all OFF).

Results:

• Test 1 (MCP active): 56,900 tokens
• Test 2 (MCP active, separate session): 56,900 tokens (identical — no cache discount)
• Test 3 (MCP disabled): 51,700 tokens

Additionally, the in-app "What's contributing to your limits" dashboard (Day view) showed 65% of usage from sessions >150k context tokens, and MCP connectors at 23% combined (Claude Browser 12% + Kling 11%).

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

unknown — see note above

Claude Code Version

2.1.221 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Screenshots of the 3 isolated tests (context window before/after + Account & Usage panel) available on request.
Cowork/Claude Code desktop app on Windows — not using a terminal for this workflow.
Bundled engine version confirmed via process inspection: 2.1.221. Could not locate an explicit "check for updates"/about version display in the desktop app settings to confirm this is the latest build.