Cite as: Real Problem AI problem “Why do account-level MCP connectors silently eat 100K tokens before you've typed a single word?”. Opportunity score 7.2 out of 10 (severity 7, AI feasibility 9, market signal 6, competition gap 7). Category AI / Agents. Trend Agents. Source signal: anthropics/claude-code GitHub issue #50062, fernandoxavier02, 17 Apr 2026.. Canonical URL: https://www.realproblem.ai/archive/why-do-account-level-mcp-connectors-silently-eat-100k-tokens-before-you-type.
Why do account-level MCP connectors silently eat 100K tokens before you've typed a single word?
Every connector configured at the Claude.ai account level auto-loads into every CLI session with no per-environment opt-out, burning roughly 90K tokens of schema cost across ~240 tools before the user does anything.
Who has it: Claude Code CLI users who also have Claude Desktop connectors configured.
Evidence
The issue describes account-level connectors silently loading into every CLI session and consuming a large share of context before the user types anything.
Our summary of the public post linked below, not a quote. Nobody submitted it to Real Problem AI.
anthropics/claude-code GitHub issue #50062, fernandoxavier02, 17 Apr 2026.Why it is archived
Trimmed to 100-cap (lowest opportunity_score)
Scoring breakdown
Existing players
- Revoke OAuth entirely · Only full-scope workaround, kills Desktop integrations too.
- Disable each connector by hand · Tedious, must be redone if connectors change.
What they are missing
A per-environment opt-out (env var, allowlist, or scope-aware login) that keeps account connectors available in Claude Desktop while excluding them from CLI sessions by default.
Stack hint
#ATB18 · Canonical URL: https://www.realproblem.ai/archive/why-do-account-level-mcp-connectors-silently-eat-100k-tokens-before-you-type