[BUG] claude.ai chat: 'Load tools when needed' setting doesn't actually defer connector schemas — context limit exceeded at ~11 connectors before first message

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?

The deferred tool loading setting in claude.ai chat ("Load tools when needed") does not actually defer connector tool schemas. With this setting enabled, connecting more than ~10 MCP connectors to a Claude.ai account causes new chats to fail immediately with:

"This request exceeds Claude's context limit by about 33,609 tokens. Try turning off some tools or shortening your message."

…before any user message is sent.

This breaks chat entirely across all surfaces (web, Claude Desktop, Claude mobile) for accounts with more than ~10 connectors enabled. Claude Code and Cowork on the same account are unaffected, which suggests they handle deferred tool loading correctly and chat does not.

What Should Happen?

With "Load tools when needed" enabled, only a lightweight registry/index of available connectors should be injected into context up front. Full tool schemas should only be loaded when Claude decides to invoke a tool in a given connector — which is how Cowork and Claude Code handle the same set of connectors without exhausting context.

A new chat should be able to start regardless of how many connectors are enabled at the account level, because none of those schemas should be in context until a tool is actually called.

Error Messages/Logs

This request exceeds Claude's context limit by about 33,609 tokens. Try turning off some tools or shortening your message.

Steps to Reproduce

1. On a Claude.ai account, go to Settings → Connectors and enable the "Load tools when needed" / deferred loading setting.
2. Connect ~11 or more MCP connectors at the account level (a mix of light and heavy: e.g. Gmail, Calendar, Drive, Slack, Notion, Linear, Asana, HubSpot, Shopify, plus any heavy one like Ahrefs or a large e-commerce/ads MCP).
3. Start a brand-new chat in claude.ai web (or Claude Desktop, or Claude mobile — all three reproduce).
4. Type any message (e.g. hello) and send.

Expected: chat responds normally; tool schemas are only loaded when a tool is actually needed.

Actual: chat fails immediately with This request exceeds Claude's context limit by about 33,609 tokens. Try turning off some tools or shortening your message. The overshoot scales with how many connectors are enabled.

Reproduction is deterministic for a given account state:

• With 0 connectors enabled, chat works.
• Re-enabling connectors one at a time, chat continues to work up to ~10.
• Adding the 11th connector (varies by which connector — heavier connectors push the limit faster) triggers the error.
• Disabling the most recently added connector restores chat.

The failure is independent of: logging out / clearing cache, switching between web / Desktop / mobile, starting a fresh chat (it happens on the first message of a brand-new chat), or rebooting. Claude Code and Cowork on the same account work fine throughout — only chat is blocked.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

N/A — this bug is in claude.ai chat (web / Claude Desktop / Claude mobile), not the Claude Code CLI. Claude Code and Cowork on the same account work normally.

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

Surfaces affected: claude.ai web, Claude Desktop, Claude mobile (iOS). All three reproduce on the same account.
Surfaces NOT affected: Claude Code (CLI), Cowork — both handle the same set of connectors without exhausting context.

Why this matters / why it's distinct from prior reports:
The "Load tools when needed" setting was the official answer to context bloat from account-level connectors. This issue is that the setting does not actually defer schemas — it appears to inject the full tool schemas anyway, just with different timing, so adding the 11th-ish connector wedges chat entirely (cannot send the first message of a new chat).

Related (mostly CLI-focused) open issues:

• #29971 — Claude Code Context Bloat tracker (umbrella)
• #39686 — Skills and Cowork plugins silently injected, no opt-out
• #45515 — Phantom ~22K tokens per session on one account vs another
• #26625 / #22301 / #45158 — feature requests to disable claude.ai connectors per-client
• #50062 (closed) — claude.ai connectors auto-loaded into CLI; the chat-side equivalent is what this issue is about

Workarounds attempted (none worked):

• Log out / log back in
• Clear browser cache / app cache
• Try web, Claude Desktop, and Claude mobile (same account state, same error on all three)
• Toggle "Load tools when needed" off and on
• Start a brand-new chat (the failure is on message 1)

What actually works:

• Disconnecting connectors until count drops to ~10 or below. The exact ceiling depends on which connectors are enabled (heavier ones with many tools push the limit sooner).

Asks:

1. Make "Load tools when needed" actually deferred so connector schemas only enter context when a tool is invoked.
2. Surface the connector context cost in Settings → Connectors (e.g. "this connector adds ~Xk tokens to every chat") so users can budget.
3. Per-chat / per-project connector enable/disable in claude.ai chat (parity with the way Cowork scopes them per session).