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

[nevilhulspas]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• 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:

• [BUG] Claude Code Context Bloat: Bug Report + Community Issue Tracker #29971 — Claude Code Context Bloat tracker (umbrella)
• claude.ai Skills and Cowork plugins silently injected into Claude Code context — no opt-out, ~6k tokens wasted per session #39686 — Skills and Cowork plugins silently injected, no opt-out
• Phantom ~22K tokens per session on one account vs another — same machine, same project #45515 — Phantom ~22K tokens per session on one account vs another
• Allow disabling claude.ai MCP connectors per-client (Claude Code vs Chat) #26625 / Add setting to disable cloud MCP connectors from claude.ai #22301 / [FEATURE] Support disabling claude.ai MCP connectors at the project level #45158 — feature requests to disable claude.ai connectors per-client
• Claude.ai account-level MCP connectors auto-loaded into CLI with no per-environment opt-out (~100K tokens of silent context bloat) #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).

[nevilhulspas]

Auto-labeled invalid, requesting re-triage. This is the chat-side manifestation of the same account-level connector-loading machinery that's already tracked in this repo via #29971, #39686, and #50062 (closed). The CLI and chat surfaces share the connector backend; the same connectors injected into Claude Code via account auth are what's blocking chat here. If there's a better Anthropic repo for claude.ai chat bugs I'm happy to move it, but I couldn't find one — every existing report on this class of bug lives here.

[itsmanoharkumar]

Cross-linking: I've refiled this in the dedicated MCP-integration tracker, which is the correct repo for claude.ai connector bugs (issues filed here tend to get auto-labeled invalid as consumer-app scope):

➡️ anthropics/claude-ai-mcp#401

It includes a clean repro and a recommended fix — when Load tools when needed is enabled, exclude connector tool schemas from the system-prompt prefix and load them on demand via the existing Tool Search / defer_loading path (as Claude Code CLI already does), which also preserves prompt caching.

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

[happyarts]

Same error class here, but account-level and independent of connector count — and far larger. Reopening consideration would be appreciated, since this closed issue is the only public record of this failure mode.

Since ~2026-07-10, nearly every message on claude.ai fails for my account (web, desktop app, iOS, incognito, all models, brand-new empty chats, prompt "test") with:

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

with N drifting between 4,625,481 and 5,863,234 tokens over the week. Claude Code and Cowork on the same account work fine.

API bisection (2026-07-17, 09:20–10:00 UTC):

• Overage is identical across models (Haiku 4.5 vs Opus 4.5: Δ10 tokens) → single fixed blob, ~4.84M tokens (~19 MB), injected server-side during prompt assembly.
• Independent of client payload: fails identically with tools: [] on a fresh conversation created via API.
• Ruled out while failing: all 11 connectors disconnected, memory deleted + both memory toggles off, skills/plugins/styles < 15 KB total, no local config (fails on mobile/incognito too).
• The blob intermittently disappears for a few minutes and returns: I directly observed a ~7-minute window where identical requests suddenly succeeded, then failed again with a new overage value (4,625,481, three identical runs).
• Side effect: account memory never rebuilds after deletion — background summarization jobs presumably die on the same oversized context.

Request IDs for Anthropic staff (2026-07-17): failing req_011Cd7Qwe45kCCijPaknN2wb (~09:25 UTC, tools:[], Haiku 4.5, overage 4,641,642) vs succeeding chatcompl_011Cd7RQ5v4VstraeuxQsV6p (~09:40 UTC, same account, near-identical payload). Diffing prompt assembly between these two should reveal the injected object. Support conversation ID 215475070708461 (unanswered for 4 days).

The paid chat product has been unusable for over a week on this Max account. Please escalate.