`/context` under-reports MCP tool schema consumption (shows 0 tokens per tool, actual ~100K hidden in "Messages")

Summary

The /context command reports each MCP tool as consuming 0 tokens, but in reality the combined cost of tool schemas (name + description + JSONSchema parameters) can exceed 100K tokens on a fresh session. The real cost is absorbed into the Messages category, making it effectively impossible for users to audit which MCP servers are bloating their context window.

Environment

• Claude Code CLI v2.1.112 (also reproducible on earlier 2.x)
• Platform: Windows 11 (also reported via bash/git-bash and WSL)
• Provider: any (tested with both native Anthropic and a third-party via ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic)
• MCPs tested: mix of plugin-registered and Claude.ai account-level connectors (Canva, ClickUp, Figma, Gmail, Google Calendar, Hugging Face, Linear, Notion, Vercel, GitHub, Chrome DevTools, Playwright, Telegram, etc.)

Reproduction

1. Have several MCP servers active (ideally 10+ servers totalling ~200+ tools — trivially reproduced for any user with Claude.ai account connectors + a few plugin MCPs).
2. Start a fresh CLI session.
3. Run /clear.
4. Send a minimal prompt, e.g. hi.
5. Run /context.

Expected

Each MCP tool's schema consumption is either:

• Shown as a non-zero per-tool cost, OR
• Aggregated into a dedicated MCP tools category with realistic tokens, separate from Messages.

Actual

Every individual MCP tool is listed as 0 tokens, yet Messages reports ~100K+ tokens on a brand-new session with nothing but hi. The user has no way, from /context, to attribute that cost to MCP.

Evidence (side-by-side measurement)

Identical fresh sessions, identical prompt (hi), only variable: which MCPs are enabled.

Before disabling account-level MCPs

Context Usage — MiniMax-M2.7-highspeed
151.6k / 200k tokens (76%)

Estimated usage by category
  System prompt:   5.8k tokens  (2.9%)
  Custom agents:   4.7k tokens  (2.3%)
  Memory files:    2.8k tokens  (1.4%)
  Skills:         16.8k tokens  (8.4%)
  Messages:      121.6k tokens  (60.8%)   ← hidden MCP cost lives here
  Free space:     15.4k         (7.7%)
  Autocompact buffer: 33k       (16.5%)

MCP tools · /mcp
  └ mcp__claude_ai_Canva__cancel-editing-transaction: 0 tokens
  └ mcp__claude_ai_Canva__comment-on-design:          0 tokens
  └ ... [~280 tool entries, EVERY ONE showing 0 tokens]

After /mcp disabling all 12 Claude.ai account-level connectors (plugin MCPs left alone)

Context Usage — MiniMax-M2.7-highspeed
51.1k / 200k tokens (26%)

Estimated usage by category
  Custom agents:   5.1k tokens  (2.6%)
  Skills:         15.2k tokens  (7.6%)
  Messages:       30.8k tokens  (15.4%)   ← dropped by ~93K
  Free space:    115.9k tokens  (58.0%)
  Autocompact buffer: 33k       (16.5%)

Delta

| Metric | Before | After | Delta |
| --------------- | ------- | ------ | -------- |
| Total used | 151.6K | 51.1K | -100.5K |
| Messages | 121.6K | 30.8K | -93.3K |
| Free space | 15.4K | 115.9K | +100.5K |

Disabling ~280 MCP tools removed ~100K tokens — averaging ~360 tokens per tool schema — yet every one of those tools was reported as "0 tokens" in /context in both runs.

Impact

• Silent exhaustion: Users see high context consumption with no way to identify the cause. New users blame the model, their prompt, or the CLI itself.
• Unmeasurable optimisation: Users cannot see which MCPs are expensive vs. cheap. Disabling the wrong ones wastes time.
• Compounds with issue #46416: When CC falls back to a 200K window for unknown third-party providers, losing 100K+ to invisible MCP schemas on startup makes the CLI effectively unusable for long sessions.
• Discoverability of connector inheritance: Without a clear per-MCP cost display, users don't realise that Claude.ai account connectors (Desktop) are loaded into CLI. (This is the subject of a separate issue.)

Suggested Fix

Pick one (or combine):

Option A — Attribute tokens per tool / per server

Show the real schema cost per tool or, at minimum, per MCP server:

MCP tools · /mcp
  └ claude.ai Canva:        12.4k tokens  (31 tools)
  └ claude.ai ClickUp:      18.0k tokens  (45 tools)
  └ plugin github:          16.2k tokens  (41 tools)
  ...

Option B — Add a dedicated category

Split Messages into Messages (conversation) and MCP tool schemas, so users can see MCP tool schemas: 93K at a glance.

Option C — At minimum, document it

Update the /context docs and the CLI help to state that MCP tools are counted inside Messages and that the displayed per-tool 0 tokens is the deferred-load incremental cost, not the baseline schema cost.

---

Additional debugging info (for Anthropic engineers)

The ~360 tokens/tool average is consistent with a full tool schema: name + description + JSONSchema parameters with types and documentation. Cross-check with ToolSearch: when a deferred tool is searched, its schema is loaded on top of whatever baseline already exists — suggesting the 0 tokens figure in /context is the incremental cost after a search, not the baseline schema catalog cost.

A user-level audit script confirmed the OAuth scope user:mcp_servers in ~/.claude/.credentials.json triggers account-level connector loading in CLI, but the per-tool cost attribution in /context remains 0 regardless of whether the connector is local (plugin), remote (account), or hybrid.

---

Happy to provide additional traces, /context dumps, or env details on request.