/context may over-state MCP token cost

You run /context, see a five-figure number next to an MCP server, and reach for /mcp to disable it. That reflex is worth a second's hesitation: at least one careful measurement says the number you are reacting to may be several times the tokens that server actually costs per request.

The measurement

A developer measured one MCP server three different ways and got three different answers:

The third row is reportedly how /context arrives at its MCP figure. The proposed explanation: every count_tokens request that carries tool definitions also carries a hidden tool-use preamble of roughly 313-346 tokens. Measure one tool, you pay it once. Measure sixty tools one at a time and add the results, and you have counted that preamble sixty times — about 20k tokens of pure double-count on top of the real ~15k.

A separate write-up on a four-server setup with tool search deferral active reported roughly 7,077 tokens of steady-state overhead for 37 tools, against roughly 20k with deferral off. Neither figure is nothing. Neither is "half your context window" either.

Why this matters less than it used to

Since tool search became the default, MCP tool definitions are deferred: only tool names and server instructions load at session start, and Claude fetches a full schema on demand when it decides to use that tool. The docs describe the mechanism in Scale with MCP tool search, and the context window walkthrough lists MCP tools as a deferred line item with a small startup cost.

So on a default setup the /context MCP line is already small, and the over-count question is mostly academic. It stops being academic in every case where tools load upfront instead:

• ENABLE_TOOL_SEARCH=false, or auto/auto:N while the definitions still fit under the threshold
• ANTHROPIC_BASE_URL pointing at a non-first-party host, where tool search is off by default because most proxies do not forward tool_reference blocks
• CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS, which strips defer_loading and cannot be overridden by setting ENABLE_TOOL_SEARCH yourself
• Microsoft Foundry deployments hosted on Azure, and Google Cloud Agent Platform models earlier than the 4.5 generation
• Individual servers you exempted with alwaysLoad: true

Those are exactly the setups where a big MCP number appears and someone starts disconnecting servers.

What to do instead of trusting the number

Measure it yourself, once, for your own configuration:

# 1. See what Claude Code reports
claude
# then: /context   — note the MCP tools figure

# 2. Force the interesting case, where tools load upfront
ENABLE_TOOL_SEARCH=false claude
# then: /context   — note the figure again

Then compare against a single count_tokens API call carrying the same tool definitions in one request. One call, all tools, one preamble. If your single-call number lands far below the /context figure, the summation hypothesis holds for your version. If the two agree, either the arithmetic changed or it never applied to your setup, and you can stop worrying about it.

What is actually unverified here

Being specific about the gaps, because that is the whole value of this entry:

• The mechanism is inferred, not confirmed. No official source states that /context sums per-tool measurements. The author reverse-engineered it from the arithmetic lining up.
• The version is unknown. The measurement predates tool search being on by default. If the summation has since been corrected, this tip is stale.
• The sample is one server. The 3x ratio is a property of tool count and average schema size, not a constant. A server with six tools would show a far smaller gap; the preamble is charged per measurement, so the error scales with how many tools you have.
• count_tokens is itself an approximation of what a live request costs, since it does not model caching or the rest of your conversation.

The cheap, correct posture: before you disconnect a server on the strength of a /context number, spend two minutes measuring it in one batched call. If the server is genuinely costing you 15k rather than 45k, that is often the difference between "cut it" and "keep it and turn deferral back on."