Two projects can configure the same MCP server and observe different usage. The client may expose different tools, defer definitions, or return different amounts of content.
MCP connects an agent to external tools and context. The Codex MCP documentation describes server configuration and tool controls. The Claude Code MCP documentation explains its tool discovery behavior.
Separate the four sources of context
| Source | What it contains | What to inspect |
|---|---|---|
| Server instructions | Guidance about when and how to use the server | Length, relevance, repeated instructions |
| Tool definitions | Names, descriptions, and argument schemas | Which definitions the client actually loads |
| Discovery results | Information returned when the agent searches for tools | Repeated discovery and excessive descriptions |
| Tool results | Retrieved documents, records, logs, or code | Relevance, size, repeated content |
Network connections and model tokens are separate measurements. A connected server does not prove that every schema entered every request. A tool call can also incur a separate service charge. Record that charge separately when it applies.
Claude Code can defer tool definitions
Anthropic documents MCP tool search as enabled by default in supported configurations. The client initially exposes tool names and server instructions, then retrieves full definitions when needed. Provider compatibility and configuration can change that behavior.
This makes the statement “every installed MCP tool always consumes its full schema” unreliable. Record the client version, model provider, and tool search configuration in any comparison. Use the official compatibility notes for your deployment.
Codex exposes server and tool controls
In Codex CLI, codex mcp list lists configured servers. The /mcp command shows active MCP information within a session.
Codex configuration supports these controls:
| Setting | Purpose |
|---|---|
enabled |
Enables or disables a configured server |
enabled_tools |
Allows selected tools |
disabled_tools |
Excludes selected tools |
See the official configuration reference. Apply changes in a test configuration before changing a shared setup. Preserve tools that provide required project knowledge or validation.
Measure startup context and useful work separately
A startup comparison answers how much context appears before useful tool work. A task comparison answers whether that context helps the agent finish. You need both to interpret overhead.
- Record the repository revision, client version, model, effort, and server configuration.
- Start a fresh session with the required tools only.
- Record the available context breakdown and first request usage.
- Repeat with one additional server under equivalent conditions.
- Run a task that can use that server in both configurations.
- Compare total usage, tool calls, elapsed time, and acceptance results.
Do not treat the difference between two context meters as a complete billing estimate. Caching, reasoning, output, and later tool use can change the result.
If the client does not expose a startup breakdown, report that measurement as unavailable. Do not estimate schema tokens from the number of servers.
Reduce low-value results first
Ask a search tool for a bounded result set. Retrieve the relevant document section instead of the entire document when the tool supports it. Use a code reference query when you need callers, and a text search when you need a simple definition.
Keep an escalation path to the full source. A short result that removes the error or its cause can create more work than it saves.
Halv’s Crux integration supports code navigation, while RTK filters command output. Try Halv on a code investigation with a clear completion check.
Decide whether the server earns its cost
Keep a server when its information improves correctness or avoids enough work to justify its overhead. Disable a server when it adds irrelevant context or duplicates a tool you already use. Review this decision per project, because the useful tool set can change.
For other causes of usage, read the Codex diagnosis guide. For instruction overhead, read AGENTS.md token usage.
Sources and scope
Official documentation checked on September 28, 2026. This article explains a measurement procedure. It does not report a measured token tax for an MCP server.