Tool call
A tool call is a structured request from a model to run something outside itself, such as reading a file or executing a command, whose result is fed back into the context.
Tools are what let a model act rather than only answer. Their context cost depends on the client and its loading strategy: some include detailed schemas up front, while others defer them. As of September 2026, Codex can defer discoverable MCP tools through tool search. Tool results then join the conversation and can be sent again on later turns.
The results cost more than the definitions. A command that prints two thousand lines contributes two thousand lines of context, and keeps contributing them for the rest of the session.
Filtering output before it reaches the model is therefore one of the few interventions that is both simple and worth a lot. head, grep, a --quiet flag or a JSON filter all cost less than asking the model to ignore the noise.