GitHub optimizes token usage in agentic workflows by removing unused tool configurations and replacing MCP calls with GitHub CLI commands.

GitHub began systematically optimizing token usage of agentic workflows in April 2026, building a Daily Token Usage Auditor to flag significant consumption increases and a Daily Token Optimizer to propose specific improvements based on workflow logs and source code.

The most common inefficiency was unused MCP tool registrations: a GitHub MCP server with 40 tools can add 10–15 KB of schema per API call even when the agent only uses two tools. Removing unused tools from configuration reduced per-call context size by 8–12 KB, saving several thousand tokens per run with no change in behavior.

GitHub replaced MCP calls for data-fetching operations (pull request diffs, file contents, review comments) with GitHub CLI commands through two methods: pre-agentic data downloads (running gh commands before the agent starts) and in-agent CLI proxy substitution (routing CLI traffic to GitHub's API without exposing credentials to the agent).

GitHub introduced an Effective Tokens (ET) metric that normalizes token consumption across different models by applying model cost multipliers (Haiku = 0.25×, Sonnet = 1.0×, Opus = 5.0×) and weighting different token types (input = 1.0×, cache-read = 0.1×, output = 4.0×) to distinguish cost reduction from model switching.