Context efficiency

​

Context ratios versus flat MCP

​

Selective output task

​

Interpretation

• Tool-definition and task-context tokens are exact for the o200k_base text encoding, but host-specific MCP wrappers and any generic shell-tool schema are excluded equally. They are not a billing estimate for every model.
• MCP wire figures are JSON HTTP body bytes only. Headers are excluded because bearer length is noise and credentials must never enter output. CLI auth exchange bytes are recorded separately in JSON.
• CLI context assumes the agent already knows the stable Passport command contract. One-time installation/help text is excluded; every per-task command and stdout byte is included.
• Latency includes localhost gateway/upstream work. CLI includes two Node subprocesses, two credential exchanges, and two MCP sessions. Latency is observational and deliberately has no pass/fail threshold.
• Automatic mode uses the current catalog-size and approximate tool-definition token-budget choice: flat for the small fixture, apps for the medium fixture, and compact for the large fixture. The budget excludes server instructions and result payloads, and compact’s fixed meta-tools may exceed a very small budget. The benchmark records the actual path rather than assuming a tier.
• Broad native searches are brief; the selective compact path performs a brief search followed by one exact schema lookup. CLI keeps its compatibility search behavior.
• Selective output is opt-in and is applied only after the complete provider result has passed Passport’s existing result safety checks. The raw-result comparison is the deterministic fixture envelope, while selected-result tokens are measured from the real gateway response. The selector pages this response; it does not use a cached snapshot or provider cursor.