pilotfish
pilotfish is a multi-model orchestration policy for Claude Code that keeps a frontier model responsible for planning, approval and final judgment while routing bounded execution work, like search, repetitive edits, tests and documentation, to cheaper models.
pilotfish configures Claude Code to use different models for different roles instead of one model for everything. A frontier opus-family model handles the main session: planning, approval, integration and final judgment. Sonnet handles bounded execution and reconnaissance, and fresh Opus contexts are used specifically for risk-triggered review, so most of a session's tokens are not spent on frontier-level judgment for routine work.
It installs either as a macOS and Linux Claude Code plugin, which packages the policy and roles under Claude Code's native plugin lifecycle with hook-based ambient activation, or through a legacy global configuration install that edits Claude Code's settings, agent definitions and instructions files directly.
Related projects extend the same policy to other hosts: remora for session-scoped GPT routing in Claude Code, pilotfish-grok for Grok Build, and pilotfish-codex for the Codex CLI.