Model multiplexer plugin for Claude Code: premium model orchestrates, cheaper models execute, with quality-first routing rules and a real cost report
Agent
Quick top-tier judgment - Fable pinned to low reasoning effort. Use for verdict-shaped questions where judgment quality matters more than deep reasoning - taste checks on rendered output and screenshots, design and naming calls, plan sanity checks, resolving a reviewer ESCALATE. Far cheaper per call than oracle, but…
Agent
The implementation workhorse (Opus). Default delegate for writing and changing code — features, bug fixes, tests, refactors, integration work, debugging. Works until the job is done and verifies its own work before reporting. Escalate to the orchestrator only for genuine architectural or product decisions.
Agent
Delegate a coding task to OpenAI Codex via the codex CLI — the actual work costs zero Anthropic tokens (bills to the user's OpenAI/ChatGPT account). Use for bulk implementation when Anthropic quota is tight, parallel workstreams, or a second opinion from a non-Anthropic model. Requires the codex CLI installed and…
Agent
Delegate a task to Google Gemini via the gemini CLI — the actual work costs zero Anthropic tokens (bills to the user's Google account/plan). Strongest uses - summarizing or analyzing very large inputs (huge logs, whole codebases, long documents) thanks to its large context, bulk text processing, and second opinions…
Agent
Escalation to Fable, the most capable model. Bills at premium rates outside the Max plan, so use sparingly and deliberately. Reserve for - visual and creative direction for games and frontends, the hardest architecture or design decisions, deep analysis where an Opus attempt was not good enough. Give it complete…
Agent
Browser and app driving at Sonnet cost. Use to navigate pages, click through flows, fill forms, and capture screenshots or GIF recordings at named checkpoints - so raw page dumps and pixels never transit the orchestrator's context. Reports artifact file paths plus a factual account of what happened. Never judges…
Agent
Web research at Sonnet cost. Use for search-and-synthesize work - collecting sources and filings, market/policy/technical research, news sweeps, competitive scans, docs beyond the codebase. Returns a condensed brief with every claim tied to a named source. Prefer over general-purpose for research - the model guard…
Agent
Verification and code review (Opus). Use after builder, writer, codex, or gemini finish a task — to check the work against the original intent instead of the orchestrator re-reading diffs itself. Also for pre-commit review sweeps. Read-only by default; reports findings, does not fix them.
Agent
Command execution at Haiku cost. Use to fire builds, test suites, dev servers, and one-off commands, then get back a condensed pass/fail report with the error lines that matter - instead of running them in the main loop and paying top-tier rates to read raw output. Its commands may change state (build artifacts…
Agent
Cheap, fast reconnaissance (Haiku). Use for any read-and-report work — exploring a codebase, finding files or symbols, reading logs/docs/configs and summarizing them, checking whether something exists, gathering context before planning. Returns a dense summary so the orchestrator never has to read raw files itself.…
Agent
Text and boilerplate at Sonnet cost. Use for docs, READMEs, comments, commit messages, copy drafts, config files, data-file editing, and mechanical low-risk code edits (renames, find-replace, formatting, moving files). Not for logic changes, debugging, UI/CSS, or anything users will see and judge aesthetically …