The multi-model toolkit

The multi-model toolkit

One thread.
Many minds.

Every job has one door: ask → Consensus · review → Second opinion or Improve · argue → Debate · build → Boss & Worker.

Three ways to put models in one room NEW

Three session types on the new-chat screen — each one looks like what it is.

Boss & Worker

Talk to the boss. It keeps a pinned project card — mission, task list, round history — dispatches the worker one task at a time, and reviews every round. Arm Autopilot and it runs the whole list on its own, capped so it can't run away. Everything survives restarts, and the plan mirrors to a file in your repo.

Debate

Two sides trade rebuttals until one concedes, and the concession turns green. You get one agreed answer.

Group chat

Cast each seat: Sonnet as a product manager, Codex as an ML engineer. They argue their corners, avatars, likes and all.

/consensus — three models, one answer

Ask once. Three models answer in parallel, each as its own card, then your session model merges them into one answer — with disagreements called out, not papered over. For the decisions you don't want a single model's blind spot on.

Second opinion, on the latest reply

Pick any model — Claude, GPT, Gemini or a free one — to critique or validate the last answer, in-thread. If your pick is capped, a live model steps in. GPT checks Claude, Claude checks GPT — the disagreements are where the bugs were hiding.

Improve — critique, then rewrite

Another model critiques the answer; your model rewrites it addressing every point. Draft → critique → revise, automated into one button.

@mentions — address any model

Start a message with @sonnet, @gpt or @qwen and the conversation hands off to that model from there — the thread stays one conversation, with every model an addressable participant.

Smart routing + Economy / Balanced / Best modes

Pick Economy and everything routine runs on free models — you pay only for genuinely complex work. Or turn on Smart routing and each message routes itself: trivial asks go to free or cheap models, hard ones to premium. Your subscription stretches before any limit hits, not after.

Guardian — a model watching the agent

In Auto mode, a deterministic policy classifies every shell command and a cheap second model reviews it before it runs; anything risky escalates to an approval card. The deterministic layer runs even when the model reviewer is off, so a risky command never slips through on a timeout.

The savings ledger — your receipt

Renitor counts, locally, every limit it survived and every token it ran free, priced at paid rates. One click turns it into a shareable receipt card. Your machine keeps the score — nothing is phoned home.

The Chain — briefs written by models

When the thread switches models, the incoming one inherits the whole conversation — and once it fills up, a model writes the summary that folds the older turns down. Context survives because a model wrote it, not a regex.

Question cards — models ask before they guess

Give a genuinely ambiguous task and the model asks up to 3 short multiple-choice questions as clickable cards in the thread — then continues the same turn with your answers. No retyped context, no wrong guess built at full speed.

Your agents & skills — on every model

The personas in .claude/agents and skills in .claude/skills you already wrote for Claude Code run here on any model: @reviewer fix this on GPT, /deploy-check on a free model. Your setup stops being locked to one vendor.