Baton

Baton

Conduct the fleet.

Baton turns Claude Code into the conductor of a fleet of coding LLMs — paid cloud CLIs, free CLIs, and local models on your own machines. You hand each task to the right agent (pass the baton), and Baton tracks what they did, what it cost, what you decided, and what you learned. Built on claude-octopus as the dispatch layer (recommended companion plugin, not a hard dependency).

Status: v1.15.0 — the fleet does the labor. The conductor now farms coding tasks out to non-Claude instruments and gets repo-applied results back — with direct /baton:codex|grok|gemini|agy lines to every model, per-model token telemetry, plan/research/acceptance gates, and a usage governor that keeps spend legible. MIT licensed. An early/experimental personal project, shared in the hope it's useful — not a turnkey product.

📖 New here? Read these

• Full guide (start to finish) — what it is, how to install, and a worked walkthrough.
• Command reference — every command and flag, in plain language.
• Decision log — every design decision and why.
• Roadmap — what's shipped and what's parked.

Install (Claude Code plugin)

claude plugin marketplace add Ryfter/baton
claude plugin install baton@ryfter

Commands surface as /baton:<command> — e.g. /baton:go "ship the feature", /baton:codex "...", /baton:route, /baton:fleet doctor, /baton:usage.

---

Visual overview

The mental model — three tiers, like an orchestra. You + Claude Code conduct; the orchestrators route, gate, and govern; the instruments (the fleet) do the labor. CLI-first — the dashboard is a control board over the same commands.

Using /baton — the happy path. Which command drives each step, from a spark to shipped.

The command map. All 50 commands, grouped by what you're doing.

The fleet loop. Every dispatch is routed, metered, journaled, and learned from — observe, govern, learn.

Diagram sources live in docs/assets/diagrams/ (SVG — editable, versioned).

---

Features

Each feature has a one-line "what it does"; the link goes to its full design spec.

• Automatic usage tracking — every AI dispatch is logged with time, cost, and token counts to a journal, plus a catalog of each model's strengths and pricing. Commands: /baton:log-routing, /baton:consolidate-routing.
• Live web dashboard — a browser view at http://localhost:8765 showing real-time activity, today's spend, a model leaderboard, and controls to stop local models. Runs fully offline. (spec)
• A dispatchable fleet of AI models — one registry of paid CLIs, free CLIs, and local models, all callable from a single command. Adding a model is a few lines of config. Commands: /baton:fleet doctor|list|test. (spec)
• Cross-machine fleet — pull models running on other machines into the fleet over a private network (Tailscale), so a beefier desktop's local models are usable from your laptop.
• Job tracking with phases — track a unit of work from research → design → code → review, each job in its own folder with a brief, a phase log, and captured lessons. Commands: /baton:job-start, /baton:job-status, /baton:job-list, /baton:job-phase, /baton:job-resume, /baton:job-lesson. (spec)
• A searchable knowledge base — your lessons and decisions become a meaning-based (semantic) search index, fully local — no cloud, no cost. Commands: /baton:kb-index, /baton:kb-search. (spec)
• Multi-model research ensembles — ask one question to several models at once and get a single synthesis of where they agree, differ, and what's uniquely useful. Commands: /baton:ensemble, /baton:research. (spec)
• Six Thinking Hats — examine a question from six fixed angles (facts, feelings, risks, benefits, creativity, process), then a synthesized conclusion. Command: /baton:six-hats. (spec)
• LLM Council — a two-round deliberation where models answer, then refine after