paxon-yang/claude-agent-tools

**Automatic model routing and a live agent dashboard for Claude Code.**
Haiku answers the lookups, Sonnet does the daily work, Opus plans the big jobs —
and you watch every agent, model and dollar in real time.
**▶ Try the live demo** · English · 简体中文
Demo data. Open the live demo · full-quality video
Why
Claude Code uses one model for everything unless you keep typing /model. That means paying Opus prices to ask *"where is the login code?"*, or asking Sonnet to architect a migration. And when a session fans out into sub-agents, you can't see who is doing what.
This repo ships two small tools that fix both:
| | What it does |
|---|---|
| **auto-router** | A Claude Code plugin that picks the model and effort **every turn**, and for every sub-agent. Simple questions go to Haiku, everyday coding to Sonnet, architecture and multi-step jobs to Opus (which plans and delegates). A built-in report shows what it saves you versus running everything on Opus — measured on your own transcripts, not promised. |
| **agent-viz** | A local dashboard at http://localhost:4321: every project, the current task, a live tree of the main session and its sub-agents, which model each one runs, a model timeline, estimated cost and savings, and an alert when Claude is waiting for your approval. Works on your phone too. |
How it compares
| | **claude-agent-tools** | claude-code-router | ccusage | /model by hand |
|---|---|---|---|---|
| What it is | Plugin + local dashboard | Proxy in front of Claude Code | Usage analyzer CLI | Built in |
| Picks the model | **Every turn and every sub-agent**, automatically | By scenario (background, thinking, long context…), to any provider | — | You, when you remember |
| Stays on Anthropic models, no proxy | ✓ | Routes through its proxy | ✓ | ✓ |
| Sees sub-agents live | ✓ | — | — | — |
| Shows cost | Live, plus a savings-vs-Opus report | — | ✓ detailed usage & cost reports | — |
Use it alongside ccusage if you like its reports; use claude-code-router if you want non-Anthropic models. This project is for people who want to stay on Claude and stop overpaying for easy turns.
Features
- **Per-turn routing** — keyword rules first, then a Haiku classifier for the rest. Tag a prompt with #opus, #sonnet, #haiku or #fable to override once; /route opus to pin.
- **Multi-model teamwork** — big tasks ("implement docs/spec.md") get Opus as the orchestrator, explorer sub-agents on Haiku to read code, worker sub-agents on Sonnet to write it.
- **Plan → execute** — in plan mode (Shift+Tab) Opus writes the plan; once you approve, Sonnet carries it out.
- **Guard rails** — Haiku hands over to Sonnet before editing a 4th file or running risky commands (rm -rf, git push, migrations, deploys); 3 tool failures in a turn escalate one tier on the spot; downgrades need two confirmations.
- **Cache-aware** — the prompt cache belongs to one model, so a switch makes the new model re-read the whole conversation at the cache-write price. While the current model's cache is warm, the router only downgrades when the turn will earn that back (it learns whether your cache lasts 5 minutes or 1 hour); the status line shows how long the cache has left.
- **Test gate** — when Haiku or Sonnet changed code, your tests run before the turn ends. If they fail, the turn steps up to Opus and keeps fixing — once per turn, never in a loop. Detects npm / pnpm / yarn / bun, pytest, cargo, go and make test; skipped when the model already ran the tests itself.
- **/route card, on your phone too** — model and why, cache time left, test gate, running sub-agents and session cost as one card in the terminal, the desktop app, and the Claude app via Remote Control.