tokenwise
Measurement-driven model router for Claude Code. Routes Haiku/Sonnet/Opus per task class, logs every routed task with real $ numbers, and A/B tests cheaper tiers before you trust the savings.
- License 许可
- MIT
- 最近更新
- 2026-08-01
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- tokenwise
- risk
- critical
- tags
- [ "model-routing", "token-optimization", "cost-reduction", "anthropic", "haiku", "sonnet", "opus", "claude-code", "ab-testing", "measurement" ]
- tools
- [ "claude" ]
- author
- CodeShuX
- plugin
- { "targets": { "codex": "blocked", "claude": "blocked" } }
- source
- community
- license
- MIT
- category
- developer-tools
- date_added
- 2026-05-12
- description
- Measurement-driven model router for Claude Code. Routes Haiku/Sonnet/Opus per task class, logs every routed task with real $ numbers, and A/B tests cheaper tiers before you trust the savings.
- source_repo
- CodeShuX/tokenwise
- source_type
- community
- license_source
- https://github.com/CodeShuX/tokenwise/blob/main/LICENSE
03 SKILL.md 正文
TokenWise — Measurement-Driven Model Router
Overview
A Claude Code skill that auto-routes subtasks to the cheapest model that can handle them (Haiku for grunt work, Sonnet for scoped reasoning, Opus only for synthesis), then logs every routed task to a local NDJSON with real token + cost numbers. Includes an A/B test subcommand that runs the same task across multiple tiers and scores quality, so the routing decisions are verified against the user's real workload — not estimated.
Anthropic's own bug tracker (Issue #27665) reports 93.8% of Max-subscriber Claude Code tokens flow to Opus. Existing routers (claude-router, wshobson, VoltAgent) either pin models statically or route by vibes-based heuristics with no measurement. TokenWise fills the measurement gap.
When to use
- Cutting Claude Code token spend without sacrificing output quality
- Validating whether Haiku/Sonnet is "good enough" for a specific task class before trusting auto-routing
- Auditing where Opus tokens are actually being burned
- Logging per-session cost data for finance or chargeback
Subcommands
/tokenwise:install— guided installer with diff preview, automatic backups, and--dry-runmode/tokenwise:report— per-session token + cost summary vs all-Opus baseline/tokenwise:summary [--week|--month|--all]— historical aggregate with trend/tokenwise:ab "<task>"— A/B test the same task at multiple tiers, generates a markdown comparison/tokenwise:undo— restore CLAUDE.md / settings.json from backup
Routing taxonomy
| Tier | Model | Task class |
|---|---|---|
| Mechanical | Haiku 4.5 | file reads, grep, format, rename, simple edits, doc lookups |
| Scoped reasoning | Sonnet 4.6 | single-file refactor, scoped research, test writing |
| Synthesis | Opus 4.7 | architecture decisions, multi-file refactor, security review |
Safety caps:
- Haiku never spawns further subagents
- Max spawn depth = 2
- Subagents that need a smarter model return to parent — they never escalate on their own
- Tasks under 100 chars with no file context run inline (subagent overhead > savings)
- Subagent context >30k tokens bumps a tier
Privacy
Zero telemetry. All logs in .tokenwise/log.ndjson local to the project. Task descriptions truncated to 80 chars and stripped of file contents before logging. No analytics endpoint exists in the source.
Install
In any Claude Code session:
/plugin marketplace add CodeShuX/tokenwise
/plugin install tokenwise@tokenwise
Then run /tokenwise:install and follow the guided prompts.
Limitations
- Token counts approximate to ±2% vs Anthropic billing
- A/B test mode costs extra tokens (one task × N tiers) — intentional one-time validation
- Anthropic-only by design (use LiteLLM or OpenRouter for cross-vendor)
- Subagent
model:param has known silent-fail bugs on some Claude Code builds — skill probes for this at install and refuses to configure if routing is broken
Source
- Repo: https://github.com/CodeShuX/tokenwise
- License: MIT
- Author: CodeShuX
04 同仓库更多 Skills
zustand-store-ts
Create Zustand stores following established patterns with proper TypeScript types and middleware.
zoom-automation
Automate Zoom meeting creation, management, recordings, webinars, and participant tracking via Rube MCP (Composio). Always search tools first for current schemas.
zoho-crm-automation
Automate Zoho CRM tasks via Rube MCP (Composio): create/update records, search contacts, manage leads, and convert leads. Always search tools first for current schemas.
zod-validation-expert
Expert in Zod — TypeScript-first schema validation. Covers parsing, custom errors, refinements, type inference, and integration with React Hook Form, Next.js, and tRPC.
zipai-optimizer
Adaptive token optimizer: intelligent filtering, surgical output, ambiguity-first, context-window-aware, VCS-aware, MCP-aware.
zeroize-audit
Detects missing zeroization of sensitive data in source code and identifies zeroization removed by compiler optimizations, with assembly-level analysis, and control-flow verification. Use for auditing C/C++/Rust code handling secrets, keys, passwords, or other sensitive data.