Every prompt finds the cheapest model that can do it.

One local board for all your repos. Routine work falls to cheap models. Hard work stays with Claude or Grok.

One-time. Your own keys. No token markup. The v0.1.0 installer is unsigned, so SmartScreen warns once.

Illustrative board. Example numbers, not measured savings.

A real cascade, not a guess.

A local classifier splits each prompt. The routine part falls to the cheapest capable model in OpenRouter's live-priced catalog. Only the hard remainder reaches Claude or Grok, with the cheap work attached. If a model fails, the next ranked one takes over.

$ rename userCount, then fix the failing auth test

Classify locally

no API call · split route

$0.000

Rename across 6 files

cheapest capable · GLM 5.3

$0.004

Auth test root cause

the hard remainder

Claude
Same prompt, all-Claude: $0.71This route: $0.21

Example costs for illustration. Real prices come from OpenRouter at run time.

One window. Fewer expensive tokens.

No more opening a terminal, changing folders, and starting another agent for every repo.

A modular board

Each chat is tied to its repository. Other chats keep running while you switch, and Ctrl+Tab moves between projects.

A savings ledger

Routed tokens and estimated cost are recorded, so every saving can be inspected.

Your keys, no markup

Pay OpenRouter, Claude, and Grok directly.

0%

Free rankings for real work.

Dated, evidence-first rankings of models, IDEs, agent skills, and MCP servers. No paywall.

Or choose your developer class and get a setup prompt →

What works today.

Shipped behavior, separated from unfinished work. The pending column is a roadmap, not a guarantee.

Works today

  • Local zero-API classifier (classify, route --dry-run)
  • Cascade routing to the cheapest capable live-priced model
  • Automatic fallback across ranked candidates
  • Savings ledger, statistics, and response cache
  • Downloadable Windows desktop app (v0.1.0)
7 more shipped features
  • Model tiering (small, medium, large)
  • Coding policy tiers with independent review and inspectable promotion rules
  • Dated bang-for-buck leaderboard with JSON and CSV feeds
  • Hard per-call token cap (the Ringer)
  • Global nudge hook and hook-log audit trail
  • Usage pace and quota estimate in the CLI and dashboard
  • Desktop launchers for Claude, Codex, Grok, and OpenCode

Being wired up

  • Skill self-triggering is near 0.1% real-route conversion on measured nudges
  • No code signing yet, so SmartScreen warns on first launch
  • No native macOS or Linux desktop package yet

Pending code or evidence

  • A week or more of organic use for a real before-and-after
  • Enough harness runs to replace every public snapshot prior
  • A general answer to long-thread input compounding

Runs on your machine.

The board binds to 127.0.0.1. This website never connects to your localhost.

CLI

Windows, macOS, and Linux. Installs to ~/.waterfall and merges the Claude Code hooks with a backup.

curl -sSL https://raw.githubusercontent.com/adamtpang/waterfall.sh/main/install.py | python3

Windows app

The v0.1.0 release as an NSIS installer or MSI. Unsigned, so SmartScreen warns once.

waterfall_0.1.0_x64-setup.exe

Read the code first

The public repo holds the site, router, tests, installers, and desktop sources.

git clone https://github.com/adamtpang/waterfall.sh
cd waterfall.sh && pip install -e .

$30 once. Yours to keep.