model-matchmaker

model-matchmaker

Effort-aware prompt routing for Claude Code. Classifies prompt complexity and suggests the right model + reasoning effort so you're not burning Opus tokens on "yes" or giving Haiku an architecture task.

What it does

model-matchmaker classifies every prompt into a complexity tier and injects a suggestion for which model and reasoning effort to use. You still switch models manually — the plugin tells you when and what to switch to.

Tier	Triggers	Suggestion
LOW (trivial)	Affirmatives, git ops, file ops, mechanical tasks	/model haiku
LOW (simple)	Short prompts (≤6 words, no question)	/model sonnet with low reasoning effort
STANDARD	Everything else	No suggestion — current model at natural depth
HIGH	Architecture, security, cross-cutting, planning, very long prompts	/model opus with high reasoning effort (or ultrawork prefix)

Override

Prefix any prompt with \ to skip the suggestion for that turn.

Why not auto-switch?

The Claude Code hooks API doesn't support programmatic model switching. Rather than running a proxy server (which breaks OAuth on Max plans), model-matchmaker works within the plugin system to give you actionable suggestions before Claude executes.

Installation

From marketplace

/plugin marketplace add jacobcxdev/cc-plugins
/plugin install model-matchmaker@jacobcxdev-cc-plugins

Local development

claude --plugin-dir ~/Developer/src/github/jacobcxdev/model-matchmaker

Configuration

With GSD (Get Shit Done)

The plugin reads your GSD profile automatically:

/gsd:set-profile quality   # Thorough by default, LOW only for trivial
/gsd:set-profile balanced  # Effort matches task (default)
/gsd:set-profile budget    # Lean responses, HIGH suggests /model opus

Without GSD

Create ~/.config/model-matchmaker/config.json:

{
  "profile": "balanced"
}

Profile Behaviour

Profile	LOW threshold	HIGH threshold	Philosophy
quality	Only affirmatives, ≤6 words (no model switch suggestion)	Silent (already on Opus)	Pay compute tax, avoid correction tax
balanced	Trivial → Haiku, simple → Sonnet+low effort	Architecture/security → Opus+high effort	Smart routing — model matches task
budget	Silent (already on Sonnet)	Complex patterns → strongly recommend Opus	Maximum savings, escalate when it matters

Decision Log

Decisions are logged to ~/.local/state/model-matchmaker/decisions.log for tuning analysis:

2026-03-04T22:35:00Z | balanced | LOW    | "yes"
2026-03-04T22:36:00Z | balanced | STD    | "fix the bug in auth..."
2026-03-04T22:37:00Z | quality  | HIGH   | "architect a new auth..."

Development

pnpm install    # Install TypeScript
pnpm build      # Compile src/ → dist/
pnpm test       # Run tests (29 tests across all profiles)

Licence

MIT