Describe your task. Get the right model. Stop over-paying for Opus on tasks Haiku handles fine.

---

Running every Claude task through Opus because you can't be bothered to think about it? That's a 15× price multiplier over Haiku on the tasks that don't need it. context-router reads your task description, scores it against 80+ complexity signals, picks the right model tier, explains why, and shows you the cost delta — before you commit a single token.

🧭 Context Router
────────────────────────────────────────
Task: "implement user authentication with JWT"

Analysis:
  Signals: sonnet (+6): implement, authentication, jwt | haiku (+1): check
  Confidence: 72% → Sonnet
  Tokens est: ~65 input, ~32 output

Routing Decision: ✓ SONNET
  Why: Implementation task with moderate complexity — Sonnet is the sweet spot
  Model ID: claude-sonnet-4-5

Cost Estimate:
  Haiku    $0.0001   ⚡ Fast & cheap — best for simple tasks
  Sonnet   $0.0004   ← selected
  Opus     $0.0020   ★ Maximum intelligence — best for complex reasoning

Savings vs Opus: $0.0016 (80% cheaper) 💰

Run with Sonnet? (y/n/h for haiku/s for sonnet/o for opus):

Install

No npm account, no global install — runs straight from GitHub with zero dependencies:

npx github:NickCirv/context-router "your task"

Usage

# Route a task — see model decision + cost estimate
npx github:NickCirv/context-router "implement user authentication with JWT"

# Route AND call the selected model, stream the response
npx github:NickCirv/context-router "fix typo in variable name" --run

# Show cost log (last 20 entries + running total)
npx github:NickCirv/context-router --log

# Weekly and monthly spend breakdown
npx github:NickCirv/context-router --stats

# Interactive calibration quiz — tune thresholds to your workflow
npx github:NickCirv/context-router --calibrate

# Clear the cost log
npx github:NickCirv/context-router --reset-log

Flag	Description
--run	Route + call the selected model, stream response, log actual cost
--log	Show last 20 logged calls with running total
--stats	Weekly and monthly spend breakdown by model
--calibrate	Interactive quiz to tune Haiku/Opus scoring boosts
--reset-log	Clear ~/.context-router-log.json

Routing logic

No AI involved — pure heuristics against 80+ keyword signals across three tiers:

Tier	Signal examples	Price range
Haiku	fix, rename, format, list, check, lookup, convert, parse, trim, validate, sort, encode, summarize, how do i, one line, trivial	$0.001/1K in · $0.001/1K out
Sonnet	implement, write, create, build, debug, refactor, review, optimize, test, integrate, deploy, migrate, authentication, api endpoint, middleware	$0.003/1K in · $0.015/1K out
Opus	architect, architecture, design system, strategy, roadmap, trade-off, security audit, threat model, compliance, distributed system, scalability, concurrency, root cause, comprehensive, full audit	$0.015/1K in · $0.075/1K out

Multi-word signals score 2 points; single-word signals score 1 point. Opus wins any tie it participates in (architectural decisions shouldn't be down-graded). Ambiguous ties default to Sonnet as the safe middle tier.

Tune the thresholds with --calibrate — answers to three questions persist score boosts to ~/.context-router-config.json.

Real-world savings

Scenario	Haiku	Sonnet	Opus
100× "fix typo" tasks	$0.10	$0.45	$1.50
50× "implement feature" tasks	—	$0.75	$7.50
10× "architect system" tasks	—	—	$1.50

Savings compound: a team running 500 tasks/day with mixed complexity can shift 40–60% to Haiku/Sonnet and halve their monthly Anthropic bill without touching model quality on tasks that actually need Opus.

API key

Set ANTHROPIC_API_KEY to use --run mode and stream real responses:

export ANTHROPIC_API_KEY=sk-ant-...

Without it, the tool still routes, explains, and estimates cost — useful as a dry-run decision aid before you paste the task into Claude.

Cost log

Every --run call is logged to ~/.context-router-log.json:

[
  {
    "date": "2026-03-02T10:00:00.000Z",
    "task": "implement user authentication with JWT",
    "model": "sonnet",
    "inputTokens": 65,
    "outputTokens": 420,
    "cost": 0.0063
  }
]

--stats aggregates this by week and month, broken down by model — so you can see whether your routing mix is drifting toward Opus over time.

What it is NOT

• Not a proxy or API wrapper. It calls the Anthropic API directly using Node's built-in https module when --run is set. No middleware, no rate-limiting layer, no request transformation.
• Not a guarantee of quality. Routing is heuristic. A task with no strong signals defaults to Sonnet. Override interactively with h, s, or o at the prompt, or force a tier with --run after accepting the suggestion.
• Not a cost billing system. The log tracks estimated + actual token costs locally. It does not connect to your Anthropic billing dashboard or enforce budgets.

---

_{Zero dependencies · Node 18+ · MIT · by NickCirv}