TokenRouter

Open-source AI model routing platform — BYOK (Bring Your Own Key), auto-select the optimal model based on prompt content. Self-hosted, zero data leakage.

TokenRouter analyzes your prompt, classifies the task type (coding, math, translation, etc.), estimates complexity, and routes to the best model — saving 60-80% on API costs without sacrificing quality.

Why TokenRouter?

Without TokenRouter	With TokenRouter
Every request → expensive model (GPT-5.2 / Claude Opus)	Simple questions → cheap model, complex ones → premium
Manual model switching in code	One API key, automatic routing
No visibility into costs	Real-time cost tracking + savings dashboard
Single provider failure = downtime	Automatic fallback chain

Features

• 🧠 Smart Routing — L1 regex + L2 model classifier, auto-selects optimal model per request
• 🔑 Unified API Key — One tr-xxx key, routes to OpenAI/Anthropic/Google/DeepSeek behind the scenes
• 💰 Cost Tracking — Every request logged with actual cost vs baseline, see exactly how much you save
• 📊 Built-in Dashboard — Dark-themed web UI with cost charts, model distribution, request logs
• 🔄 Fallback Chain — Automatic retry with fallback models on failure (429/5xx)
• 🌐 OpenAI-Compatible — Drop-in replacement for any OpenAI SDK
• 🏠 Self-Hosted — Your keys never leave your server
• ⚡ 3 Strategies — cheapest, best, balanced (quality/cost ratio)

Quick Start

Install

pip install byok-router

Option 1: Self-Hosted Platform (Recommended)

Set up once, use everywhere:

# 1. Create a unified API key
tokenrouter keys create --name "my-project" --strategy balanced

# Output: Created key tr-a1b2c3d4... (id: 1)

# 2. Add your provider keys
tokenrouter keys add-provider 1 openai sk-...
tokenrouter keys add-provider 1 anthropic sk-ant-...
tokenrouter keys add-provider 1 deepseek sk-...

# 3. Start the server
tokenrouter serve --port 8000

# Dashboard: http://localhost:8000
# API: http://localhost:8000/v1

Now use your tr-xxx key with any OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="tr-a1b2c3d4..."  # your unified key
)

response = client.chat.completions.create(
    model="auto",  # TokenRouter picks the best model
    messages=[{"role": "user", "content": "Write a Python quicksort"}]
)

print(response.model)  # "deepseek-chat" (cheapest for coding)
# Check headers for cost info:
# X-TokenRouter-Cost: 0.00032
# X-TokenRouter-Saved: 0.04500
# X-TokenRouter-Saved-Pct: 99.3

Option 2: Python SDK (Direct)

from tokenrouter import TokenRouter

router = TokenRouter(
    keys={
        "openai": "sk-...",
        "anthropic": "sk-ant-...",
        "deepseek": "sk-...",
    },
    strategy="balanced",
)

response = router.chat([
    {"role": "user", "content": "Write a Python quicksort function"}
])
print(response.model_used)  # "deepseek-chat"
print(response.choices[0].message["content"])

# Streaming
for chunk in router.chat_stream([
    {"role": "user", "content": "Explain quantum computing"}
]):
    delta = chunk.choices[0].get("delta", {})
    print(delta.get("content", ""), end="")

Option 3: YAML Config

# config.yaml
admin_token: ${TOKENROUTER_ADMIN_TOKEN}
database: ~/.tokenrouter/keys.db
baseline_model: claude-opus-4  # cost comparison baseline

# Legacy mode: direct keys (single user, no tr-key needed)
keys:
  openai: ${OPENAI_API_KEY}
  anthropic: ${ANTHROPIC_API_KEY}
  deepseek: ${DEEPSEEK_API_KEY}

strategy: balanced

rules:
  - task: coding
    model: deepseek-chat
  - task: chinese_language
    model: qwen-max

exclude_models:
  - claude-opus-4

proxy:
  host: 0.0.0.0
  port: 8000
  cors: true

tokenrouter serve --config config.yaml

Dashboard

Access the built-in dashboard at http://localhost:8000 after starting the server.

What you see:

• 📈 Total requests, cost, savings, and savings percentage
• 📉 Daily cost vs baseline line chart
• 🍩 Model usage distribution
• 📋 Recent request logs with task type, model, cost, and latency
• 🔑 Key management (create, view, delete)

Cost Tracking

Every request is tracked with:

Field	Description
actual_cost	What you actually paid
baseline_cost	What it would cost with Claude Opus / GPT-5.2
saved	baseline_cost - actual_cost
saved_pct	Savings percentage

# View usage from CLI
tokenrouter usage 1  # key id

# API endpoint
curl http://localhost:8000/v1/usage/1/savings \
  -H "Authorization: Bearer $ADMIN_TOKEN"

Example savings report:

{
  "total_requests": 1247,
  "total_cost": 1.83,
  "baseline_cost": 12.45,
  "saved": 10.62,
  "saved_pct": 85.3,
  "period": "month"
}

API Key Management

CLI

tokenrouter keys create --name "production" --strategy balanced
tokenrouter keys list
tokenrouter keys add-provider 1 openai sk-...
tokenrouter keys add-provider 1 anthropic sk-ant-...
tokenrouter keys delete 1

REST API

# Create key (requires admin token)
curl -X POST http://localhost:8000/v1/keys \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"name": "my-key", "strategy": "balanced"}'

# List keys
curl http://localhost:8000/v1/keys \
  -H "Authorization: Bearer $ADMIN_TOKEN"

# Add provider key
curl -X POST http://localhost:8000/v1/keys/1/providers \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"provider": "openai", "api_key": "sk-..."}'

# Get savings report
curl http://localhost:8000/v1/usage/1/savings \
  -H "Authorization: Bearer $ADMIN_TOKEN"

How Routing Works

User Request → L1 Classifier → L2 Classifier (if needed) → Routing Table → Strategy Selection → Provider
                  (<1ms)           (~200ms)                  (24 combos)     (cheapest/best/     → Fallback
                                                                              balanced)             Chain

1. L1 Classifier — Regex heuristic scores 8 task categories (coding, math, translation, creative writing, summarization, reasoning, QA, Chinese)
2. L2 Classifier — When L1 confidence < 0.7, uses a cheap model to refine classification
3. Routing Table — Maps task_type:complexity to candidate models
4. Strategy Selection — Picks from candidates using capability matrix (each model scored 1-5 on 8 dimensions)
5. Fallback Chain — If primary model returns 429/5xx, auto-retries with next candidate

Supported Models

Provider	Models	Best For	Cost (per 1M tokens)
OpenAI	GPT-5.2, GPT-5 Mini	General, coding	$1.75-14 / $0.25-2
Anthropic	Opus 4/4.5, Sonnet 4, Haiku 4.5	Coding, writing	$1-15 / $5-75
Google	Gemini 3 Flash, 2.5 Flash/Pro	Quick tasks	$0.30-1.25 / $2.50-10
DeepSeek	V3.2, R1	Coding, math	$0.28 / $0.42
Moonshot	Kimi K2.5	Chinese, coding	$0.60 / $2.50
Qwen	Turbo, Plus, Max	Chinese tasks	$0.05-1.60 / $0.20-6.40
Zhipu	GLM-4 Plus	Chinese, QA	$0.50 / $1.50

Use with OpenClaw

TokenRouter is designed to work seamlessly with OpenClaw. Set up your tr-xxx key and point OpenClaw to your TokenRouter instance:

# OpenClaw config
model:
  provider: openai-compatible
  baseUrl: http://localhost:8000/v1
  apiKey: tr-a1b2c3d4...
  model: auto

Development

git clone https://github.com/aplomb2/TokenRouter.git
cd TokenRouter
pip install -e ".[dev]"
pytest  # 182 tests

License

MIT — see LICENSE.