Optimize Your API Costs - Route tasks to cost-effective models via TokenRouter. Dynamically builds model tiers from live pricing, classifies tasks into 6 categories (supporting English & Chinese), and orchestrates multi-agent workflows with adaptive fallback.
Pre-flight Check: PaleBlueDot AI TokenRouter Provider Configuration
This skill requires a PaleBlueDot AI TokenRouter provider to be configured. Before using ANY feature, the agent MUST perform the following check-and-setup flow.
Phase 1: Check Existing Provider
Read the configuration file at ~/.openclaw/openclaw.json.
Look under models.providers for any provider whose baseUrlcontainshttps://open.palebluedot.ai (partial match is sufficient).
If found → Skip to Phase 3 (the provider is already configured).
If NOT found → Proceed to Phase 2.
Phase 2: Guide User & Write Configuration
When no TokenRouter provider is detected, the agent should:
Inform the user that TokenRouter is not configured and prompt them to register:
TokenRouter provider not configured. To use this skill, you need a PaleBlueDot AI TokenRouter account.
Please visit https://www.palebluedot.ai to register and log in, then go to the TokenRouter section to get your configuration (base URL and API key).
Once you have the information, please provide it to me and I will complete the configuration for you.
Wait for the user to provide the following information:
Base URL (e.g., https://open.palebluedot.ai/v1)
API Key (the user's real API key from PaleBlueDot AI)
(Optional)Provider name (defaults to tokenrouter if not specified)
Agent writes the configuration by running the setup command:
This writes the provider into ~/.openclaw/openclaw.json under models.providers with a backup of the original config.
Do NOT use placeholder API keys. The agent must only write the actual credentials provided by the user.
Phase 3: Re-verify & Auto-Sync Models
After the provider is confirmed (either pre-existing or just written in Phase 2), the agent MUST:
Re-verify the provider configuration by running:
text
check
This confirms the provider with baseUrl containing https://open.palebluedot.ai is properly saved in the config.
If verification fails, inform the user and go back to Phase 2.
If verification succeeds, automatically sync models by running:
text
sync
This will:
Fetch all available models from the TokenRouter API.
Add all model names to the TokenRouter provider's models array.
Add all models to the models.allowed list.
All models will be routed through the configured TokenRouter provider.
Display the synced model list to the user.
Confirm completion to the user:
TokenRouter configuration complete. {N} models have been synced and added to your allow list. All models are routed through the TokenRouter provider. You can now use list to view pricing or start planning tasks.
Automatic Pre-Planning Sync
Before executing any plan command, the system will automatically:
Verify TokenRouter Provider: Checks if a provider whose baseUrl contains https://open.palebluedot.ai exists in models.providers. If not found, the agent enters the Phase 2 setup flow described above.
Auto-Sync Models: Fetches the latest model list and updates the provider's models array and the models.allowed list.
Set Default Model: If no default model is set, picks the first available model from the synced list.
This ensures that the plan command always has access to the most up-to-date model information, and all models are routed through the user's TokenRouter provider.
Dynamic Model Tier System
Model tiers are not hardcoded. On every plan invocation, the system:
Sorts by price descending and splits into 3 equal buckets (high / mid / low).
Picks the median-priced model from each bucket to avoid outliers.
Tier
Role
Selection Rule
tier1 (high)
Architect / Reasoning
Median of top-third by price
tier2 (mid)
Coder / Drafter
Median of middle-third by price
tier3 (low)
Reviewer / Quick tasks
Median of bottom-third by price
If the API is unreachable or fewer than 3 known models are available, the system falls back to hardcoded defaults (claude-opus-4.6 / gpt-4o-mini / deepseek-v3.2).
6-Category Task Classification Engine
The plan command uses an enhanced classifier that supports both Chinese and English keywords. Tasks are scored against 6 categories; the highest-scoring category wins.
When the user's task is complex (e.g., building an application, designing a system, multi-step workflows), proactively suggest multi-agent routing:
Identify complexity — If the task involves multiple phases, recommend running plan.
Show the pricing — Run list to display the full price list.
Run the planner — Run plan "<task>" to show the recommended routing with projected savings.
Offer to enable models — Ask the user if they want to enable the recommended models as fallbacks.
Example prompt to user:
Your task involves multiple phases. Let me run the planner to find the optimal routing:
text
build a REST API with authentication / 帮我构建一个REST API的鉴权功能
The planner will automatically pick the best models for each phase based on current pricing and show projected savings.
Quick Start
text
# Step 1: Check if TokenRouter is configured
check
# Step 2: If not configured, set it up (agent does this with user-provided credentials)
setup --name tokenrouter --base-url https://open.palebluedot.ai/v1 --api-key sk-xxx...
# Step 3: Verify and sync all models
check
sync
# List all models with real-time pricing
list / 列出TokenRouter的模型价格列表
# Get routing recommendations for a task (Chinese or English)
# NOTE: This will automatically sync models before planning
write a Python script
帮我开发一个用户管理后端接口
analyze and compare the data reports of three competing products
把这段中译英
# Enable a model by index or name
enable 1
enable openai/gpt-4o-mini
# Generate execution plan for host agent to dispatch sub-agents
build a todo app
Core Functions
1. check - Verify Provider Configuration
text
check / 检查TokenRouter配置
Verifies that a TokenRouter provider (baseUrl containing https://open.palebluedot.ai) exists in the config. Displays provider details if found, or guides the user to set up if not.
Writes the TokenRouter provider into ~/.openclaw/openclaw.json (with automatic backup). The agent uses this command after the user provides their credentials from PaleBlueDot AI. The agent must never use placeholder API keys — only real credentials provided by the user.
3. sync - Fetch & Sync All Models
text
sync / 同步TokenRouter模型
Fetches all available models from the TokenRouter API, adds them to the provider's models array and the models.allowed list. All models are routed through the configured TokenRouter provider. This is automatically called before plan.
4. list - Real-Time Model Pricing
text
list / 列出模型的价格列表
Fetches current TokenRouter pricing and displays all available models with input/output/cache prices.
5. plan - Smart Task Routing
text
plan "<task description>" / 计划 "<任务描述>"
plan "<task description>" --execute / 计划 "<任务描述>" 并执行
Classifies the task, builds a multi-phase pipeline, assigns dynamic model tiers, and shows projected savings. Automatically syncs models before planning. Add --execute (or -x) to generate a structured JSON execution plan (swarm_plan.json) that the host agent uses to dispatch sub-agents via the internal sessions_spawn API.
How --execute works:
The --execute flag does NOT call sub-agents directly via CLI. Instead, it outputs a JSON plan to ~/.openclaw/workspace/swarm_plan.json and to stdout. The host agent (OpenClaw) reads this plan and dispatches sub-agents internally using sessions_spawn. Each step in the plan contains:
model: which model to use for this phase
system_prompt: the role prompt for the sub-agent
task_prompt: the task instruction
expected_artifact: the file the sub-agent should produce
max_retries and timeout_seconds: retry and timeout policy
Steps must be executed sequentially — each step's artifact is context for the next.
6. enable - Auto-Configuration
text
enable 1 # Enable model by index
enable openai/gpt-4o-mini # Enable by name
启用 1
启用 openai/gpt-4o-mini
Writes the model directly into ~/.openclaw/openclaw.json (with automatic backup).
Adaptive Stability Fallback
The system tracks historical model performance via swarm_memory.json and swarm_insights.json:
Tier selection: If a tier2/tier3 model's historical success rate drops below 50%, it is automatically replaced by the tier1 model for that phase during plan generation.
Retry policy: The generated execution plan specifies max_retries: 2 per step. The host agent should inject the error context into the retry prompt when re-dispatching a failed step.
Logging: The host agent should append execution results to swarm_memory.json. Run consolidate_memory.py to generate performance insights for future adaptation.
Advanced Usage
Custom Routing Rules
You can override the dynamic tier selection for specific categories by creating ~/.openclaw/model-routing.json. Each category maps to an ordered list of model IDs, one per phase (matching the pipeline order). An optional "fallback" model is used for any phase without an explicit override.
