Iblai Openclaw Router

Cost-optimizing model router for OpenClaw. Automatically routes each request to the cheapest capable Claude model (Haiku/Sonnet/Opus) using weighted scoring....

Security Scan

The skill largely does what it says (a local routing proxy) but the package metadata, install behavior, and secret handling are inconsistent and risky — review before installing.

Purpose & Capability

The code and README implement a local Anthropic/OpenRouter proxy which matches the stated purpose. However the registry metadata declared 'no required env vars' and 'instruction-only' while the package includes server.js and install scripts that require an ANTHROPIC_API_KEY and create a systemd service — an incoherence between claims and actual requirements.

Instruction Scope

SKILL.md and README instruct running scripts that will copy files to ~/.openclaw/workspace/router, create/enable a systemd service, and modify ~/.openclaw/openclaw.json to register the provider. The installer also attempts to auto-detect an Anthropic key by grepping ~/.openclaw/agents/main/agent/auth-profiles.json. Reading/writing other agent config and extracting API keys is outside a purely 'instruction-only' description and broadens scope.

Install Mechanism

Install is local (no external downloads) and just copies files and creates a systemd unit. That itself is reasonable for a local proxy, but the installer writes the detected Anthropic API key directly into the systemd unit (Environment=ANTHROPIC_API_KEY=...). Embedding secrets into unit files can expose them to other users/processes and is poor secret management.

Credentials

The package metadata declares no required env vars, but server.js requires ANTHROPIC_API_KEY and the install script actively searches for the user's Anthropic key and places it into the unit file. The README also suggests configuring 'apiKey passthrough' in OpenClaw, but the router ignores incoming x-api-key and uses the environment key — this mismatch is confusing and potentially leads to unexpected key usage/exposure.

Persistence & Privilege

The installer creates a persistent systemd service (Restart=always) and registers a model provider in OpenClaw config. Persistent local services and config changes are expected for this functionality, but combined with the secret-handling behavior above this persistence increases the blast radius if misconfigured.

What to consider before installing

This package implements a local proxy and will install a systemd service and modify your OpenClaw config. Before installing: 1) Inspect server.js and config.json yourself to confirm behavior and upstream URLs. 2) Edit scripts/install.sh to avoid auto-extracting your API key (the script greps ~/.openclaw/agents/main/agent/auth-profiles.json) — do not allow the installer to embed secrets into /etc/systemd/system/*.service. Prefer using an EnvironmentFile with restricted permissions or systemd secret mechanisms, or export ANTHROPIC_API_KEY at runtime instead. 3) Back up ~/.openclaw/openclaw.json and test registration manually rather than letting the installer modify it automatically. 4) If you must run it, run in an isolated environment (container or VM) first and verify that the router only forwards to your configured LLM provider. 5) After install, check /etc/systemd/system/iblai-router.service for embedded secrets and remove/rotate any secret that was written there. The inconsistencies (metadata claiming no env vars / instruction-only vs included code that needs a key and writes configs) are likely sloppy engineering but could expose secrets if installed without care.

iblai-router

A zero-dependency proxy that sits between OpenClaw and the Anthropic API, routing each request to the cheapest capable model using a 14-dimension weighted scorer (<1ms overhead).

Install

Run the install script to set up everything automatically:

bash "$(dirname "$0")/scripts/install.sh"

This will:

1. Copy server.js and config.json to ~/.openclaw/workspace/router/
2. Create and start a systemd service (iblai-router) on port 8402
3. Register iblai-router/auto as an OpenClaw model provider

After install, iblai-router/auto is available anywhere OpenClaw accepts a model ID.

Verify

curl -s http://127.0.0.1:8402/health | jq .
curl -s http://127.0.0.1:8402/stats | jq .

Use

Set iblai-router/auto as the model for any scope:

Scope	How
Cron job	Set model to iblai-router/auto in job config
Subagents	agents.defaults.subagents.model = "iblai-router/auto"
Per-session	/model iblai-router/auto
All sessions	agents.defaults.model.primary = "iblai-router/auto"

Tip: Keep the main interactive session on a fixed model (e.g. Opus). Use the router for cron jobs, subagents, and background tasks where cost savings compound.

Customize

All config lives in ~/.openclaw/workspace/router/config.json and hot-reloads on save — no restart needed.

Models

Change the models per tier:

{
  "models": {
    "LIGHT":  "claude-3-5-haiku-20241022",
    "MEDIUM": "claude-sonnet-4-20250514",
    "HEAVY":  "claude-opus-4-20250514"
  }
}

Non-Anthropic models (OpenAI, Google)

Set apiBaseUrl to route through OpenRouter:

{
  "models": {
    "LIGHT":  "openai/gpt-4.1-mini",
    "MEDIUM": "openai/gpt-4.1",
    "HEAVY":  "openai/o3"
  },
  "apiBaseUrl": "https://openrouter.ai/api/v1"
}

Update the API key in the systemd service when switching providers, then systemctl daemon-reload && systemctl restart iblai-router.

Scoring

Keyword lists control which tier handles a request:

• simpleKeywords, relayKeywords → push toward LIGHT (cheap)
• imperativeVerbs, codeKeywords, agenticKeywords → push toward MEDIUM
• technicalKeywords, reasoningKeywords, domainKeywords → push toward HEAVY (capable)

Tune boundaries and weights in config.json to match your workload. See the full README for details.

Uninstall

bash "$(dirname "$0")/scripts/uninstall.sh"

Stops the service, removes the systemd unit, and deletes router files. Reminder: switch any workloads using iblai-router/auto back to a direct model first.