Benchmark-Driven Routing
for OpenClaw

What the Router Does

Most routing solutions use "simple vs complex" heuristics or generic capability tiers to pick a model. That's a guess dressed as a system. The OpenMark router for OpenClaw takes a different approach: it uses your actual benchmark results to make every routing decision.

You benchmark your recurring tasks on OpenMark AI, export the results, and the router uses that data to match incoming prompts to the best-performing model for each task category. No keyword matching, no complexity scoring — just measured performance on your real work.

What You See

You send a prompt. The router classifies it, finds the matching benchmark, picks the winner, and the routed model answers — all in a single turn. A routing card shows what happened:

The routed model generates the real reply. The classifier only identifies the task category — it never produces the user-visible answer.

How It Works

The plugin uses an internal two-phase architecture. To the user, it looks like a single reply.

No direct provider API calls from the plugin. Classification goes through the OpenClaw gateway. Provider authentication and model execution stay inside OpenClaw — you don't hand API keys to the router.

Quick Start

1. Benchmark your recurring tasks on OpenMark AI — test across 100+ models with deterministic scoring.
2. Export — click Export → OpenClaw on the Results tab. The CSV includes dual model keys, scores, costs, and metadata.
3. Install — run openclaw plugins install openmark-router and restart the gateway.
4. Import — place CSVs in the benchmarks directory, or use the local dashboard's import flow. The router activates automatically.

Five Routing Strategies

Choose how the router ranks models from your benchmark data:

All strategies use a 6-step cascade sort and a viability floor (max(top_score - 15pp, top_score * 0.5)) to exclude underperforming models. Fallback models are ranked from the same benchmark data.

Why Custom Benchmarks Matter for Routing

Every routing solution that uses generic categorization breaks in practice. "Email tasks" lumps cold outreach, complaint triage, and legal notices together — but model performance varies dramatically across these subtypes.

Generic benchmarks are equally broad. MMLU, Arena Elo, and HumanEval test general capabilities. A model scoring well on "writing" tells you nothing about your email templates with your tone requirements.

When you benchmark on OpenMark AI, you test models on your specific task, with your prompts, against your criteria. That's the data the router needs to make decisions you can trust.

Local Dashboard

The router ships with a local dashboard at http://127.0.0.1:2098/dashboard. From there you can:

• View router health, version, and detected providers
• See active benchmark categories and data freshness
• Change routing strategy and toggle routing cards
• Import and manage benchmark CSVs directly
• View benchmark category descriptions from loaded metadata

Works With Your Existing Setup

The router detects which providers your OpenClaw install can use and filters benchmark candidates accordingly. Direct provider keys are preferred first; if a model's direct provider isn't available but OpenRouter is, the router falls back to the OpenRouter key for that row.

Single-provider setups still benefit — you can benchmark and route within one provider's model lineup. The router is also useful with subscriptions, hosted access, or OAuth-backed providers, as long as OpenClaw can execute the model IDs involved.