Can an Open‑Source Router Cut AI Agent Costs by 60% and Keep Sensitive Data Local?

AI agents are increasingly integrated into workflows, but their current usage exposes three critical issues: (1) sending sensitive customer data to third‑party cloud servers risks privacy leaks; (2) even simple tasks trigger expensive top‑tier models, inflating costs; (3) local models are cheap but often lack the capacity to handle complex tasks, leading to failures.

ClawXRouter, jointly released by Tsinghua University THUNLP, Renmin University, AI9Stars, MianBan Intelligent, and OpenBMB, addresses these problems through end‑cloud collaboration. It routes each request to the most appropriate execution path, keeping privacy‑sensitive data on the device while delegating complex inference to powerful cloud models.

Routing mechanism consists of three levels:

S3 (Private) : Physically isolates highly sensitive data such as SSH keys or passwords; the request is processed entirely offline on the local model, never reaching the cloud.

S2 (Sensitive) : Detects data like internal‑network IP logs or contact phone numbers, applies intelligent redaction (e.g., replacing "王小二" with [REDACTED:NAME]), then forwards the sanitized request to the cloud.

S1 (Secure) : Handles ordinary queries (e.g., HTTP 403 vs 401) by sending them directly to the cloud for full‑strength inference.

A rule‑plus‑model dual‑detection engine ensures fast and accurate routing decisions.

Cost‑aware routing employs a lightweight local model as an "LLM‑as‑Judge" to evaluate task complexity and dispatch it to the most cost‑effective model, avoiding the "using a cannon to kill a chicken" scenario.

Dual‑track memory preserves privacy while maintaining workflow continuity: the cloud sees only the redacted conversation history ( MEMORY.md), whereas the local side retains the full context ( MEMORY‑FULL.md).

The plugin also offers a composable routing pipeline with ten hooks that cover the entire request lifecycle without invasive changes to existing OpenClaw flows, and a bilingual dashboard visualizing usage, session logs, routing rules, and model configuration.

Installation (example commands):

pnpm add -w @openbmb/clawxrouter

openclaw plugins install clawhub:clawxrouter

ollama pull openbmb/minicpm4.1

ollama serve

openclaw gateway

After deployment, benchmark tests on PinchBench (23 OpenClaw Agent tasks) show a 58% reduction in cost and a 6.3% performance improvement.

The authors conclude that developers no longer need to choose between cloud and edge; ClawXRouter enables each side to play to its strengths, and the project will continue to evolve as an open‑source effort.