ACRouter
Category: AI Infrastructure
ACRouter is an open-source AI model router that uses execution feedback (Context-Action-Feedback loop) to dynamically select the best LLM per coding task, achieving 2.6x cost savings over always-using-premium-model setups. ACRouter was founded in 2026. Based in Singapore / Hangzhou, China. Latest round: Bootstrapped.
- Founded
- 2026
- Headquarters
- Singapore / Hangzhou, China
Value proposition
ACRouter closes the information deficit in static LLM routing by accumulating execution-grounded experience during deployment via a Context-Action-Feedback loop, enabling adaptive model selection that outperforms static routers and single-model baselines on both cost and quality.
Products and solutions
ACRouter (agentic model routing framework), CodeRouterBench (~10K task benchmark with 8 frontier LLMs), Qwen3.5-0.8B PEFT/LoRA router adapter, Claude Code Router integration, cc-switch proxy-level integration, API Coding Solver demo
Unique value
First model routing framework that treats routing as an online learning problem (C-A-F loop) rather than static classification, achieving 15.3% relative gain just by adding performance statistics, and lowest cumulative regret on both in-distribution and out-of-distribution coding tasks.
Target customer
AI/ML engineers and developers using coding agents (Claude Code, Codex, Opencode) who want to optimize cost and performance across multiple LLM providers
Industries served
AI/ML Infrastructure, Software Development, Coding Agents
Technology advantage
Context-Action-Feedback (C-A-F) loop architecture with Orchestrator, Verifier, and Memory modules; contextual multi-armed bandit formulation with cumulative regret metric; execution-grounded feedback accumulation; open-source with ready-to-use integrations for Claude Code Router and cc-switch
How they differentiate
Unlike static routers (RouteLLM, OpenRouter) that treat model selection as one-shot classification, ACRouter uses an iterative C-A-F loop that learns from execution feedback, accumulating verified experience across tasks. It also introduces cumulative regret as a streaming metric and provides CodeRouterBench for standardized evaluation.
Main competitors
RouteLLM (open-source LLM routing), OpenRouter (commercial model routing API), Inworld Router (AI gateway/router), vLLM Semantic Router (Red Hat)
Key partnerships
National University of Singapore (NUS), Alibaba DAMO Academy, Zhejiang University, UC Berkeley, HKUST, Hugging Face (CodeRouterBench dataset and router adapter hosted)
Major milestones
June 2026: Paper released on arXiv (arXiv:2606.22902), June 2026: CodeRouterBench released with ~10K task instances and 8 frontier LLMs, July 2026: Covered by VentureBeat for 2.6x cost savings over Opus-only setups, Claude Code Router and cc-switch integrations released
Growth metrics
GitHub: ~514 stars, ~15 forks (agent-as-a-router repo); HuggingFace: 27 downloads (router adapter); Paper submitted June 2026 (arXiv:2606.22902)
Market positioning
Research-stage open-source project positioned as the most advanced agentic model routing framework for coding tasks, with demonstrated superiority over both static routers and single-model baselines in academic benchmarks
Geographic focus
Global (open-source project with contributors from Singapore, China, US)
Latest news about ACRouter
More AI Infrastructure companies
Official website: https://omnisource.cn/agent-as-a-router