WorkWeave Router: High-Performance LLM Routing for Agentic Systems

The WorkWeave Router is a high-speed proxy that dynamically selects the optimal LLM for each request to balance cost and performance. It supports major providers like Anthropic and OpenAI while keeping API keys secure and encrypted on the user's infrastructure. Developers can quickly integrate it into tools like Claude Code or self-host the full stack for complete control and observability.

Key Points

• Dynamic request-level routing uses a local embedder to select the best model for every turn based on performance and efficiency.
• The router acts as a drop-in replacement for Anthropic, OpenAI, and Gemini APIs, ensuring compatibility with existing agentic tools.
• Security is prioritized through a BYOK (Bring Your Own Key) architecture where provider credentials stay local and encrypted.
• Built-in observability features include OTLP tracing and a Next.js-based admin dashboard for tracking cost savings and model metrics.
• Seamless integration is provided for popular developer environments including Claude Code, Cursor, and OpenAI Codex.

Sentiment

The overall sentiment is mixed and technically skeptical. HN largely agrees with the article that model selection and agentic coding cost are real problems, and there is clear interest in a router that can exploit cheaper models without sacrificing quality. However, the community does not fully accept the article's core claim until WorkWeave publishes stronger evidence around cache-aware savings, routing accuracy, generalization, and comparisons with simpler workflows.

In Agreement

• Many commenters agree that AI coding costs and token budgets have become painful enough to justify tooling that chooses cheaper models when quality allows.
• Several participants find cache-aware routing more credible than stateless routing and say the product should foreground that distinction.
• Supporters see subagents and fresh context windows as natural places to route simple work to cheaper models without damaging the main agent session.
• Some users welcome automation because they are tired of manually changing model intelligence, speed, and provider choices during coding sessions.
• A few commenters view enterprise environments with large AI adoption and limited token budgets as a strong fit for this kind of router.
• Commenters who already split work across models or use cheaper open models see the router as a way to systematize a workflow they are doing manually.

Opposed

• The strongest objection is that switching models can destroy prompt or KV cache benefits, making routing more expensive than staying with a cached model.
• Several commenters argue that coding harnesses already understand planning, exploration, implementation, and review phases better than an external proxy can.
• Skeptics question whether a router trained on limited traces can generalize across unfamiliar codebases, ambiguous tasks, prompt styles, and rapidly changing model releases.
• Some prefer eval-driven model locking, explicit artifact-based workflows, or manual planner-and-executor splits over opaque real-time routing decisions.
• Commenters ask for public benchmarks, A/B tests, and cost-quality evidence before accepting the claimed savings and quality neutrality.
• Privacy and confidentiality concerns remain for hosted routing because prompts, outputs, and code may pass through another service, even if local deployment is possible.
• A few participants suspect first-party model providers or existing tools such as Cursor, OpenCode, and semantic routers may offer simpler or cheaper alternatives.