router

Product Introduction

1. Definition: The Weave Router is an intelligent model router for AI agentic systems and AI-powered coding assistants. It operates as a software-based routing proxy that sits between a developer's IDE (like Cursor or Claude Code) and various Large Language Model (LLM) providers.
2. Core Value Proposition: It exists to dramatically reduce AI infrastructure costs by automatically and intelligently routing each coding prompt to the most cost-effective model capable of handling the task at full quality, solving the problem of overspending on expensive frontier models for simple requests.

Main Features

1. Intelligent Prompt Classification & Routing: The router analyzes each prompt in real-time (in under 50ms) using a distilled classification model hosted on an edge layer. It determines the complexity of the request (e.g., "explain stack trace" vs. "refactor auth module") and routes it to the smallest, cheapest model in its configured pool that can complete the job successfully.
2. Configurable Routing Policies: Users can set enterprise-grade routing policies on a per-team or per-repo basis. Policies include "Quality-first" (uses frontier models when quality scores separate), "Balanced" (default, uses the cheapest model within a quality parity band), and "Cost-first" (uses the lowest-cost model that clears a set quality bar for every request).
3. Unified Cross-Platform Management: Unlike harness-native routers, Weave Router operates above individual tools. It applies a single, consistent routing policy across all supported AI coding assistants (Claude Code, Codex, Cursor, VS Code, etc.) from a central dashboard, providing unified cost control and observability.
4. Performance-Optimized Integration: The system adds minimal latency (~3ms of added routing overhead) and maintains a 99.9% request completion rate with automatic failover. Integration is designed for developers, requiring just one command (npx -y @workweave/router) to automatically configure environment variables for providers like Anthropic, OpenAI, and Google.

Problems Solved

1. Pain Point: Inefficient and costly AI token spend, where engineering teams default to using expensive, over-powered frontier LLMs (like Claude Opus or GPT-4) for simple, repetitive coding tasks that cheaper models could handle perfectly well.
2. Target Audience: Engineering leaders, DevOps/Platform teams, and developers in organizations using AI coding assistants who are responsible for managing and optimizing cloud/AI infrastructure budgets. Specifically, teams using Claude Code, Codex, Cursor, Zed, or Neovim.
3. Use Cases: Automatically routing simple tasks like summarizing pull requests, drafting release notes, explaining stack traces, or triaging tickets to cost-efficient models (e.g., DeepSeek, Gemini), while reserving frontier models for complex tasks like refactoring critical modules or debugging flaky integration tests.

Unique Advantages

1. Differentiation: Unlike simple API proxies or manual model selection, Weave Router uses real-time quality-per-token scoring to make routing decisions. Unlike other observability tools, it takes active control of routing. It also differs from vendor-locked solutions by being client-agnostic, working across multiple AI coding harnesses with one policy.
2. Key Innovation: Its routing decisions are scored and validated against the actual engineering metrics Weave already tracks for a team—such as code quality score, keep rate (percentage of AI-suggested code that is merged), and cost per merged PR. This creates a closed-loop system where routing optimizes for real business outcomes, not just hypothetical latency or cost.

Frequently Asked Questions (FAQ)

1. How does the Weave Router reduce AI model costs without sacrificing code quality? The router uses a small, fast classifier to analyze each prompt's complexity in real-time. It then routes the task to the least expensive model in its pool that has been proven capable of delivering a high-quality output for that type of task, as measured by code quality scores and keep rates. This ensures simple tasks don't waste expensive tokens, while complex tasks still get the power they need.
2. What is the latency impact of adding an intelligent model router? According to Weave, the classification adds under 50ms, with only about 3ms of additional routing overhead. The first token from the selected model streams back in the same timeframe, often faster, as the router can bypass congested API endpoints.
3. Which AI models and coding assistants does Weave Router support? The router supports a wide pool of models, as indicated by examples like Gemini, Kimi, GPT-4, Claude Sonnet/Opus, DeepSeek, and Llama. It integrates natively with popular AI coding assistants including Cursor, Claude Code, Codex, Zed, Neovim, and VS Code.
4. Is my code and prompt data private when using the Weave Router? Data privacy is a key concern. The router is designed so that the bearer token (your key) never leaves your device unless you explicitly export it. The service likely operates on a proxy model where prompts are routed but policies can be enforced with attention to data security, though specifics should be verified with their security documentation.
5. How difficult is it to integrate the Weave Router into our existing developer workflow? Integration is designed for minimal friction. Developers run a single command (npx @workweave/router init) which auto-detects installed clients and automatically configures the necessary environment variables. No changes to developer habits or code are required.