AI tool: Cursor Router

Introducing Cursor Router · Cursor

Roughly 60% of developers using Cursor pick a single model as their daily driver. This results in routine work being completed at frontier prices, and AI spend growing much faster than output quality. Cursor Router fixes that by classifying each request before a model runs.

At its core, Cursor Router is a classifier that routes users to the best model option based on their query. trained on 600k+ live requests and evaluated performance in an online A/B test across millions of live requests directed by Cursor Router, optimizing for user satisfaction (AFC) as a reward.

Cursor Router is an intelligent model router designed for teams and enterprises using the Cursor editor. It automatically routes coding requests to the most capable and cost-effective AI model based on task complexity, domain, and context.

---

Key Features & How It Works

• Smart Classification: Analyzes query context, task complexity, and domain to select the best model. For example, simple tasks go to price-efficient models, UI tweaks to design-focused models, and complex reasoning to frontier models.

• Cache-Aware Routing: Trained and evaluated to account for cache-miss costs across multi-turn conversations.

• Three Optimization Modes:

  • Intelligence: Matches frontier-model quality for complex tasks.

  • Balance: Delivers strong performance comparable to popular daily-driver models at lower cost.

  • Cost: Optimizes token spend while maintaining solid output quality.

• Admin Controls: Enterprise admins can roll out Cursor Router per team/group, set default modes, and allow or block specific underlying models.

---

Results & Impact

• Cost Savings: Early access testing showed 30%–50% cost reductions compared to routing all requests to Opus 4.8 API rates, with online A/B testing demonstrating up to 60% savings for frontier-quality output.

• Cost Per Commit: Reduced commit generation costs to $4.63 (Balance mode) and $6.76 (Intelligence mode), compared to higher rates for fixed frontier models.

• Availability: Available across Desktop, Web, iOS, CLI, and SDK for Teams and Enterprise plans.

tools to optimize AI usage: rtk, CodeGraph

How I Stopped Running out of Tokens · Daniela Baron

After hitting her organization's monthly spend limit while using Claude Code for daily engineering work, Daniela Baron established a setup to track, optimize, and drastically reduce her token consumption without sacrificing code quality or speed.

---

1. Monitoring Token Usage

• Claude Desktop Usage Meter: Built-in settings window to keep an eye on official account usage limits.

• Claude Code Usage Monitor: A live CLI tool that sits alongside terminal sessions to estimate current burn rates, cost, and estimated time before limits are hit.

---

2. Optimization Tools

• rtk (Rust Token Killer) — Saves Input Tokens: Acts as a proxy for CLI commands (like git log), compressing raw terminal output before it gets sent as input context to the model.

• Caveman — Saves Output Tokens: A Claude Code plugin that strips filler, pleasantries, and hedging from AI responses, enforcing concise fragments, bullet points, and front-loaded file/line references.

• CodeGraph — Saves Both: Maps the codebase into a local SQLite graph database so Claude can query relationships directly instead of grepping and reading large files.

---

3. Key Habits & Settings

• Default to Sonnet with Opus as Advisor: Uses Sonnet for standard tasks and configures the /advisor feature to automatically delegate complex edge cases to Opus only when needed.

• Audit CLAUDE.md: Replaced broad, auto-loaded @ file imports in the project's config with conditional triggers (e.g., "When asked about X, load Y"), sharply reducing ambient context size across the team.

• Clear Context Frequently (/clear): Resets stateless conversation history between tasks so long chat histories aren't repeatedly resent as costly input tokens on every turn.

• Scope Prompts & Check /context: Inspects active context regularly and strictly limits prompts to only the files relevant to the immediate task.

rtk-ai/rtk: CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies @GitHub

colbymchenry/codegraph: Pre-indexed code knowledge graph, auto syncs on code changes, for Claude Code, Codex, Gemini, Cursor, OpenCode, AntiGravity, Kiro, and Hermes Agent — fewer tokens, fewer tool calls, 100% local @GitHub

The fastest complete code graph · surgical context · built for how agents actually work · 100% local
**Kernel powered by Rust**

codegraph — Understand any codebase as a graph

CodeGraph is a local-first code-intelligence tool. It parses your codebase with tree-sitter, stores every symbol, edge, and file in a local SQLite database, and exposes the result as a queryable knowledge graph — over the Model Context Protocol (MCP), a CLI, and a TypeScript library.

A local-first code-intelligence tool that turns any codebase into a queryable knowledge graph for AI coding agents.

Introduction - Tree-sitter

Tree-sitter is a parser generator tool and an incremental parsing library. It can build a concrete syntax tree for a source file and efficiently update the syntax tree as the source file is edited.