CostHawk Launches Public Leaderboard for Claude Code, Codex, and Cursor Token Consumption
CostHawk released a public AI Tools Leaderboard that ranks users of Claude Code, OpenAI Codex CLI, and Cursor by total token consumption. The leaderboard updates every 5 minutes and includes per-tool breakdowns by token share, operator mix (single-tool vs. multi-tool), and daily momentum. Monthly seasons reset at midnight Central Time.
Token counts, model identifiers, sync timestamps, and hashed project IDs are tracked, but the service explicitly does not store prompts, responses, or source code. Data is parsed locally from developer-tool data directories before leaving the machine, with AES-256-GCM encryption for API keys. Connection options include Admin API keys, wrapped proxy keys, or MCP telemetry. The auto-sync feature is opt-in (COSTHAWK_AUTO_SYNC=true) and defaults to off.
Connecting a Local Tool
Run the following one-liner in a terminal:
npm exec --yes costhawk@latest -- --loginThis authenticates the machine and installs the CostHawk MCP server for Claude Code, Codex CLI, Cursor, and Gemini CLI. After connecting, tokens are reported under a hashed project identifier.
The leaderboard currently shows no activity (all zeroes) because no operators have synced yet—the site notes “No leaderboard activity yet” and “LOADING DATA SHOWING 0 OF 0 OPERATORS.” Monthly rankings will become populated as users opt in and sync.
📖 Read the full source: HN AI Agents
👀 See Also
IUM: MCP Symbol Indexer Cuts AI Agent Token Usage by 15.9x vs grep
IUM indexes codebases into an SQLite matrix of symbol events, exposing exact file:line coordinates, call graph tracing, and semantic search via MCP. Benchmarked against DataFusion (1,538 files) showing 15.9x fewer tokens than grep for equivalent queries.
Otterly: Route OpenClaw Through Your Claude Code Subscription
Otterly is a small npm package that exposes the local Claude CLI as an OpenAI-compatible HTTP server, letting you bill OpenClaw requests to your Claude Code subscription instead of pay-per-token API rates.
Measuring Off-Task Token Spend in Claude Code: The 'Undeclared-Intent' Metric
A developer built a metric to quantify compute spent on unintended execution paths in Claude Code sessions, finding that 22.8% of tokens went to off-task work.
NexQuant: Rust-native 3-bit KV-cache engine for edge deployment
NexQuant is a production-hardened Rust engine that enables running high-context models on consumer hardware with 3-5x memory reduction. It supports Metal, CUDA, Vulkan, and CPU backends.