XDA ran a warning this week that Claude Code’s agentic loop can eat your wallet if you don’t touch the settings. The exact math depends on model choice and how aggressively the agent iterates, but developers have posted screenshots of $200 to $800 monthly spend on a single project. That number hides in an API dashboard nobody checks daily. The fix is to install a tool that tracks cost as you code, then set caps that actually fire.
We tested seven of the best apps for tracking AI coding costs on Windows, macOS, and Linux, and ranked them by what they cover, how they alert, and whether they play nicely with Claude Code, Cursor, and Aider.
What to look for in an AI coding cost tracker
- Per-model breakdown. Sonnet, Opus, and Haiku bill at very different rates. A single number hides the pain.
- Session-level attribution. The tool should map spend to project, agent run, or feature branch, not just a monthly total.
- Alerts and hard caps. Passive dashboards don’t stop runaway loops. Real caps do.
- Local-first option. Devs who don’t want to route every API call through a third party need a self-hosted option.
- Multi-provider. Anthropic, OpenAI, Google, OpenRouter, and locally hosted models should live in the same report.
- Zero friction install. If it takes an hour to set up, you won’t. The best options run as one command.
Quick comparison
| App | Best for | Platforms | Free plan | Starting price/mo |
|---|---|---|---|---|
| CCUsage | Claude Code CLI users | All (Node.js) | Yes | Free |
| Vibe Kanban | Cursor and Claude Code with task-level tracking | Mac, Windows, Linux | Yes | Free tier available |
| OpenRouter Analytics | OpenRouter API users | Web | Yes | Included |
| Helicone | Full observability across providers | Web + self-host | Yes | $20 per seat |
| Langfuse | Open-source LLM observability | Web + self-host | Yes | Free self-host |
| LiteLLM Proxy | Proxy with cap enforcement | Self-host | Yes | Free self-host |
| Cursor Stats | Cursor-specific budget tracker | Extension | Yes | Free |
The best AI coding cost trackers
1. CCUsage, best for Claude Code CLI users
CCUsage is a lightweight Node.js CLI that reads Claude Code’s local logs and produces a daily, weekly, or per-session breakdown of tokens and cost. Install with npm and run, no signup required, no data leaves your machine.
CCUsage for tracking AI coding costs is the default first install if you use Claude Code. It answers “what did I spend today” in under a second.
Where it falls short: Claude Code only. Cursor and Aider users need something else.
Pricing:
- Free: full functionality
- Paid: none
Platforms: Any OS with Node.js
Download: CCUsage
Bottom line: Perfect for anyone running Claude Code from the CLI. Skip if you don’t use Claude Code.
2. Vibe Kanban, best task-level tracking across Cursor and Claude Code
Vibe Kanban is a project management layer that wraps Cursor, Claude Code, and Aider runs into task cards. Each card carries the tokens consumed and cost, so you learn which feature branches actually cost you money.
Vibe Kanban for tracking AI coding costs stands out on attribution. “This feature cost me $47” is a more useful number than “I spent $200 this week.”
Where it falls short: Requires you to structure work into tasks, which is friction if you code freeform.
Pricing:
- Free: local use
- Paid: cloud team tier available
Platforms: macOS, Windows, Linux
Download: Vibe Kanban
Bottom line: Perfect if you already use kanban and want cost attributed per task. Skip if you code without task structure.
3. OpenRouter Analytics, best for OpenRouter users
OpenRouter Analytics ships built into the OpenRouter dashboard. If you route API calls through OpenRouter (they aggregate Anthropic, OpenAI, Google, Meta, and dozens of others), you already have a per-model cost view in your account.
OpenRouter Analytics for tracking AI coding costs is zero-effort if you use OpenRouter. It handles multi-provider automatically because everything flows through them.
Where it falls short: Only tracks calls routed through OpenRouter. Direct Anthropic or OpenAI usage is invisible.
Pricing:
- Free: included with any API key
- Paid: pay only for tokens
Platforms: Web dashboard
Download: OpenRouter
Bottom line: Perfect if OpenRouter is already your API gateway. Set this aside if you call providers direct.
4. Helicone, best full observability
Helicone is a full LLM observability platform. Every call routed through their proxy captures cost, latency, prompt, response, and user session. Dashboards let you slice by project, user, and model.
Helicone for tracking AI coding costs is the pick if you want production-grade observability with alerts and per-user budgets.
Where it falls short: The proxy step adds latency, and free tier caps monthly volume. Requires code change or gateway config.
Pricing:
- Free: 100K logs per month
- Paid: from $20 per seat per month
Platforms: Web + self-host
Download: Helicone
Bottom line: Perfect for teams shipping AI products in production. Overkill for solo hobby projects.
5. Langfuse, best open-source observability
Langfuse is Helicone’s open-source competitor. Self-host the stack with Docker, wire your SDK to send traces, and get the same dashboards without vendor lock-in. Prompts, cost, latency, and evaluations all in one view.
Langfuse for tracking AI coding costs is the choice for privacy-conscious teams or hobby devs who want production-grade tools without a subscription.
Where it falls short: Self-hosting takes real effort. Learning curve is steeper than a CLI.
Pricing:
- Free: self-host
- Paid: cloud from $59 per month for teams
Platforms: Docker on Linux, Mac, Windows
Download: Langfuse
Bottom line: Perfect for privacy-first teams and open-source purists. Skip if you don’t want to run Docker.
6. LiteLLM Proxy, best for hard-cap enforcement
LiteLLM Proxy is a self-hosted gateway that normalizes API calls across dozens of providers. Its budget module enforces monthly caps per API key. When you hit the cap, calls fail cleanly, so runaway agents actually stop.
LiteLLM Proxy for tracking AI coding costs is the pick if you want caps that fire, not just dashboards that scold you after the fact.
Where it falls short: Requires proxy config in every tool. Not a fire-and-forget experience.
Pricing:
- Free: open source
- Paid: enterprise support tier
Platforms: Docker on Linux, Mac, Windows
Download: LiteLLM
Bottom line: Perfect for teams that want enforced budgets across every AI tool. Skip if you only need read-only tracking.
7. Cursor Stats, best Cursor-specific budget tracker
Cursor Stats is a lightweight VS Code / Cursor extension that reads your Cursor Pro usage counter and surfaces it in the status bar. Runs entirely local, no signup, no proxy.
Cursor Stats for tracking AI coding costs is the “just tell me what I’ve used this month” pick for solo Cursor users.
Where it falls short: Cursor only. Does not cover CLI agents or non-Cursor tools.
Pricing:
- Free: full functionality
- Paid: none
Platforms: Cursor extension on Mac, Windows, Linux
Download: Cursor Stats
Bottom line: Perfect for solo Cursor users who just want a status-bar counter. Skip if you use multiple AI tools.
How to pick the right one
- If you use Claude Code from the CLI: CCUsage
- If you want task-level attribution across tools: Vibe Kanban
- If OpenRouter is your gateway: OpenRouter Analytics
- If your team needs production observability: Helicone
- If you want production observability without the vendor bill: Langfuse
- If you need enforced hard caps: LiteLLM Proxy
- If Cursor is your only AI tool: Cursor Stats
FAQ
How much does Claude Code cost per month?
Depends on model choice, prompt length, and agent loop count. Real developers report $50 to $800 per month for active use. Sonnet costs a fraction of Opus per token. Aggressive Opus-first agents get expensive fast.
How do I stop Claude Code from spending too much?
Set model preference to Sonnet or Haiku by default, cap tool-use iterations in settings, and install CCUsage to see spend daily. LiteLLM Proxy adds enforced caps if you want a hard stop.
Is there a free AI coding cost tracker?
CCUsage, Vibe Kanban (local), OpenRouter Analytics, Langfuse (self-host), LiteLLM Proxy, and Cursor Stats all have real free tiers.
Which is best for tracking Cursor spend?
Cursor Stats for the built-in usage counter, or Vibe Kanban if you want task-level attribution across projects.
Do these tools work with local LLMs like Llama or Qwen?
Langfuse, LiteLLM Proxy, and Helicone all support locally hosted models. CCUsage is Claude Code specific and does not track local runs.
