How do you track time spent on AI tools?
Use one of three methods: a manual log, the usage pages AI vendors give you, or an automatic tracker that records sessions as they happen. Manual logs drift within days, and vendor pages show tokens and limits, not hours. Only an automatic tracker gives you one weekly number across tools.
Last reviewed: November 2026. This post covers all three methods, with a worked weekly log you can copy.
| Situation | Best method |
|---|---|
| You use AI only on a phone | Manual log (TimeUno has no mobile app) |
| You only need spend or limits | Vendor usage pages |
| You use several AI tools on a computer | Automatic tracker |
Why measure AI time at all?
"I use AI a lot" is not a number, and memory is a poor instrument. In a 2025 randomized trial by METR, 16 experienced developers expected AI tools to speed them up by 24%. Afterwards they still believed it had sped them up by 20%, yet they took 19% longer on the tasks where AI was allowed (METR).
METR labels those results out of date for current tools, so the 19% is not a verdict on AI. It does show that people misjudge what AI does to their own time.
"AI time" also hides two different numbers:
- Your time with the tool: minutes the app or tab is in front of you, typing, reading, reviewing.
- AI working time: minutes the tool runs while you do something else, such as an agent in a background terminal.
The first is your attention. The second is capacity you did not spend. Neither equals the wall clock.
Is tracking time on AI tools the same as logging time with AI?
No. Tracking time on AI tools measures how long you and your agents spend in ChatGPT, Claude Code, Cursor and similar tools. Logging time with AI means telling an assistant to create entries in a time tracker, usually through MCP, the Model Context Protocol. This post is about the first. For the second, TimeUno runs an MCP server at https://app.timeuno.io/mcp, documented at timeuno.io/mcp.
Can you track AI time manually?
You can, with a spreadsheet or a timer, but it breaks down in about a week. AI use is many short bursts: a 90-second question, a 4-minute code fix, a long agent run. Nobody starts a timer for each one, so the log undercounts the small sessions that add up the most. Background agent runs have no natural start or stop, and browser tabs stay open for hours without being in use.
If you start by hand, keep one row per session (date, tool, minutes, task) for three days. The total is a lower bound, and the gap between it and an automatic count shows how much you forget.
Developers have a free half-way option for one tool. Claude Code stores conversation history with timestamps on disk, and open-source scripts such as cctimesheet group those messages into time blocks (write-up by its author). It covers Claude Code only, on one machine, and only as far back as the history goes.
What do AI vendors' usage pages show?
They show what the vendor bills or limits: tokens, requests, spend and plan limits. They do not show how long you sat in the tool, and they cover only that vendor.
Claude Code's /usage command shows token usage for the current session, an estimated cost and, for subscribers, plan usage bars and a breakdown by skill, subagent and plugin. Anthropic's documentation says the figures are approximate and computed from local session history on that machine, so usage from other devices or claude.ai is not included (Claude Code cost docs).
As of November 2026:
| Source | What it shows | What it leaves out |
|---|---|---|
Claude Code /usage | Session tokens, estimated cost, plan limit bars, 24-hour and 7-day breakdown | Other tools, other devices, claude.ai, your time away from the session |
| claude.ai settings, Usage | Current session usage, weekly reset time and usage credits (Claude help) | Hours of use; other vendors |
| Cursor dashboard | Included usage and on-demand spend (Cursor help) | Hours in the editor, other tools |
| ChatGPT | A data export of your chat history and account data (OpenAI help); not offered in Business, Enterprise or Healthcare workspaces | The page describes no time-spent report |
These pages answer "what did this cost me" and "how close am I to my limit". They cannot answer "how many hours did I spend in AI across all my tools".
How does automatic AI time tracking work?
An automatic tracker watches which app or tab is in front of you, reads session activity from supported AI tools, and adds it up. You never start or stop anything. You get one timeline of your attention and one of your AI tools, side by side.
TimeUno is an automatic time tracker for macOS and Windows that records the app in front of you and session metadata from AI tools. A browser extension (Chrome, Firefox, Edge) covers web tools, and a headless agent covers remote servers. There is no mobile app and no free tier. Every plan starts with a 14-day trial that starts with a card, charged only when the 14 days end, and paid plans start at $9.99 a month as of November 2026 (see pricing).
Which AI tools does TimeUno track?
| Where | Tools |
|---|---|
| Desktop app | Claude Code, Cursor, Codex CLI, OpenCode, GitHub Copilot, Claude desktop app, Windsurf, Aider, Ollama, Gemini CLI, oh-my-pi, Pi |
| Browser extension | claude.ai, ChatGPT, Gemini, Perplexity, DeepSeek, v0, Grok, Bolt, Lovable, Replit |
| Remote server (headless agent) | Claude Code, Cursor CLI, Codex CLI, Gemini CLI, OpenCode, Aider, Ollama, oh-my-pi, Pi |
Details are on the download page and in the FAQ.
What does it record from AI tools?
From the desktop AI-tool integrations it sends metadata only: when a session starts and ends, whether the AI is working or waiting for you, the model, token counts and cost, and the project folder. Those integrations do not read your prompts, the AI's answers or your code. For Aider, Windsurf and Ollama, TimeUno sees only whether they are active, so there is no token count or cost.
To describe sessions on a timesheet, the desktop app also sends the session title, a task type from a fixed list, relative paths of edited files and subjects of commits the AI made. Switch that off under Settings, Your data, Share task details; the time is still tracked.
Separately, the desktop app and the browser extension read on-screen or page text (up to 4,000 characters per activity record) by default, which can include a prompt or an AI answer if a chat is in front. Turn that off under Settings, Read on-screen text, or have a workspace owner or admin turn it off for everyone.
Can it count AI sessions from before you installed it?
Partly. When you start tracking, the app can import earlier sessions from the history your tools keep on disk:
- Full detail (times, model, tokens, cost, project): Claude Code, oh-my-pi, Pi, OpenCode, Codex CLI.
- Start and end times only: Cursor, Gemini CLI, GitHub Copilot.
- Not imported: Windsurf, Ollama, Aider and the Claude desktop app.
It finds only what your tools still have. Claude Code deletes its history after about 30 days by default. Parallel sessions each count, so imported hours can exceed elapsed time.
What does a weekly AI time log look like?
One row per tool, with sessions, your time, AI working time and one output you care about. The numbers below are made up to show the shape. They are not measured from a real user or a benchmark.
| Tool | Sessions | Your time (h) | AI working (h) | Output you add by hand |
|---|---|---|---|---|
| Claude Code | 11 | 9.5 | 6.0 | 6 pull requests merged |
| Cursor | 8 | 7.0 | 2.5 | 3 pull requests merged |
| claude.ai | 14 | 3.5 | 1.0 | 4 drafts sent |
| ChatGPT | 9 | 2.0 | 0.5 | 2 drafts sent |
| Perplexity | 6 | 1.0 | 0.2 | 3 research notes |
| Total | 48 | 23.0 | 10.2 |
Three numbers come out of it. If the computer was tracked for 41.5 hours in total that week, your AI share is 23.0 of 41.5 hours, about 55%. Hours per output is 9.5 hours for 6 pull requests in Claude Code, about 1.6 each, against 7.0 hours for 3 in Cursor, about 2.3. The working ratio is 10.2 AI hours beside 23.0 of yours, about 0.44.
TimeUno fills the first four columns. The output column is yours, and takes a few minutes each Friday. To see what a tool changes, track a week or two before you adopt it, then compare.
Should you bill clients for the time AI tools spend working?
If you bill by the hour, bill the hours you worked and treat AI working time as your own capacity. That is a rule of thumb, not a legal rule. Your contract and any professional guidance for your field decide what counts as billable. Your time with the tool is the figure you can stand behind with a client. AI working time explains why a task finished faster than the hours suggest.
How do you set it up?
Install the desktop app from the download page, click Start tracking and say yes to "Import your past AI sessions?". Add the browser extension if you use ChatGPT, claude.ai, Gemini or Perplexity in a tab. Work normally for a week without timers, then copy the AI totals from the dashboard into a table like the one above. If a stretch lands in the wrong category, correct it once and TimeUno remembers the correction.
Who should not use this approach?
Skip an automatic tracker if you only use AI on a phone, since TimeUno has no mobile app, or if you want a free tool, since a spreadsheet costs nothing. If you only need a spend figure, vendor usage pages are enough. Choose an automatic tracker if you use several AI tools on a computer and want one number across them. The full write-up is on the original post.
Questions
How do I track how much time I spend on ChatGPT and Claude?
Use an automatic tracker, because a manual log undercounts short sessions. TimeUno tracks ChatGPT and claude.ai through its browser extension and Claude Code through the desktop app, and adds them into one weekly total.
Why can't I use vendor usage pages to measure AI hours?
They show tokens, spend and plan limits, not hours, and each covers only one vendor. They cannot give you one number across all your AI tools.
Does TimeUno record my prompts and AI answers?
Not through its AI-tool integrations, which send only metadata from desktop AI tools. The desktop app and the browser extension do read on-screen or page text by default, and that text can include a prompt or an AI answer if a chat is in front. You can turn it off under Settings, Read on-screen text.
Can I ask my AI assistant to log my hours instead?
Yes, through an MCP connection, but that records what you tell it and does not measure the time. TimeUno's MCP server is described at timeuno.io/mcp.
Does TimeUno have a free plan or a mobile app?
No to both. It has a 14-day trial without a card, desktop apps for macOS and Windows, browser extensions and a headless Linux agent. Questions about plans are answered in the FAQ.
Sources
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025, checked November 2026.
- Anthropic, Manage costs effectively (Claude Code docs), checked November 2026.
- Anthropic, How do usage and length limits work (Claude help center), checked November 2026.
- Cursor, Overages and checking usage, checked November 2026.
- OpenAI, Exporting your ChatGPT history and data, checked November 2026.
- Paul Denya, Claude Code Timesheets, November 2025, and cctimesheet on GitHub.
I work on TimeUno, the tool this post discusses. The full version is at https://timeuno.io/blog/how-to-track-time-spent-on-ai-tools.
