Know where your coding time and money go.
An open-source macOS app for people who code with agents. See agent activity, estimated human time and API-equivalent spend from the Claude Code, Codex and Cursor history already on your Mac. Track Claude and Codex account usage and pace in the sidebar.
Download for Mac · Setup guide · Report a bug
See it in 44 seconds
xtrace-launch-44s-logo.mp4
Screenshots and the video show the app with made-up sample projects and account usage, not real work.
You run several agents, try a new tool, or change how you prompt. XTrace helps you see what changed in your time, spend and activity:
- Where did my time go? See agent activity alongside estimated human time, the ratio between them, and how many sessions ran at once.
- What would this cost at API prices? Explore spend by day and session, estimated from published API rates rather than your subscription bill.
- What can I improve? Run scoring skills on a project, read their findings, and add your own skills or ones you discover.
- Will my usage last until the reset? See account limits, whether you are ahead of or behind even use, and a forecast at your current pace.
- What happened during a session? Inspect activity spans, prompts and recorded context compactions on a timeline, then read the whole conversation.
- Where did a session go wrong? See a warning where you corrected or interrupted the agent, repeated a prompt, a call kept failing, or the agent admitted a mistake.
Use the daily charts and previous-period comparisons to investigate changes after adopting a tool or workflow. Compare similar work and consider what you shipped and the review it needed: more agent hours alone do not prove a productivity gain.
The Dashboard's Effort card opens on the human + agent view: one row per day, with shaded bands for when you were working, a line for how many agents ran in each 15 minutes, and that day's hours and leverage. A day at or below 1× shows in red. Hover a date for the day's summary and its sessions, and click a session to open it. Switch to agent hours, human hours or API-equivalent cost for the last 7, 14 or 30 days. Hover the favorite model to see each model's share of output.
The overview shows leverage (agent hours divided by estimated human hours), concurrency and hands-off time, with previous-period comparisons. Definitions explain how each number is counted and what is missing. Values the app could not measure show as a dash, not zero.
Use scoring skills to review your project instructions, comments, repeated code, documentation and agent setup. Open a skill to grade a project, then click a score to see how it was given: the summary, the points for each rubric criterion, and every finding with its explanation and suggested fix. Gradings you start while another one runs wait in line and start by themselves. The screenshot shows the built-in CLAUDE.md Doctor alongside four example added skills.
Sample scores and findings across 14 fictional projects, not measured grading results.
Use Add a grading skill to add your own installed skills or discover skills through a public GitHub link. A skill needs a scoring rubric; Make a rubric can create one for a suitable review skill before you add it. Skills that need to modify a project or access the network may not be suitable. See adding skills and rubrics.
Grading and rubric creation run Claude Code on your Claude account and use your Claude usage. Grading reads project content and keeps the project read-only; these are actions you start, separate from local history indexing. See Your data for what is sent and stored.
Claude and Codex usage stays in the sidebar when your local tools can report it. Green and orange indicators show usage behind or ahead of an even pace. Expand a provider to see remaining usage, daily readings and a linear forecast based on average use since the weekly window began. The forecast assumes that pace continues; it is an estimate, not a guaranteed run-out time. If Claude Code is signed out, its row says so; click it to open Terminal and sign in again.
See active spans and gaps across your sessions, then hover a span for its prompt and activity details. Compaction counts show how often the saved context was summarized. Counts cover the session's saved history; timeline markers show recorded compaction times within the visible range. Hover a marker to see when it happened and its trigger, when the source provides one. Unknown or unsupported readings stay distinct from a confirmed zero.
The Sessions page brings your indexed Claude Code, Codex and Cursor sessions into one list. Find a session by ID, repo or branch, filter by tool, and open it to read the whole conversation, with tool activity you can expand. Codex and Cursor need the bundled readers and Python 3.10 or newer; missing readers and unsupported formats show an unavailable state. Cursor SQLite transcripts are not supported. A sub-session whose parent session can't be confirmed is left out of the list.
XTrace checks each session's whole history for signs it went wrong: you corrected the agent after a reply, interrupted it or declined a tool call, sent the same prompt twice, the same call failed more than once, or the agent's reply admitted a mistake ("you're right", "my mistake"). These are simple rules, not a model's judgment, so some warnings will be false alarms.
Warnings show as red dots on the Dashboard's session rows and human + agent days, and as a count on the Sessions page. Hover a dot to see what was flagged; on a day, it also names the session:
Open a warning to see the conversation around it and why it was flagged. Dismiss hides a warning everywhere, and Undo brings it back. A dismissed repeated prompt or repeated failure comes back if it happens again. Diagnose asks Claude Code (or Codex) to explain what happened and can suggest a skill or an instruction line; it sends messages only after you confirm (see Your data). A session's page lists its warnings and says which checks could not read everything.
Merged pull requests show the effort of their linked sessions: tokens, your
messages, agent time and hands-off time. A session linked to several pull requests
counts in full toward each of them. Pull request details come from GitHub's
command-line tool (gh) when it is installed and signed in. XTrace never posts
to GitHub.
The menu bar keeps today's activity one click away. Back and Forward beside each page's title (⌘[ and ⌘]) return to pages you opened. Settings includes light, dark or system appearance, the human-time break threshold, index status, storage controls and the telemetry switch.
How is agent time counted? XTrace groups recorded events within each session into activity spans. Gaps longer than 20 minutes are excluded. Overlapping sessions add together, so this measures session activity rather than pure model compute time or elapsed time on a clock.
How is human time estimated? XTrace uses messages classified as sent by you, excluding recognized automated messages, across conversations and tools on one timeline. Consecutive messages at most 60 minutes apart count as one working stretch. You can change that break threshold in Settings. A single isolated message adds no time; an unknown message classification makes the estimate unavailable rather than silently lowering it.
Does reviewing diffs count? Reviewing is not detected directly. If you send messages at 2:00 and 2:40 and review in between, those 40 minutes count with the default threshold. Review after your last message is not automatically included. Other activity between those messages can count too, so this is an estimate of your work with agents, not a record of every minute worked.
Does this identify who edited a file? No. Agent time measures recorded agent activity, and human time uses message timing. They can overlap. These metrics do not attribute individual edits or lines to you or an agent.
- Download the
.dmgfrom the latest release. - Drag XTrace into Applications.
- Open it. The first scan reads your existing history and shows its progress.
Requires an Apple Silicon Mac with macOS 14 or newer. Claude Code history is read directly; Codex and Cursor history needs Python 3.10 or newer. Cursor conversations kept only in the Cursor app's own database are not read. The app downloads signed updates and installs them when you click Restart to update. See the setup guide for first launch and troubleshooting.
- Local history indexing. XTrace reads the history files your agents already save without editing them or automatically uploading their transcripts.
- Metrics, not full transcripts. By default the local database keeps metrics, repository paths, IDs and a preview (up to 280 characters) of each message you sent and of each recognized Claude Code background-task summary. Previews can't be turned off. Saved warning decisions can also keep a few matched message words in either storage mode. Full transcripts are saved only if you turn that on. Delete stored content clears saved text and warning decisions, even when no other saved text remains. Later imports, including older messages, can create new previews and warning decisions.
- Diagnose sends selected messages. Only after you confirm Send & diagnose, selected messages around a warning go to Claude Code's hosted model, or Codex's if Claude Code is unavailable before receiving the messages. Cursor is not tried automatically. XTrace keeps the result in the window's cache and does not save it in the database. Save as skill is a separate click that writes the suggested skill to the conversation's project.
- Scoring uses Claude. When you start a grade, Claude Code reads the selected project and skill; that content can be sent to Claude on your account. Making a rubric similarly uses the selected skill's content. Grading keeps the project read-only and blocks network access for sandboxed commands; that does not make Claude's model requests offline. Grading and rubric runs also cannot read well-known secret folders such as SSH keys, cloud credentials, keychains and browser profiles. The run's transcript is not saved by Claude Code; XTrace keeps the resulting scores, summaries, findings and reports. See grading and rubric details.
- Other network use. XTrace uses the network for anonymous usage telemetry (see Telemetry), account usage readings, GitHub pull request checks, app updates, public GitHub look-ups and downloads you start with Add from a link, and the public leaderboard if you join it. Keeping a skill stores it locally and links it into normal Claude Code sessions when its name is free there. Claude usage checks run your local Claude Code about every ten minutes, which adds small entries to Claude Code's own settings and history.
More about local storage and account-usage checks is in data and network.
XTrace sends a few anonymous usage events, such as when the app is opened
and roughly how often it is used, with the app version, the macOS version, which
coding agents have history on this Mac and a random ID. It never sends repository
names, file paths, prompts or other conversation text, pull-request links, token
counts or costs. Telemetry is on by default, and nothing is sent until the app has
shown a notice about it. Turn it off with Turn off in that notice, the switch in
Settings, or by starting the app with XTRACE_TELEMETRY=0. TELEMETRY.md
lists every event and property.
XTrace is a public beta. Some numbers can be incomplete when part of your history is missing. Rulebook is marked as coming soon in the sidebar. On the Leaderboard page you can choose to join the public leaderboard at leaderboard.xtrace.ai. Once you have joined, the page also shows the public list with your row marked; TELEMETRY.md says what joining sends and reads. See the beta testing limits.
You need macOS 14 or newer, Xcode command-line tools, Rust 1.94.1, Node 22.23.1 and pnpm 10.33.0.
pnpm install --frozen-lockfile
pnpm check
cargo test --workspace --locked
pnpm tauri devA development build reads your local history by default. DEVELOPING.md covers sample data, all required checks and native verification, and the architecture map explains how the pieces fit together.
Issues and pull requests are welcome. Read the contributor guide for review and DCO sign-off, and the code of conduct. Security reports go through SECURITY.md.
Source code is licensed under Apache-2.0. The XTrace name and logo are covered by the trademark terms.