Time Machine has one job. When a coding agent breaks your project, help you get back to the last version that worked, with proof that it works.
Coding agents now run shell commands, delete folders and touch databases. The tools around them were built to undo an edit, not to bring back a project that works.
Its docs say: "A file that only Bash commands changed isn't tracked and stays as it is." A file that only a command like rm or mv changed can't be brought back through rewind.
Sources: Claude Code docs, eon.io
A user reported that an agent permanently deleted an entire production resource folder, and asked for help getting it back.
Source: Cursor community forum
Asked to reorganize a folder, the agent ran commands that lost the user's files, then reported what it had done.
Source: AI Incident Database
During a code freeze, the agent deleted a live database, then called it a catastrophic failure.
Source: Fortune
of developers trust AI with important decisions.
trust it when the output is easy to check.
Figures as reported by InfoWorld. About 7 percent and 48 percent.
Trust follows proof. So Time Machine does not ask you to trust the agent, or us. It runs your check and shows you the result.
Everything below marked Built works today and is covered by tests. Planned means not built.
Time Machinerm -rf, git reset --hard, git clean -f and DROP TABLE. Check at the end of each turn.Built~/.timemachine, with a Git bundle, so save points survive the project folder being deleted.Builttimetm, command tm. Build from source until then.Not published yettm ui opens the app in your browser, served from your own computer. History, save point detail, Rewind and Compare, on the same engine. Free for up to 3 projects at a time; $50 once for more.Builttm command line tool is open source (MIT) and stays that way. Restores are never paywalled.Checkout not open yetThat is the whole idea of Time Machine, so the codebase lives by it. Before a fix, there is a test that shows the bug. The fix is done when that test passes and nothing else broke.
Two independent reviews of the engine found 17 bugs. Every one was fixed test first. The reviews also changed the design: verdicts are now bound to the exact save point they were made for, a store inside the project is refused, and restores never write through a symbolic link.
Each stage starts when the one before it has met its gate. A gate is an outcome you can check, not a day on a calendar.
Gate 1 · Engine and CLI
Done when: every acceptance test passes on macOS and Linux CI, the demo runs end to end, and your folder and current branch are byte for byte unchanged after every command.
Status: met. 404 automated tests pass in CI on macOS and Linux. Publishing timetm to npm is the last step.
Gate 2 · Web app
Done when: someone who does not use a terminal can go back to the last working version with one button, on the same engine as the CLI, in the web app.
Status: built. tm ui opens the web app from a source install and runs on your computer. Still to do: publish timetm to npm, and open checkout for the $50 key.
Gate 3 · Action connector
Starts only when: timetm is on npm and a stranger can reach a verified recovery branch, the design is signed off in writing, and the legal and data questions are settled.
What it would do: record what agents do in apps like Gmail, Slack, Notion and Calendar through Claude and ChatGPT, and reverse only what each app can really reverse. Every result would say what was not reversed.
The CLI works. Turning it into something people can trust with their projects takes four things that time alone cannot buy.
We write to the list when the CLI is on npm. Nothing else.