What is Tadoru?
An open-source tool that records your on-screen activity locally and turns it into LLM-ready timelines for AI agents.
Tadoru records what you are doing on your computer — from OS-level events, not screen recordings — and lets you retrieve it in a form AI agents can read directly.
tadoru timeline --since 2hThat single command returns the last two hours of your activity as a session-structured, deduplicated Markdown timeline. Coding agents like Claude Code or Codex CLI can read it to understand what you were just working on.
What you can do with it
Resume work
“Pick up where I left off” just works. Your agent reconstructs the context — the PR you were reviewing, the files you were editing, the pages you were reading — from the timeline.
Search your activity
“Which Stripe doc was I reading yesterday?” Query raw events filtered by time range, app, or event type.
Standups and work logs
Feed a day’s timeline to your agent and get a draft of your daily report in seconds.
Context for agents
Two integration surfaces — CLI and MCP — plus shipped skill files cover major agents, including shell-less MCP clients like Claude Desktop.
Design principles
Activity history is sensitive data. Tadoru is designed to be safe by default, not merely safe when configured carefully.
| Principle | Implementation |
|---|---|
| Local-only | There is no egress feature to turn off — the code has no network path for sending your data anywhere. |
| No screen recording | No screenshots, no screen-capture APIs. Only OS accessibility and event information. |
| No keystroke content by default | Only the fact that typing happened and the field type are recorded. Capturing content is an explicit opt-in. |
| Sensitive data excluded automatically | Password fields, Chrome Incognito windows, and password managers are excluded out of the box. |
| Short-lived retention | Events are deleted automatically after 48 hours by default. |
See the privacy model for details.
How it works
Tadoru captures OS events — app switches, window focus, UI interactions, browser URL changes — normalizes them, applies privacy processing, and writes them to a local SQLite store. Everything that reads data back (CLI, MCP) is a thin wrapper over the same store.
Two interfaces
| Interface | Role |
|---|---|
| CLI | The core surface. Recording control, permission diagnostics, timeline retrieval, and filter management — all in a single tadoru binary. |
| MCP server | A read-only view over the same store. The only path for shell-less MCP clients such as Claude Desktop. |
For agents, tadoru setup additionally provides a skill — instructions telling the agent when to call Tadoru. It installs the skill for agents with native support and prints derived instructions for you to place yourself for other CLI agents. A skill is not a third interface, just usage instructions for the two above; MCP tools carry their own descriptions inside the protocol.
Platform support
- macOS, currently. Release binaries are signed and notarized, with no external runtime dependencies. Homebrew distribution is in preparation — for now, install by building from source.
- Browser URL capture currently supports Google Chrome only. See the event reference for the reasoning and per-browser status.
- Windows and Linux are on the roadmap.