Sifu records what you do on your Mac, then hands the recording to your own AI — which turns it into a skill it can run for you. Local. Open source.
Paste the prompt below into Claude Code or Codex and your agent handles the rest — no manual steps required.
Install Sifu locally so I can train my replacement.
---
## Platform
macOS only. Sifu uses CGEventTap and the macOS Accessibility API to observe
input events. It will not run on Linux or Windows.
---
## Steps
### 1 · Clone or update the repository
If `~/sifu` does not exist:
```
git clone https://github.com/heymitch/sifu.git ~/sifu
```
If `~/sifu` already exists:
```
cd ~/sifu && git pull --ff-only
```
### 2 · Install the package
```
pip install -e ~/sifu
```
This installs Sifu in editable mode. The `sifu` command will be available in
the current Python environment.
### 3 · Run bootstrap
```
python3 -c "from sifu.install.bootstrap import run; print(run())"
```
Bootstrap creates the local library directory and prints two keys:
- `library` — the path where Sifu stores workflow data
- `next` — the next command to run
Surface the `next` value to the user verbatim.
### 4 · Grant Accessibility permission
Sifu cannot capture input events without explicit user approval.
Go to: System Settings → Privacy & Security → Accessibility
Add SifuBar to the list and enable the toggle. Sifu will not prompt for this
automatically — it must be granted before capture begins.
### 5 · Report back
Tell the user:
- The `library` path printed by bootstrap
- The `next` step printed by bootstrap
- That Accessibility permission must be granted before starting capture
Do not launch capture automatically. The user decides when Sifu begins
observing. That decision belongs to them.
---
## Notes
- `pip install -e ~/sifu` is idempotent. Safe to re-run.
- Bootstrap is also idempotent. Running it again will not overwrite existing
library data.
- If `pip` resolves to Python 2, use `pip3` instead.
- The `sifu start` command begins capture. The `sifu context <query>` command
takes a free-text query, finds the best-matching library workflow by substring
score, and prints the workflow markdown plus an instruction line for your agent
to drive the macro via NavMacro/Battleship. Neither runs automatically during
install.
Sifu sits beside you and records the small repeated motions of your work — clicks, keystrokes, the apps you move between, the pages you visit. Nothing leaves your machine.
When a task is worth keeping, copy it to your own AI in one click — the full recording, the screenshots, and Sifu's notes on how to rebuild it as a skill.
Your agent reads the recording and builds an installable skill — driving real APIs and tools, not brittle clicking. You graduate it to autonomy as you come to trust it.
Most work is a pattern you repeat without naming it. Sifu records it, and your agent turns it into a skill — yours to keep, edit, or hand off.