A manual-session "Recall" for developers. Start a session, work normally, and every click captures the window you're working in. When you're done, review the steps, let a local vision model describe them, add your own notes, and export everything as a Markdown document for Obsidian or GitHub.
Built with Wails (Go + React + TypeScript + Tailwind). Runs on macOS and Windows; Linux (X11) builds from source.
- Manual sessions. Start recording from the app; the window hides itself so it never captures its own UI.
- Click-triggered capture. Each click grabs only the focused window, throttled to one capture per 1.5 s.
- Global stop shortcut. Stops recording and opens the review screen (
ctrl+shift+eon Windows/Linux,⌘⌥Eon macOS; configurable). - Review before AI. Nothing is sent to a model until you start the analysis. Remove steps, add notes, set instructions for the whole batch, then analyze all or individual steps.
- Local LLM. Works with any OpenAI-compatible vision endpoint (Ollama, LM Studio, vLLM, a DGX server, …). Each request carries exactly one screenshot.
- Markdown export. One folder per session with the
.mddocument and the screenshots, using relative links. - Private by default. Screenshots, the database and settings stay on your machine.
Download the latest build from Releases, or build it yourself (below).
- macOS: grant Accessibility (global click listener) and Screen Recording (window capture) in System Settings → Privacy & Security, then restart the app.
- Windows: no extra setup. Some antivirus tools block global mouse hooks; allow the app if clicks aren't captured.
- Linux: X11 only (Wayland is not supported).
- Open ⚙ Settings and set your endpoint URL, model ID and (optional) API key.
- Click Start Session. The window hides.
- Work as usual. Every click captures the focused window.
- Press the stop shortcut. Recording stops and the review screen opens.
- Remove steps you don't need, add notes, optionally write instructions for all steps, then click Analyze all unanalyzed. A step's own notes are sent along as extra instructions for that step.
- Click Export Markdown and choose a folder.
Settings are stored in settings.json in the data folder:
| OS | Data folder |
|---|---|
| macOS | ~/Library/Application Support/my-recall-app/ |
| Windows | %LOCALAPPDATA%\my-recall-app\ |
| Linux | ~/.local/share/my-recall-app/ (or $XDG_DATA_HOME) |
The same folder holds the screenshots, the SQLite database and app.log. RECALL_LLM_URL, RECALL_LLM_MODEL and
RECALL_LLM_API_KEY set the defaults used before you save settings for the first time.
Requirements: Go (see go.mod), Node.js 22+, and the Wails CLI.
go install github.com/wailsapp/wails/v2/cmd/wails@latest
wails doctor # checks platform dependencies
make dev # hot-reload dev mode (also generates frontend/wailsjs)
make check # go vet (native + Windows) and go test -race
make build # native build into build/bin/
make build-windows # Windows .exe from any OS, no C compiler neededmacOS and Linux builds need CGO and must be built on that OS. Linux additionally needs
libx11-dev libxtst-dev libxkbcommon-dev plus the GTK/WebKit packages listed by wails doctor.
Pushing a tag like v1.0.0 runs the release workflow, which builds macOS (universal) and Windows binaries and attaches them to a GitHub release.
main.go wiring: config, logging, storage, services → Wails
app.go Wails binding layer (thin; every exported method is callable from TS)
emitter.go forwards backend events to the frontend
internal/
recorder/ session lifecycle: click hook, 1.5 s throttle, capture → JPEG → disk → DB
analysis/ vision-model jobs: worker limit, cancel-on-reanalyze, per-step status/errors
llm/ OpenAI-compatible client, exactly one image per request
storage/ SQLite repository (pure-Go driver)
settings/ user settings (endpoint, model, API key, shortcut), atomic JSON persistence
export/ Markdown bundle writer (doc + copied images)
hotkey/ global stop shortcut, configurable, safe rebinding
capture/ focused-window grab: Windows (user32/gdi32), macOS (CoreGraphics), Linux (X11)
clickhook/ global mouse-down hook: Windows pure Go, macOS/Linux gohook
imaging/ downscale to ≤1280 px + JPEG
config/ data dir + static tuning values
events/ backend → frontend event names (mirrored in frontend/src/lib/events.ts)
models/ wire types shared with the UI
frontend/ React + TS + Tailwind dashboard
Services depend on small consumer-side interfaces (recorder.Store, analysis.Analyzer, …), so each is unit-tested with fakes.