talkthru

talkthru.dev · npm · @dr_edon · MIT

An automated feedback loop for any app. Screen-record your app on your phone and talk while you use it. Send it to your machine and your coding agent gets everything you said, attached to the screen you were looking at when you said it.

you:    *tapping through checkout* "this button is too small, I keep missing it"
                                  "and this error doesn't say what went wrong"

agent:  sees both screens, finds both components, fixes them.

Runs entirely on your machine. No accounts, no uploads, no API keys.

Works with a mac screen recording (⌘⇧5), an iPhone one over AirDrop, or Windows Game Bar. Any video with sound works — point the watcher at a folder.

70-second demo — install, record, and the agent picking it up.

Install

npx talkthru doctor --fix

Pulls ffmpeg, whisper.cpp and the speech model. Nothing else to set up.

Start the watcher

npx talkthru watch

Leave it running. It only touches files named like a screen recording — the rest of your Downloads folder is invisible to it.

Connect your agent

talkthru is an MCP server, so any MCP client works — Cursor, Windsurf, Cline, Zed, Claude Desktop. Add it to your client's config:

{
  "mcpServers": {
    "talkthru": {
      "command": "npx",
      "args": ["talkthru-mcp"]
    }
  }
}

Claude Code has a one-liner for the same thing:

claude mcp add --scope user talkthru -- npx talkthru-mcp

Restart your client afterwards so it picks up the server.

Built and tested against Claude Code. Other MCP clients should work — tell me if yours doesn't.

Set up your mac

Once. Press ⌘⇧5, pick a record mode (the icons with the ◉ dot), then open Options and set Save to → Downloads and your Microphone. Both stick.

No mic selected means a silent video and nothing to transcribe. Worth checking before your first recording.

Set up your phone

Optional — only if you want to record a phone app rather than your desktop.

Once. Control Centre → + → Add a Control → Screen Recording. Then press and hold it and turn the microphone on.

Use it

1. Record your app, talk as you go
2. Get the video onto your machine (AirDrop is easiest)
3. Ask your agent: "check the feedback in my last talkthru session and implement it"

Ready in about twenty seconds. Your original video is kept in ~/.talkthru/archive/.

Mute your app if it talks. Its narration lands in the transcript next to yours and whisper can't tell you apart.

Output

Your words, attached to the screen that was up when you said them:

## 00:34 · f11 — `frames/f11.jpg`
- [00:39] "this button is too small, I keep missing it"

## 00:52 · f13 — `frames/f13.jpg`
- [00:55] "and this error doesn't say what actually went wrong"

Plus the frames. A two-minute session is about 600 tokens.

Element context (optional)

Point talkthru at a UI hierarchy file and your agent gets element types, labels and stable test ids instead of guessing which component you meant from pixels:

talkthru process recording.mp4 --hierarchy hierarchy.json

## 00:34 · f11 — `frames/f11.jpg`
ui: button#checkout-submit "Place order" @24,680 342x48
- [00:39] "this button is too small, I keep missing it"

It is a timestamped sidecar your app writes — nothing here generates it, and the format is plain JSON with no iOS-specific assumption:

{ "snapshots": [
  { "tMs": 1200, "nodes": [
    { "type": "button", "testId": "checkout-submit", "label": "Place order",
      "rect": [24, 680, 342, 48], "depth": 4 }
  ] }
] }

Common alternate spellings (tag, id, text, bounds, elements, …) are accepted as-is, so a DOM client can usually emit its natural shape. Malformed files are ignored with a warning rather than failing the recording.

Full format: docs/hierarchy.md — fields, aliases, ranking rules, and how to write a client.

Network and console (optional)

Drop an events.json next to your recording and failed requests land under the frame you were looking at:

## 00:34 · f11 — `frames/f11.jpg`
ui: button#checkout-submit "Place order" @24,680 342x48
net: POST https://api.example.com/checkout 500 340ms
- [00:39] "I tapped place order and nothing happened"

{ "events": [
  { "tMs": 12400, "kind": "network", "method": "POST",
    "url": "https://api.example.com/checkout", "status": 500, "durationMs": 340 },
  { "tMs": 12450, "kind": "console", "level": "error", "text": "TypeError: ..." }
] }

No headers, no bodies — auth tokens live in headers and the point of this tool is that your data stays on your machine. Secret-looking query values are replaced with REDACTED at parse time. Everything you send is stored; only failures and error/warn reach the agent by default, because two minutes of XHR is more text than the frames and narration combined.

Full format: docs/events.md

Windows

Win+Alt+R (Game Bar) records to Videos\Captures, which is where the watcher looks by default — no name filter there, because that folder holds nothing else:

winget install -e --id Gyan.FFmpeg
npx talkthru watch

whisper.cpp has no winget package. Build it from ggml-org/whisper.cpp and point TALKTHRU_WHISPER at the binary.

Written and unit-tested for Windows, but not yet run on a Windows machine — I do not have one. If you try it, an issue either way is useful.

Commands

talkthru watch                  # watch ~/Downloads for screen recordings
talkthru watch ~/some-folder    # watch your own folder, any video
talkthru process video.mp4      # process one file
talkthru list                   # recent sessions
talkthru show latest            # print one
talkthru doctor                 # check the setup
talkthru compact                # reclaim disk: drops old originals, keeps every frame + transcript
talkthru prune                  # delete old sessions entirely

tt also works.

How it works

ffmpeg pulls the frames where your screen actually changed. whisper.cpp transcribes your voice locally, split on the real silence between sentences, so each thing you said lands on the screen you were looking at. An MCP server hands it to your agent.

Needs Node 20.11+.

Why

I built this because I test my own app on my phone every day and kept losing the feedback on the way back to my editor — screenshotting, cropping, trying to describe which screen I meant. Now I just record, talk, and Claude Code picks it up. Sharing it in case it saves you the same trip.