Transcripted

Transcripted is a local Mac app for dictation and meeting capture that turns spoken work into structured files your agent can actually use.

Use it today as a practical dictation and meeting tool. Longer term, we think audio is the first useful layer of local AI context, because it captures real work without asking you to maintain a second brain by hand.

What You Get Today

• Dictate into any app and paste text back where you were already working
• Record meetings locally with mic and system audio capture
• Save human-readable Markdown and machine-readable JSON artifacts on disk
• Point Claude, Codex, or another local agent at those artifacts with a starter prompt

Connect Your Agent

Transcripted now has one main connection path plus two optional fallbacks:

• Copy one prompt for the default setup flow
• MCP for supported agents that can use direct tools
• Folders for any file-reading agent that should work from local artifacts directly

The end-user setup guide lives in docs/agent-connect.md.

Why This Exists

A lot of important context never makes it into docs or chat logs.

It shows up in:

• meetings
• dictated messages
• voice notes
• half-formed spoken thinking

Most of that context disappears as soon as the conversation ends.

Transcripted tries to preserve that high-signal spoken work locally, then turn it into files that stay inspectable, reusable, and easy to load on demand.

How It Works

1. Capture spoken work

Transcripted supports two concrete workflows today:

• dictation into any app
• local meeting recording and transcription

2. Process it locally

Transcripted transcribes audio on-device and keeps the resulting artifacts on your Mac.

3. Save durable artifacts

Meeting recordings become:

• a Markdown transcript
• a structured JSON sidecar
• a transcripted.json index of saved meetings
• AGENT.md and CLAUDE.md helper docs for external agents

Dictation becomes:

• daily Markdown logs like Dictations_2026-04-07.md
• timestamped sections with source-app metadata and delivery status

4. Load the right slice later

Instead of asking an agent to read everything all the time, Transcripted gives you durable files that can be loaded when needed:

• latest meeting
• meetings for a named speaker
• dictations from a specific day
• transcripts related to a topic or decision

Artifacts, Not A Black Box

Transcripted is opinionated about file output because inspectable artifacts are more useful than opaque app state.

Example meeting transcript:

# Meeting with Alex

## Full Transcript

**[00:00] [Mic/You]**
Thanks for making time today.

**[00:04] [System/Alex]**
Happy to help. Let's get started.

Example meeting sidecar:

{
  "version": "1.0",
  "recording": {
    "duration_seconds": 750,
    "engines": {
      "stt": "parakeet-tdt-v3",
      "diarization": "pyannote-offline"
    }
  },
  "speakers": [
    { "id": "mic_0", "name": "You" },
    { "id": "system_0", "name": "Alex" }
  ],
  "utterances": [
    { "start": 0.0, "end": 4.0, "speaker_id": "mic_0", "text": "Thanks for making time today." }
  ]
}

Example dictation artifact:

---
title: "Dictations for April 7, 2026"
date: 2026-04-07
capture_type: dictation_day
---

## 9:15 AM - First note from the morning

Source app: Messages
Timestamp: 2026-04-07 09:15:00

first note from the morning

Why Audio First

We do not think the answer is "capture everything."

We think the better path is:

• start with a signal people already produce during real work
• structure it into useful artifacts automatically
• let agents load more context only when they need it

Audio is a good first wedge because it is:

• high signal
• already part of meetings and messaging
• less invasive than full screen capture
• easier to structure than a full visual memory system

That does not make Transcripted an "ambient context layer" yet. It means audio is the first practical layer of one.

Local By Default

Transcripted keeps its core workflows on-device:

• dictation capture and saved dictation logs stay on your Mac
• meeting capture, transcription, and saved transcripts stay on your Mac
• agent-facing artifacts are plain local files you can inspect directly

Fresh installs use:

• ~/Library/Application Support/Transcripted/dictations/
• ~/Library/Application Support/Transcripted/meetings/

If a legacy Draft Application Support folder already exists, current builds continue using that location for compatibility while the rename settles.

Operational caveats:

• first launch may download local models from HuggingFace if they are not cached
• beta builds can optionally contact the update/log proxy for update checks and diagnostics

What Exists Today

Today, Transcripted is already useful as:

• a local dictation tool
• a local meeting recorder and transcription tool
• a file-based handoff point for external agents

The current product is not a full passive memory system. Capture is still explicitly user-invoked. The value is that what you capture becomes durable, structured, and reusable instead of disappearing.

Where This Goes Next

The broader direction is to improve the audio context layer before expanding scope.

That likely means:

• better summarization and extraction from saved artifacts
• more selective retrieval for agents
• stronger cross-meeting speaker and topic navigation
• light non-audio context later, where it clearly improves usefulness

The goal is not to make context gathering feel like a job. The goal is to make useful context accumulate quietly, then make the right slice easy to load.

Build

bash build-deps.sh
bash build.sh

Tests

bash run-tests.sh

If you touch meeting integration or TranscriptedCore, also run:

bash run-integration-smoke.sh

Transition Notes

The old standalone Transcripted app is preserved on:

• branch: legacy/transcripted-standalone
• tag: pre-draft-takeover-2026-04-06

This repo currently uses the manual migration path:

• existing standalone Transcripted installs do not auto-upgrade into this app
• fresh installs use Transcripted-named Application Support paths
• existing Draft-named Application Support folders are still reused for compatibility
• permissions and settings do not carry over automatically

Contributing And Security

• See CONTRIBUTING.md for development setup and architecture notes
• See SECURITY.md for privacy architecture and vulnerability reporting