Agent Flow

Speech to text, with your highlights and screenshots

Highlight text or take a screenshot while you talk. Agent Flow pastes your words with each reference near what you were saying, so your agent can see what you were looking at.

Download for Mac

v2.0.366 · macOS 14.4+ · Apple silicon and Intel
Developer ID signed and Apple notarized.

Build from source

Rename this function Selected Text: TextEdit — “func loadRows(from cache: Cache)” and make the empty state look like this. Screenshot: Screenshot 2026-09-24 at 10.41.12.png

Call it fetchRows.

Pasted into Codex

Rename this function

<app_selection index="1" source="TextEdit" characters="32" middle_omitted="false" truncated="false" bundle_id="com.apple.TextEdit">
  <text>func loadRows(from cache: Cache)</text>
</app_selection>

and make the empty state look like this.

<local_screenshot path="/Users/you/Desktop/Screenshot 2026-09-24 at 10.41.12.png"/>

Call it fetchRows.

  • Live words stay in the recorder until you stop, then paste once.
  • Mouse highlights keep up to 8,000 characters with whitespace intact. Separate highlights stay in order, even without speech between them.
  • Click the recorder to type alongside your speech and references. Enter adds a line break; finish normally to paste. The microphone stays active.
  • Context is pasted only into the native Codex, ChatGPT, Claude and Telegram apps. Other apps receive your spoken and typed words without the tags.
  • Saved screenshots add a file path, not an upload. Your agent needs access to that file.
  • A Mode's Colored context option colours your opening words in a rainbow and each XML block by type: speech and typing silver, selections cyan, screenshots magenta with a local link. It's off by default, and messages too long to colour stay plain.
  • Next release: Codex gets a short preview linking to the full prompt, as shown in the video.
  • Placement is approximate: live recognition can revise earlier words.
  • Selection capture depends on the app. Codex can add its task label; Chrome can add page details.
  • Highlights are read without pressing Copy. Password fields are skipped.

For Codex, install the Agent Flow context skill so it reads highlights and screenshot paths as context, not instructions.

Setup

No Agent Flow account or subscription. Bring your own OpenAI API key for live words.

  1. Download, unzip, and move Agent Flow to Applications.
  2. Allow Microphone and Accessibility when asked.
  3. Add an OpenAI API key with API billing for GPT Live transcription. Your ChatGPT subscription doesn't include API credit.
  4. Finish setup, which also downloads a local model for transcription without a key.

Auto-send presses Return after pasting. It starts off; enable it per app in Modes if you want it. For Codex CLI or Claude Code, configure the terminal or editor it runs in.

Updates aren't automatic yet. New versions are on GitHub releases.

Optional extras
Codex skill
Ask Codex: “Install interpret-agent-flow-context from EthanSK/AgentFlow/.agents/skills/interpret-agent-flow-context.”
YouTube Bridge
Pauses the playing YouTube video in Chrome when recording starts, then resumes only that video. Its helper and login LaunchAgent need a source checkout; you load the Chrome extension once.
Chrome details
For page titles and addresses with highlights, enable Chrome's View → Developer → Allow JavaScript from Apple Events, then grant Agent Flow Automation access if asked.
Mouse controls
Map a mouse button to your recording shortcut if you want. Agentic Mouse is separate and isn't required.
VS Code
The Better Git bridge reads highlights in local code editors. Terminals and chat panels aren't covered.

Setup guide and app support

Controls

Use your keyboard shortcut or a mapped mouse button.

  • Press once to start, then again to stop and paste.
  • Press both Command and both Option keys together to type with the microphone off, then again to paste.
  • Press twice while recording to finish without pasting. The text stays on your clipboard and in History.
  • Press three times while recording to pause, then once to resume.
  • Start another recording while the last one is still transcribing.
Next Track and saved inputs

While the recorder is visible, Next Track won't skip music. During recording, it stops and sends the text to the input where you started. After a normal stop, while transcription is still running, it can instead choose the input focused now.

These saved-input actions use that app's settings and show an error if the target can't be verified. A normal stop simply pastes into the input focused when the result arrives.

Destination guide

Build from source

Needs macOS 14.4 or later, Xcode and Git.

git clone https://github.com/EthanSK/AgentFlow.git
cd AgentFlow
./scripts/install-first-use.sh
  • Builds Agent Flow, installs it in ~/Applications and opens it. It won't replace an existing installation.
  • Use --all to also install the Codex context skill and YouTube helper. Chrome still needs its unpacked extension loaded manually.
  • The upstream voiceink Homebrew cask installs VoiceInk, not Agent Flow.

Full setup guide Build details GitHub

Ethan's setup shows his hardware, mouse mappings and companion apps.