FluidVoice: The Free Local Dictation App for Agent Workflows

Dictation looks like a small productivity tool until you spend all day working with Claude Code, Codex, Cursor, Hermes, ChatGPT, or any other agentic workspace. Then it becomes an input layer.

Andrew Warner and Adam Brakhane's FluidVoice walkthrough is useful because it is not really about saving $10 or $15 per month. It is about owning a fast, local, customizable way to turn messy speech into usable instructions, messages, notes, and prompts.

Source Note

This post is based on The Next New Thing's FluidVoice demo, the attached transcript, FluidVoice's official site, the GitHub repository, GitHub releases, and public docs from Wispr Flow, Superwhisper, and Apple Dictation. Product details can change quickly, especially for open-source Mac utilities, so check the repo and releases before installing.

I am not claiming FluidVoice is universally better than every paid dictation app. The stronger claim is narrower: for Mac users who want a free, local-first, open-source dictation layer with customization, FluidVoice is worth testing.

Link Map

Why Dictation Matters

The AI world keeps obsessing over model quality, but daily speed often comes from the layer right before the model: how fast you can explain what you want.

Typing is precise, but it makes people compress their thoughts. Speaking is messier, but it carries more context: why the task matters, what you tried already, what not to do, what tone you want, which edge cases you remember, and how the result should be used.

That is why dictation is so useful for AI agents. A better voice layer turns "fix this" into a real brief:

• what broke,
• where the relevant files probably are,
• what the user journey should feel like,
• what output format you need,
• what to verify before claiming done.

For agent work, better input becomes better output. FluidVoice is interesting because it tries to make that input layer fast, visible, local, and customizable.

What FluidVoice Is

FluidVoice is a free, open-source macOS dictation app. The GitHub README describes it as a voice-to-text dictation app for macOS with on-device AI enhancement and lists installation through Homebrew:

brew install --cask fluidvoice

The product page frames FluidVoice as two layers:

• FluidVoice: the Mac app with hotkeys, overlay, local speech models, system-wide typing, modes, history, settings, and source code.
• Fluid-1 / Fluid Intelligence: an optional local enhancement layer that can clean up rough dictation, format structure, fix casing, adapt tone by app, and handle names/numbers/dates.

The important distinction is that FluidVoice is not only a transcription app. It is closer to a system-wide voice input layer. You can dictate into Claude Code, Cursor, Slack, Notion, a browser, an email, or a terminal prompt. The transcript highlights one feature Andrew liked immediately: seeing the words appear in real time before they paste into the target app.

FluidVoice vs Paid Tools

The video compares FluidVoice against the paid tools Andrew and Adam were already using: Wispr Flow and Superwhisper. The point is not that paid tools are bad. The point is that many paid dictation apps package open or available speech models into a polished app, while FluidVoice makes a serious version of that workflow open-source and local-first.

For me, the decision would not be "which app is best?" It would be "which app lets me speak useful instructions into my actual work surfaces with the least friction and the least privacy compromise?"

Setup Notes

From the video and official docs, the basic setup path is:

1. Download FluidVoice from the official site or GitHub releases, or install with Homebrew.
2. Grant microphone and accessibility permissions so it can listen and type into other apps.
3. Choose a voice engine or model profile, balancing speed and accuracy.
4. Pick the hotkey and continuation behavior that fits your muscle memory.
5. Add custom dictionary terms: names, company names, product names, client acronyms, and weird spellings.
6. Optional: enable AI enhancement and define prompts for email, Slack, coding, notes, or client messages.

The official FluidVoice page says macOS 15.0 or later is required, with Apple Silicon and Intel support. It also says iOS and Windows are coming, so Windows users should treat this as "watchlist" rather than "install today."

Agent Workflows

The most useful way to think about FluidVoice is not as a replacement keyboard. It is a prompt-capture layer.

Here are the workflows I would test first:

• Claude Code brief: dictate the bug, what changed recently, what not to refactor, and what tests to run.
• Codex task: speak a goal, constraints, files to inspect, and verification steps before starting a local agent run.
• Client email: speak rough notes, then let app-specific cleanup turn them into a professional message.
• Meeting follow-up: dictate quick action items immediately after a call while the context is still fresh.
• Research capture: speak a source note into Obsidian, Notion, or a Markdown file without breaking reading flow.
• Prompt library building: dictate raw workflow patterns, then clean them into reusable skills or templates.

Custom dictionaries matter here. If your AI tools constantly misspell client names, repo names, product names, or Portuguese/French/English mixed terms, the prompt quality quietly drops. Dictation is only useful if the text lands close enough that you trust it.

Privacy And Caveats

FluidVoice is local-first, but privacy still depends on configuration. The official FAQ says local speech models keep dictation on-device, while AI enhancement can use optional providers such as OpenAI, Groq, or custom providers. If you enable cloud providers, your privacy model changes.

Before using any dictation app for sensitive work, check:

• whether raw audio leaves the machine,
• whether transcripts are saved in history,
• whether AI enhancement uses a local model or cloud API,
• what analytics are enabled,
• whether accessibility permissions are acceptable for your threat model,
• whether your company allows third-party dictation tools.

For normal solo-builder work, local dictation is a strong default. For regulated client work, legal work, healthcare, finance, or confidential product strategy, test the settings carefully and document what is local versus cloud.

Builder Checklist

1. Install FluidVoice on a non-critical Mac first.
2. Test raw dictation into Notes, browser, Claude Code, and your editor.
3. Add 20 custom dictionary words: your name, clients, product names, repo names, and acronyms.
4. Create one cleanup prompt for Slack-style messages and one for technical prompts.
5. Run a privacy check: local model only, cloud providers off unless intentionally enabled.
6. Dictate one full agent task and compare it to your normal typed prompt.
7. Keep the tool only if it reduces friction without making you over-explain or over-edit.

FluidVoice is exciting because it makes the interface to AI feel less like a text box and more like a conversation with your own work system. That is where these tools become useful: not when they are flashy, but when they remove one small bottleneck every hour.

Sources

• The Next New Thing: Free, local audio dictation
• The Next New Thing
• FluidVoice official site
• FluidVoice GitHub repository
• FluidVoice releases
• Homebrew
• Wispr Flow
• Wispr Flow pricing
• Superwhisper
• Superwhisper vs Wispr Flow
• Apple Support: Dictate messages and documents on Mac
• Apple Support: Commands for dictating text on Mac
• JQ AI SYSTEMS: Open-source AI stack
• JQ AI SYSTEMS: Free AI business stack