flowflow

Most of my ideas come when I'm walking or between two tasks. And they vanish just as fast.

FlowFlow is a voice notes app that captures what you say, transcribes it, and lets you chat with your notes later. You ask "what was that pricing idea?" and the right passages come up - with links to the original notes.

No manual searching. No folders to dig through. Just talk, and find it later.

Get it

• iOS - on the App Store, available worldwide.
• macOS (Apple Silicon) - grab the DMG from the latest release, drag FlowFlow to Applications. Signed and notarized by Apple: it opens like any Mac app.
• From source - see Build below; one make installs on your own iPhone or Mac.

Highlights

Capture

• One composer for notes and chat: type, or tap the mic and a dark voice capsule records, transcribes and hands the text back where you were typing
• Once there is something to send, the orange mic stretches into a pill with a send button, so you can dictate again and the words land after your text
• Live 60fps waveform, pause/resume, Dynamic Island live timer; the capsule stays on screen until the text is ready, with retry if a transcription fails
• Cloud (Soniox) or fully offline transcription with local Whisper models
• Long recordings finish: local Whisper works in one-minute chunks cut at a silence, saves after each one and resumes after a lock, crash or kill; on iOS 26 it keeps going with the screen off, its progress on the Lock Screen
• Soniox uploads go up compressed (AAC) and are sent again after a kill or lock
• Personal dictionary: your names and brands spelled right, everywhere

Read a long meeting

• A recorded note reads like a document: paragraphs split at pauses, every word plays the audio from there, the word being read highlighted
• Chapters in the note's language, a reading bar stuck at the top (play, pause, current chapter) and a sheet to jump to any chapter
• On a long note, a scrubber on the right edge snaps onto each chapter and names where you land
• "Summarize the meeting": key points, decisions, timed actions and open questions in the app's language, kept with the note and synced; tap a time to play from there
• Share a note by link from its "…" menu; "Private" cuts the link in one tap

Ask your notes

• RAG chat with tappable sources: hybrid search (BM25 + vector + rerank)
• Answer with Claude or OpenAI: pick the model (Claude Haiku, Sonnet or Opus, GPT Sol or Luna) and how hard it thinks, from the + or the empty chat; each conversation keeps its own, a new one starts from the last
• Chat through a Claude or OpenAI API key, or a ChatGPT subscription; embeddings still use the OpenAI API
• Settings > AI: one row per provider (Claude, OpenAI, Exa), its keys and subscription one level down
• Optional web search (Exa) fused into the same answer
• Save any answer or thread back as a note

Organized for you

• AI titles, tags and themes as you write; searchable chats
• Theme search from the menu or the title; every folder starts open and the ones you close stay closed
• Threads: related notes as one chronological story; right-click or long-press a note within a thread to copy its full text or share that note
• Smart filters in the search bar: dictated, reminder, document, thread

Your devices, one brain

• Encrypted LAN P2P sync (Noise protocol): no server, no cloud
• One account for up to 3 paired devices; premium follows the pairing automatically
• One-archive backup, validated crash-safe atomic restore

Act, not just record

• "Pick up the kids at 5pm" becomes a calendar event, one tap to confirm
• Notes as actions: the assistant executes with your connected tools, every write holds for approval
• One + menu in notes and chat, native on iOS 26: discuss this note, add it to a thread, tools, agents and connectors, with live connection state and each product's own icon
• Scoped agents: installed packages are verified and pinned, each capability checked against its owner, every native action confirmed before it runs
• Governed connectors (Google Sheets), each through its own MCP peer; groundwork for a signed-agent marketplace

Talk to your Hermes Agent

• Chats > New conversation > With Hermes: your own Hermes Agent answers, its tool calls folded into one "N steps" row
• Link it by scanning one QR code; Hermes stays on your private Tailscale network
• Pick the model and reasoning level under the title; a reply keeps running when you leave the app and is there when you come back
• Attach photos (or take one), a PDF, Word or text file from the +: Hermes sees the photos, reads the file's text
• Your Hermes skills in the +, most used first, then by category; type / or a word that names a skill and it is suggested, one tap to call it
• Send to Hermes from any note: a new conversation opens with the note attached

Share a space with Hermes Agent

• Grant a Hermes Agent scoped, revocable access to one shared space from the space menu
• Hermes reads folders, notes and threads, writes its own notes and threads, and every change lands on your devices through the normal space pull
• Everything is a plain note on arrival: searchable and chat-ready

Native feel

• A real Mac app: ⌘N, ⌘F, ⌘⌘, view history, native file dialogs
• iPhone: the menu lies under the screen, which slides aside as a card; swipe from anywhere to open it
• iOS 26 Liquid Glass controls and native options menus (press, slide, release); haptics on every commit, drag-to-dismiss sheets
• English + French, down to error messages; word-level transcript with tap-to-seek

The full tour, one paragraph per feature: docs/FEATURES.md.

How it works

Talk   → Record → Transcribe (cloud or on-device) → Clean fillers → Apply dictionary → Auto-embed → Store → AI title

Ask    → Embed query → Hybrid search (BM25 + vector)  ∥  Web search (Exa, when enabled)
       → RRF fusion → LLM rerank → Temporal boost → Tag-enriched context → Agent with tools → Answer with sources

Sync   → Save → debounced trigger → Noise-encrypted LAN session → version-vector merge → UI refresh < 1 s

Hermes → question + photos (image parts) + files and notes (text) + skills (named)
  chat   → your Hermes API server (HTTPS, Tailscale only) → streamed run → steps + answer
         history stays in the Hermes session; FlowFlow keeps a pointer on the device

Hermes → mcps_ token (one space, read or read_write) → MCP tools on api.flowflow.be/v1/mcp-spaces
         → notes and threads written by the agent → pulled by every member device

Backup → Export scrubbed SQLite snapshot + WAV + manifest (zip) → share
         Import → read-only validation → atomic swap at next launch → vector index rebuilt offline

Hermes Agent

Chat with it. On the Hermes host, with Tailscale on both machines:

./scripts/enable-flowflow-chat.sh

It turns on Hermes' API server, exposes it on your Tailscale network only and prints a QR code. Scan it with the iPhone camera: FlowFlow opens Settings > Connections > Hermes with the address and key filled in. Check the address, then tap Save and test. Step by step: flowflow.be/hermes.

Share a space with it. FlowFlow ships a Hermes skill in skills/flowflow-spaces. Install it on the Hermes host, store the one-time token in ~/.hermes/.env, and add one MCP server per shared space:

./scripts/install-flowflow-hermes-skill.sh

mcp_servers:
  flowflow_projects:
    url: "https://api.flowflow.be/v1/mcp-spaces"
    headers:
      Authorization: "Bearer ${FLOWFLOW_TOKEN_PROJECTS}"

The MCP server reports a contract_version; the skill states the version it was written for and asks for an update when they diverge. Update with git pull && ./scripts/install-flowflow-hermes-skill.sh, or track the skill with hermes skills install <raw SKILL.md url> and hermes skills update. Full walkthrough, safety rules, rotation and revocation: docs/guides/hermes-flowflow.md.

Built with

Mainly Rust: the app itself is Rust end to end, UI included, with a few deliberate exceptions where another tool does the job better - a small TypeScript layer for webview gestures and the native menu bridge, Swift for the Live Activity and background transcription, and the web sites in Astro.