# What Siphon is

> Voice capture from Apple Watch Ultra to the AI agent on your Mac. What it does, how the pieces fit, and what it needs.

Siphon turns the Action Button on Apple Watch Ultra into a direct line to your AI agent. Press it
and talk. Your words appear live on the watch and on your Mac. When you stop, the transcript goes
to your agent, a short answer comes back to your wrist, and the whole exchange is kept as a
thread on your Mac that you can carry on later.

Siphon is in early testing. The Mac app is a public download (the current version is
0.1.0-1027; the [Quickstart](quickstart.md) installs it). The watch app is still built from
source.

## Requirements

- Apple Watch Ultra (watchOS 26 or later), paired with an iPhone
- A Mac with Apple silicon running macOS 26 or later
- An agent to send captures to. [Hermes](https://github.com/NousResearch/hermes-agent) works
  today. Claude Code and Codex can also answer inside Siphon using their local sign-in.

## How it works

```
Apple Watch Ultra ──audio──▶ Mac: siphon-server ──text──▶ your agent (Hermes)
   live transcript ◀──────── speech-to-text on the       ──reply──▶ watch + Mac thread
                              Neural Engine
```

1. **Watch** (`apps/watch`): the Action Button starts recording in under a second. Audio streams
   to the Mac in half-second chunks, and the watch shows the transcript as it fills in.
   Recording stops after you've been quiet for about two seconds, or when you drop your wrist.
   Press again for the next thought as soon as one stops: it records straight away, while the
   one before finishes sending behind it. The iPhone's mic does the same.
2. **Server** (`crates/server`, on your Mac): stores each capture and runs speech-to-text locally.
   It uses NVIDIA Nemotron Speech Streaming 0.6B through FluidAudio, on the Neural Engine.
3. **Agent**: the server sends the finished transcript to your agent and asks for a one-line
   reply that fits a watch screen, followed by the full answer.
4. **Desktop app** (`apps/desktop`, Siphon.app): shows your conversations and live
   transcription, handles watch pairing, and keeps the Mac reachable when you're away from home.

## Choose your agent and model

On desktop, open **Settings → Connectors → Agent & Model**. On iPhone, open
**Settings → Agent & Model**. These settings choose the default for new threads on your host.
On desktop, use the separate **Agent** and **Model** dropdowns in the composer footer; the model dropdown
supports searching the selected agent’s catalog. On iPhone, tap the agent/model chip in the composer to open
that thread’s settings. Changing the default does not change existing threads. The watch uses the thread’s choice, or the host default for a
new thread.

Siphon discovers the agents installed and signed in on the host, and reads their model catalogs.
For Claude Code, install `claude` and run `claude auth login` on the host. For Codex, install
`codex` and run `codex login` there. Both use your own login; credentials stay on the host. Use
**Refresh** (desktop) or **Refresh agents** (iPhone) after changing setup (discovery results may be cached for up to 30 seconds).
An unavailable agent shows why it cannot be selected. **Agent default** uses that agent’s own
configured model; exact versions are shown on replies when the agent reports them.

Until you choose, Siphon picks for you: a new thread goes to the first of Hermes, Claude Code
and Codex that answers a check when the message is sent (Settings says "Picked automatically").
Being installed isn't enough: the agent has to be signed in and answer. Once you choose, your
choice is kept, even if that agent stops answering; its messages then fail with its own error
rather than going to another agent. With nothing chosen and no agent answering, captures are
still transcribed and kept as threads. They say "Saved. No agent connected yet." on the watch,
the iPhone and the Mac, and are sent by themselves once an agent is connected.

The visible Siphon thread stays together when you change agent or model. The next message starts
a fresh native session with a compact digest of the opening and recent exchanges. Long messages
and older details can be omitted, so repeat any essential detail if needed. Matching agent/model
choices reuse their native session. Messages already active or queued retain the choice they were
sent with, and previous replies keep their original attribution. A failed request never silently
switches to a different agent.

Claude Code runs as a chat agent with its tools disabled. Codex runs with its read-only sandbox
and no approval prompts. This is separate from handing work off to another app or repository.
T3 Code appears as unavailable in the agent picker until a Siphon conversation connector is implemented.
For coding tasks, the separate desktop **Hand off** action can [start a T3 Code task in a new
worktree](t3-handoff.md), after you review the brief and choose a project.

## Approving what the agent wants to run

When Hermes flags a command (a script run with `python3 -c`, a recursive delete), the turn waits
for you instead of being refused. You see it on all three devices:

- **Mac:** a card in the thread with the reason, the command, what the agent was doing, and every
  answer Hermes offers. **Always allow** asks again before it applies.
- **iPhone:** a notification with **Approve once** and **Deny** (approving needs the phone
  unlocked). The thread shows the full command and offers **Allow for this session** when Hermes
  does.
- **Apple Watch:** a notification and a card with **Approve once** and **Deny**.

A notification shows the command itself. If one shows only "Hermes needs your approval", **Approve
once** doesn't approve: "Open the thread to approve" appears instead, and tapping it opens the
thread, where you see the command and answer.

The first answer from any device counts. The others then show how it was answered, for example
"Approved once from Watch". If nobody answers within 10 minutes, it's denied and the agent is told
why. The thread records every answer.

## Privacy

- Speech recognition runs on your Mac. Audio goes from your watch to your own Mac, not to a
  Siphon cloud.
- Raw audio is deleted after 30 days. Transcripts and threads stay until you delete them. A
  recording that couldn't be transcribed keeps its audio until it has been, or you delete it.
- At home, the watch talks to the Mac directly over your network, using sealed (encrypted)
  requests once it's paired. Pairing happens on your network too, with no relay: you compare a
  code on both screens.
- Away from home, an optional paid relay carries those same sealed requests. It forwards them
  without being able to read them, and your Mac still does all the work. Notifications when your
  agent replies, or when your Mac goes offline, come through the relay too: without one, the
  watch and iPhone get none, and what you say away from home waits on the device until it's
  back.
- A Mac signed in to a relay account sends a daily summary of usage counts (never content). You
  can turn it off in Settings. The [privacy policy](privacy.md) has the details.

## Pricing

- **At home:** free. Self-hosted on your Mac and your network, pairing included.
- **Away from home, and notifications:** a paid relay subscription (planned). A watch or iPhone
  paired at home picks it up when you add it, with no pairing again.

## Next

- [Quickstart](quickstart.md): set Siphon up on a Mac and a watch.
- [Troubleshooting](troubleshooting.md)
- [Privacy policy](privacy.md)
