Skip to main content

Compositions

A composition is a saved, named configuration of voices. Create one to quickly launch consistent multi-voice sessions without reconfiguring voices each time.


Creating a Composition

Click the New Composition button in the sidebar (below the session list).

Sidebar showing the New Composition button below the session list

The composition builder opens. Give your composition a name, then add voices.

Composition Builder in empty state with name field, mode selector, and Add Voice button


Choosing a Mode

The first setting in the Composition Builder is the session mode. This controls how your messages are dispatched when you start a session from this composition.

ModeBehavior
Conductor-DirectedYou direct each message to a specific voice. Other voices see the exchange but stay silent.
BroadcastYour message goes to all voices simultaneously. This is the default.

The continuation policy (below) is only available in Broadcast mode.


Continuation Policy

When Broadcast mode is selected, the Continuation policy section appears below the mode selector. This controls what happens after the first round of responses completes.

PolicyBehavior
NoneVoices respond once and wait. No additional rounds start automatically.
Prompt meAfter each round, a nudge banner appears in the session asking whether to continue. You click Allow or Dismiss.
AutoVoices continue responding automatically until the max rounds limit is reached.

When Auto is selected, a Max rounds slider appears (range 1–3). This caps the total number of automatic continuation rounds.

Composition Builder showing Broadcast mode with continuation policy set to Auto and Max rounds slider visible


Adding Voices

Click Add Voice to add a voice to the composition. For each voice, configure:

FieldDescription
ProviderThe AI provider (Anthropic, OpenAI, Gemini, Claude CLI, custom provider, etc.)
Voice typeAPI or CLI — shown for providers that support both
ModelThe specific model to use (e.g. claude-sonnet-4-6); populated from the provider
Display nameAn optional display name for this voice in sessions
Avatar iconAn optional emoji or symbol shown next to this voice’s messages in the feed
ColorA color swatch used to identify this voice in the message feed
System prompt templateAn optional saved template to attach (see System Prompt Templates)
System promptOptional inline instructions that shape this voice’s behavior
TonePer-voice tone override — or “Use conductor default” to inherit from your profile
ToolsFor API voices: an optional set of host-brokered tools the voice can invoke — file reads, writes, directory listing, command execution, and URL fetching. Not shown for CLI voices. See Filesystem Tools for the full list.

Voice configuration panel fully configured with provider, model, display name, avatar icon, color, tone, and Tools section visible

You can add as many voices as you like. There is no hard limit, though more voices means more tokens and longer wait times per round.


Voice Type Availability

When you select a provider in the Composition Builder, the Voice type toggle (API / CLI) shows whether each type is available based on your current configuration:

  • If an API key is not configured for that provider, the API button is disabled with the tooltip “No API key configured”.
  • If the CLI binary for that provider is not found in your PATH, the CLI button is disabled with the tooltip “[binary] not found” (e.g., “claude not found”).
  • If neither type is available, the provider button in the grid is grayed out and cannot be selected.

This prevents adding a voice that would fail when a session starts.

To configure credentials, see Voice Providers (API keys and CLI tools) or Custom Providers (custom endpoints).


Attaching a System Prompt Template

In the voice configuration panel, the System prompt template dropdown lets you attach a saved template to a voice:

  1. Select a template from the dropdown. The system prompt textarea is pre-filled with the template’s content and a “Template attached” badge appears.
  2. To go inline instead, edit the system prompt textarea directly — editing automatically detaches the template.
  3. To re-attach a template after editing inline, select it again from the dropdown.

See System Prompt Templates for how to create and manage templates.

Voice configuration panel with Security Reviewer template attached and Template attached badge visible


Adding a Custom Provider Voice

Custom providers configured in Settings → Custom Providers appear in the provider grid alongside built-in providers. Select one to open the voice configuration form. The model list is populated from that provider’s endpoint.

See Custom Providers for setup instructions.


Enabling Tools on a Voice

API voices can be granted access to host-brokered filesystem tools — file reads, writes, directory listing, command execution, and URL fetching. The voice decides when to use them; Polyphon executes the operation on your machine and returns the result.

In the voice configuration panel, the Tools section appears below the system prompt for API voices. Enable only the tools the voice needs for the task.

Write-capable tools (Write File, Move / Rename File, Copy File, Delete File, Run Command) show an amber warning when selected — they can modify or delete files on your machine.

CLI voices do not have a Tools section. They run as autonomous subprocess agents with their own file access.

Voice configuration panel showing Tools section with read-only tools checked and write-capable tools unchecked

See Filesystem Tools for descriptions of every tool, sandboxing, and security considerations.


Reordering Voices

Drag voices in the voice list to change their order. The order determines how voice responses appear in the message feed — voices respond in parallel, but are displayed in composition order.

Composition Builder voice list with drag handles on each voice row


Saving a Composition

Click Save to save the composition. It will appear in the sidebar under your session list.

Compositions are saved locally to your SQLite database. They are not synced to any cloud service.


Launching a Session from a Composition

Click a composition in the sidebar to open it, then click Start Session. A new session is created with all the voices from the composition pre-configured.

You can launch as many sessions from the same composition as you like. Each session is independent.

Saved composition detail view showing name, voice list, mode, continuation policy, and Start Session button


Editing a Composition

Open a composition from the sidebar and click Edit. The composition builder reopens with the current configuration. Make your changes and click Save.

Editing a composition does not affect sessions that were already started from it.


Archiving a Composition

To remove a composition from the sidebar without deleting it, right-click it and select Archive. Archived compositions are hidden but not deleted.

Right-click context menu on a composition showing Archive and Delete options