Skip to main content

Huddle

Huddle is a small desktop application for talking to a local agent. A live voice model handles the conversation. When a request needs work, Huddle passes it to an ACP agent and sends the result back to the voice model to speak aloud.

Built with GDK + ACP: the live provider interface from goose-providers, with goose or another ACP agent doing the delegated work.

Source

The clever bit​

The voice conversation and the working agent have different jobs. The voice model handles speech; a separate agent has the tools and context to carry out the task. They do not need to be the same model or run in the same process.

Delegation events identify a position in the spoken conversation. Huddle turns the transcript up to that point into an agent prompt, then returns the answer to the corresponding delegation. Requests queue while the working agent is busy.

This is a small example of composing provider capabilities with an existing agent instead of rebuilding the entire stack for a new interface.

How it works​

  1. The webview captures microphone audio and plays the response over WebRTC.
  2. Rust starts the live session through GDK and listens on its sideband control connection for transcript and delegation events.
  3. A delegation becomes a prompt to the persistent ACP session.
  4. The agent's result is returned through the live provider connection.

In voice.rs, the return path uses the provider's delegation interface:

let delivery = connection
    .send_delegation_update(DelegationUpdate {
        provider_delegation_id: delegation_id.clone(),
        text: text.clone(),
    })
    .await?;

The transcript code chooses the context for each request. The ACP client owns the agent subprocess and serializes prompts.

Run it yourself​

You need Rust 1.94 or later, an OpenAI API key with GPT-Live access, microphone permission, and an ACP agent on your PATH. The default is goose acp. Configure goose's model provider separately if you use that default.

With OPENAI_API_KEY set in your environment:

git clone https://github.com/jamadeo/goose-huddle.git
cd goose-huddle/src-tauri
cargo run

Open Settings to choose the agent command and working directory, then start a huddle. See the README for packaging and configuration details.

Prototype permissions

Huddle automatically approves agent tool requests and can save the API key in browser local storage. Try it in a disposable working directory with credentials and resources appropriate for a prototype.