Set up standalone mode

Set up standalone mode, where the satellite answers requests itself with the speech services, language model or AI agent you choose.

In standalone mode the satellite answers voice requests itself. After the wake word it sends what you say to services that you set up on the dashboard: speech-to-text, a brain (a language model or an AI agent such as Hermes or OpenClaw) and text-to-speech. It keeps timers itself, tells the weather and plays radio. AI agents can also use it over MCP, the protocol they use to call tools.

Standalone mode is new. Its tests run against simulated services, and it has been tried on an Echo Dot (2nd gen) with OpenRouter and ElevenLabs, not yet with every service; see Status and roadmap. Until Pinna 1.0 it can change from one release to the next. Say what works for you and what doesn’t in Discussions.

Before you begin

  • Pinna runs on the device, and you can open its dashboard. If Assistant has no Services, the device runs a Pinna from before standalone mode: update it first.

  • You have an account with a key for each hosted service you want to use, or your own servers on your network. Each service offers some of the three parts:

    Service Speech-to-text Brain Text-to-speech
    OpenAI Yes Yes Yes
    OpenRouter Yes
    Anthropic Yes
    Google Gemini Yes
    Groq Yes Yes
    Mistral Yes
    ElevenLabs Yes Yes
    Ollama, LM Studio Yes
    Hermes, OpenClaw (agents) Yes
    Other server Yes Yes Yes

    One OpenAI key can serve all three. Other server is any server that speaks OpenAI’s API, such as a self-hosted speech server.

  • Hosted services charge for what you use. The satellite uploads audio only after the wake word, and only when it heard speech. A TV or a conversation that sounds like the wake word still costs a request.

After the wake word, the audio goes to the services you chose, and nowhere else. Two more services are fixed: Pinna looks up places and the weather at Open-Meteo, which gets the name and coordinates of your place and of any place you ask about, and searches for radio stations at radio-browser.info, which gets what you searched for. Leave Place empty and turn off the Weather and Search radio tools to use neither. Neither needs an account.

Anyone who can use the dashboard can change your services and use them; see Who can use the dashboard.

Add the services

  1. Open the dashboard, select Assistant, and then select Services.

  2. Under Connections, select Add… and choose the service.

    • For a hosted service, enter its API key. You make the key in your account with the service, for ElevenLabs under its API keys. One ElevenLabs connection serves both speech-to-text and text-to-speech. A key with limited permissions needs speech-to-text and text-to-speech, and reading voices and models for the lists below.
    • For Ollama, LM Studio or Hermes, enter the computer it runs on, such as 192.168.1.20. Pinna adds the port and the path.
    • For OpenClaw, enter the computer and the port its gateway listens on, such as 192.168.1.20:8080.
    • For Other server, enter its whole address, up to and including /v1.
  3. Select Add.

  4. Under Speech-to-text, Brain and Text-to-speech, choose a connection, enter the model, and select Save and check. Text-to-speech asks for a voice, and optionally the model that speaks it; empty uses the service’s usual one. Each field offers what the service lists when you open it with the arrow at its end or the arrow keys, and typing narrows the list. You can also type a name it doesn’t list. For example:

    Part OpenAI ElevenLabs OpenRouter
    Speech-to-text A transcription model, such as whisper-1 or gpt-4o-mini-transcribe A Scribe model, such as scribe_v2. ElevenLabs doesn’t list them (none)
    Brain A chat model, such as gpt-4.1-mini (none) A model with its provider, such as deepseek/deepseek-v4.1-flash
    Text-to-speech A voice, such as alloy or nova; the model is tts-1 unless you choose another A voice from your own voice library: the list shows the names and enters the ID. The model is eleven_multilingual_v2 unless you choose another, such as eleven_flash_v2_5 (none)

    The check makes the cheapest call that proves the service works: it lists the models, transcribes half a second of silence, or synthesises one word. It never sends a prompt. If it fails, the page says why, for example that the key is wrong, or what the service said about a model it doesn’t have. The brain’s model must be in the service’s list.

  5. Under Text-to-speech, select Play sample to hear the voice on the satellite.

Some reasoning models refuse requests with tools until you choose how long they think. OpenAI’s gpt-6-luna, for example, uses tools only with thinking off. Under Brain, select Advanced, under Thinking choose None, and select Save. None also answers fastest. Thinking is there for OpenAI and the other services that speak OpenAI’s API; Service default sends nothing.

A key goes to the satellite and never comes back: the page shows only that it is set, with its last four characters. A key belongs to the address it was entered for. If you change a connection’s address, enter the key again.

Your own servers

  • Ollama listens only on its own computer until you set OLLAMA_HOST=0.0.0.0 for it. Enter the model as ollama list shows it; llama3.2 works for llama3.2:latest. Choose a model that supports tool calls: the satellite offers the brain its tools with every request, and a model without tool support refuses every request.
  • LM Studio has to serve on the local network, not only on its own computer. The same rule about tool calls applies.
  • Hermes and OpenClaw: see Use an AI agent.

Set up the conversation

Select Assistant, then Conversation.

  • Language: a tag such as en, de or en-US. Speech-to-text listens for the language, and the brain answers in it.
  • Place: a town or city, for the weather and so the brain knows where it is. Select Find. You can still ask for the weather somewhere else, such as “the weather in Paris”; add the country when a name is ambiguous. If the place is in another time zone than the satellite, the page offers to use the place’s.
  • What the brain is told: the prompt, at most 4,000 characters. It starts as the satellite’s voice rules: short, plain spoken answers. Reset to the voice rules brings them back. With each request the satellite adds the local time, the place, the language and its name, and in a satellite group the other satellites’ names, with those in Home Assistant mode named as out of reach.
  • Follow-ups: on at first. After each reply, the satellite listens for 4 seconds more, without the wake word, so you can answer. If you say nothing, it stops listening.
  • Acknowledgment: on at first. A short phrase, “One moment.” in your language, that the satellite says when an agent hasn’t answered after 3 seconds, so you know it heard you.
  • Ask a question: type a question and select Ask. The brain gets it as if you had said it, and the satellite answers aloud. Use it to try the setup from your computer.

The conversation stays in the satellite’s memory only. A request within 5 minutes of the last one continues it. It is forgotten after 5 minutes without a request, when you change the brain or the mode, or when the satellite restarts.

Switch to standalone mode

Before you switch, set the language and the place under Conversation (above): the weather needs a place unless you name one when you ask.

  1. Select Assistant.
  2. Under Mode, select Standalone. The button works once all three services have passed their check.
  3. Select Switch.

The satellite ends any request it is working on, stops a ringing alarm, cancels the timers that run and forgets the conversation. In standalone mode the dashboard chooses the wake words, under Voice & sound.

To go back, select Home Assistant under Mode, then Switch. If Home Assistant isn’t connected, nothing answers requests until it connects.

If Home Assistant is connected, you can also switch there, with the satellite’s Voice mode select. It stays in both modes, and Standalone works only once the three services have passed their check. If the satellite is answering a request when the select changes, it switches once the request has ended, after the reply and any follow-up.

Switch by voice

To switch back to Home Assistant, say the wake word and “switch to Home Assistant”. The satellite switches once its reply has played, and doesn’t listen for a follow-up. As with Switch on the dashboard, the timers are deleted, and if Home Assistant isn’t connected, nothing answers requests until it connects. This uses the switch_assistant tool under Tools, so it works while that tool is on. With Hermes or OpenClaw as the brain, it works only if the agent uses Pinna’s MCP address.

To switch to standalone mode while Home Assistant answers, Home Assistant needs something that sets the Voice mode select, because Assist has no command that changes a select:

  • With an AI conversation agent that may control Home Assistant: in Home Assistant, create a script under Settings > Automations & scenes > Scripts, and expose it under Settings > Voice assistants > Expose. In its Description, say when to use it. Replace the entity ID with your satellite’s: copy it from the Voice mode select on the device page.

    alias: Switch the Echo Dot to standalone mode
    description: >-
      Switches the Echo Dot to Pinna's standalone mode. Use it when the person
      asks to switch to standalone mode.
    sequence:
      - action: select.select_option
        target:
          entity_id: select.echo_dot_voice_mode
        data:
          option: Standalone
  • With Home Assistant’s built-in agent, or an AI conversation agent set to prefer handling commands locally: create an automation with a sentence trigger such as “switch to standalone mode”, and the same select.select_option action.

If the three services haven’t passed their check, the satellite refuses and the select goes back to Home Assistant, even if the agent says it switched. Switching deletes the timers that Home Assistant set on the satellite too.

If you also use Home Assistant

Home Assistant can stay connected in standalone mode. It keeps the media player, announcements, volume, mute, the ring and the buttons, but it no longer hears requests: it removes the satellite’s voice assistant and wake word choices until you switch back. When you switch back, Home Assistant answers requests again as soon as it is connected.

Choose what the brain may do

Select Assistant, then Tools.

  • Pinna’s tools lists what a brain can do on the satellite: timers, the time, the weather, the volume and the reply volume, the ring light, playing, pausing and stopping music and radio, clearing a notification, muting the microphones, switching back to Home Assistant mode (switch_assistant), saying which version runs, checking for updates, and passing messages to other satellites (send_message). In a satellite group, the brain can also control another satellite. Only a person can turn the microphones back on, with the Microphone off button. Turn off a tool you don’t want used. Install update is the one tool that starts off: turn it on to let the brain install an update, as Ask the satellite to update describes.
  • Radio stations: add a station with its name and stream address. Then ask for it by name, such as “play Radio Eins”. The brain can also search the radio directory for stations.
  • Your MCP servers: tools from MCP servers on your network, next to Pinna’s. See Add an MCP server.

Add an MCP server

The brain can call the tools of MCP servers on your network. The satellite connects to them over MCP’s Streamable HTTP.

  1. Under Your MCP servers, select Add….
  2. Enter a Name, such as home, and the server’s whole Address, such as http://192.168.1.20:8080/mcp. A token that is part of the address works. A server that needs a sign-in or a token in a header, such as Home Assistant’s own MCP Server integration, can’t be added yet.
  3. Select Add. The satellite asks the server for its tools, and the row shows how many of them are on, such as 0 of 77 tools on, or why the server can’t be reached.
  4. Select Its tools under the server. Each tool has a switch and the server’s description of it. All on and All off switch them all.

A server that lists more than 32 tools when you add it, such as ha-mcp, starts with all of them off. Most of ha-mcp’s tools manage Home Assistant itself, such as removing automations, so turn on only the few a voice request needs, such as its search, state and service tools. If the server doesn’t answer when you add it, or lists new tools later, those tools start on.

The brain takes at most 64 tools in a request: Pinna’s tools that are on first, then each server’s in the order the page lists them. Tools past 64 are left out of every request, so a server below a big one can lose all of its tools. While more than 64 are on, the page says how many are left out, and System › Logs names the first one. Switch off tools, or a whole server with the switch on its row.

Use timers

Ask for a timer as you would ask anyone, such as “set a tea timer for four minutes”. The ring shows an amber arc while it runs, and the alarm rings when it ends. Say the wake word or press the Action button to stop it. With an agent as the brain, the agent sets the timer through Pinna’s MCP server.

Up to 10 timers run at once, each for up to 24 hours. They survive a restart of the satellite.

When a request fails

The satellite plays the error sound and the ring flashes red. Assistant shows the last failure, with what failed, the service and the time, until a request is answered again. System › Logs has a line for each failure, with what the service said, a line for each tool the brain called, and one line for each request with how long each part took and how it ended; select All there to see the last two. The log never contains what was said, and it masks keys. See Standalone mode doesn’t answer for what the messages mean.

What’s next

Edit this page on GitHub