Status and roadmap

What works in Pinna today, what is known to be missing or rough, and what is planned next.

Pinna is in early development. You can build it, install it on an Echo (2nd gen) or an Echo Dot (2nd gen) and use it with Home Assistant, or on its own in standalone mode, and with its own dashboard. Parts of it are still being tuned on real hardware.

Until Pinna 1.0, how it works and how you set it up can change from one release to the next, standalone mode most of all. The release notes say what changed. What owners report shapes these changes; see How you can help.

What works

These parts are covered by tests that run without hardware:

  • The kernel, the signed root filesystem image and the boot image build from the repository. The kernel is Debian’s 6.12 with the MT8163 board support ported to it.
  • The image boots in a virtual machine, in both the ARM and the x86_64 version, and its services, SSH, the Home Assistant API, the dashboard and the Prometheus metrics come up.
  • A device with no system it can start stays in rescue mode instead of restarting.
  • An update installs a kernel and a root filesystem into the second slot and boots on trial, and a broken update rolls back to the previous slot, kernel included, automatically. The bootloader’s part of this is tested on the Echo Dot (2nd gen) only.
  • The device checks an update channel for a newer signed release, shows it on the dashboard and in Home Assistant, and installs the release it showed (tested against a simulated release service and in the x86_64 virtual machine).
  • With Install updates set to Automatically, the satellite installs a release by itself at a random minute in a window that you choose, and restarts in the window once it is Idle. Releases that need something by hand wait for you. Its tests run against a simulated release service; no device has installed a release this way yet.
  • The release service’s channels and downloads pass their tests against recorded GitHub answers. The browser installer checks a release’s signature and files in its tests.
  • The Home Assistant connection passes tests with Home Assistant’s own client library, including encryption and key rotation.
  • The wake word detection gives the same results as the reference implementation, within 0.001.
  • A complete voice request, from the wake word to the played reply, works against a simulated Home Assistant. The reply plays while it downloads, the wake word interrupts a reply and stops a ringing timer alarm, and changes on the dashboard reach Home Assistant. A wake word uploaded on the dashboard is tested, offered to Home Assistant and wakes the device.
  • Standalone mode runs a request against simulated services: a spoken or typed request goes through speech-to-text, the brain’s tool loop or an agent, and text-to-speech that plays while it streams. The adapters for OpenAI-compatible services, Anthropic, Gemini and ElevenLabs, the save-time checks, the MCP server and client, device timers, the update tools and the dashboard’s setup API pass their tests.
  • In a satellite group, satellites add and remove each other, pass messages and act on each other in tests that run several satellites in one process. No test has run on two devices yet.
  • The setup hotspot runs in the x86_64 virtual machine with simulated Wi-Fi radios: a phone joins it, gets an address, is sent to the dashboard and scans for networks there; a join turns it off, a wrong password brings it back and the phone reads why, and a kept network keeps it off.
  • AirPlay turns on and off in the x86_64 virtual machine: the receiver starts, publishes its AirPlay record, reports that it is ready, and stops again.
  • The Sendspin player, which is experimental, plays sessions with a test server over a WebSocket on the same computer: the encrypted handshake, audio on the server’s timeline, volume and mute in both directions, a paused group and its controls, a rename, a second server taking over from the first, and refusing a browser or a third connection. In a simulation of the speaker, timed playback stays within 1 ms of the server’s clock from 10 seconds after a stream starts.

On the devices:

  • The microphone path with beamforming and echo cancellation runs; on the Echo Dot (2nd gen) it lowers the device’s own music by 23 to 31 dB. See Audio.
  • scripts/backup-gen2.sh backs up the device, and scripts/verify-backup.sh checks the backup. See What you need.
  • A hardware test (just test-device) checks the speaker, the microphones, the light ring and the Home Assistant connection on a running device.
  • Music plays over AirPlay and Bluetooth from phones and computers: turning a source on, pairing a phone, playing, the sender’s volume, reconnecting, forgetting, turning it off, and starting after a restart.

Known gaps

  • The dashboard’s password is optional, and it travels unencrypted. Until you set one, anyone on your network can use the dashboard, including restarting and updating the device. The satellite serves plain HTTP, so someone who can read the traffic on your network can read the password too. A reverse proxy with TLS in front of the dashboard protects the path from your browser to the proxy; the hop from the proxy to the satellite stays plain HTTP over your network. See Who can use the dashboard.
  • The built-in wake words’ licence doesn’t allow commercial use. Pinna listens for up to two of Alexa, Hey Jarvis, Hey Mycroft, Hey Rhasspy (openWakeWord) and Okay Nabu at once, and the openWakeWord models’ licence doesn’t allow commercial use. You can add your own.
  • The test of an uploaded wake word doesn’t say whether it hears you. The satellite checks that the model loads, keeps up and stays quiet on silence and noise, not that it wakes on its phrase. Models that read more than the standard 16 embeddings, which openWakeWord trains for some long phrases, are refused.
  • On a Kaeru from before 22 October 2026, updates don’t change the kernel. That Kaeru loads boot_a for both slots, and updates leave boot_a alone. Install new kernels from fastboot.
  • The setup hotspot hasn’t run on an Echo yet, and its password, pinnasetup, is the same on every device. While a device has no saved network, anyone in range who knows the password can join its hotspot, use its dashboard, and see what you type there.
  • No automated test runs. The tests run when contributors run them; there is no CI for them yet. The release workflow builds the images but doesn’t run the tests.
  • Standalone mode is new. Its tests run against simulated services. It has been tried on an Echo Dot (2nd gen) with OpenRouter and ElevenLabs, not yet with each hosted service.
    • MCP servers that need a sign-in or a token in a header can’t be added.
  • Sendspin is experimental. Its specification is still a draft. Music Assistant connects without pairing, so anyone on your network can play to the satellite. How closely the satellite keeps step with other speakers isn’t measured on a device yet: the delay from the speaker’s hardware clock to the sound counts as 0 until it is. See Play music with Sendspin.

Planned

  • A score for each wake word: the highest score of the last thing each word heard, so that you can tell a word that never wakes from one that nearly does.
  • More for standalone mode: MCP servers that need a sign-in, and several satellites that share one conversation. In a satellite group, one answer when several hear the wake word.
  • A setup hotspot password that only the owner can learn, instead of one that every device shares.
  • 64-bit ARM: the processor supports it, but nobody has started on it.
  • Pairing for Sendspin through Home Assistant: Home Assistant would hand the satellite’s pairing key to Music Assistant, so that only your Music Assistant could play to it.

These plans follow the project’s goals.

How you can help

  • Say what you think. Before 1.0, feedback changes the most. Say what works for you, what doesn’t and what you miss, in Discussions. Standalone mode needs it most.
  • Test it on a device. Problems show up on real hardware first. Report what you see in an issue, with the version and the log from the dashboard, as Report a problem describes.
  • Send changes. See Contribute.
  • Bring up a new device. Supporting another Echo device means work on the device tree, clocks, audio, microphones and LED ring, and a device to test on. Echo devices lists which ones can be unlocked. Open an issue with the hardware you have.

Edit this page on GitHub