A New Brain for an Old Echo Dot

The model number is on the bottom of the Dot.

Model	Status (Nov 2026)	What it takes
Echo Dot 2nd gen (2016), RS03QR	Works; this guide walks through it	A micro-USB data cable and a button press
Echo 2nd gen (2017)	Supported by firebreak and TECHO5 (tested on one unit so far)	USB and a button gesture; see the TECHO5 docs
Echo Dot 3rd gen, 2018 board	Supported by firebreak and TECHO5 (tested on one unit so far)	USB and a button gesture; some older guides solder to test pads instead
Echo Dot 3rd gen 2019 and later, Dot with clock	Not compatible	—
Echo Dot 4th/5th gen, Echo Pop	No known way in	—

What you give up: Alexa's cloud services, such as skills and shopping. Those only ever lived in Amazon's cloud.

• Keep the Dot offline until it's unlocked. If it reaches Amazon it can update itself to a build the unlock doesn't support. A Dot that has sat unused in a drawer is ideal.
• Use a cable that carries data, plugged into a port on the computer itself, not a hub. A charge-only cable gives power and nothing else.
• Install the tools: adb, fastboot, git, Python 3 and uv. You also need Home Assistant on your network.

sudo apt install adb fastboot git libusb-1.0-0
curl -LsSf https://astral.sh/uv/install.sh | sh

On Linux, add the udev rules from the firebreak README so your user can talk to the Dot in each of its USB modes, then reload udev and log out and back in. Stop ModemManager while you work: it grabs new serial devices and can steal the Dot's brief USB window.

sudo systemctl stop ModemManager

Start firebreak first, with the Dot unplugged. It waits until it sees an Echo.

uvx firebreak@latest --verbose unlock

Know about the one-second window. When a stock Dot powers up over USB, it shows up as a MediaTek preloader for about one second, then boots into Fire OS and goes quiet on USB. The Dot isn't broken; the computer sees it for that one second. Watch it with journalctl -k -f: seeing the preloader appear confirms your cable carries data.

1. Unplug the Dot.
2. Press and hold the action button (the one with a dot).
3. While holding it, plug the cable back in.
4. Let go when the light ring turns green.

Ring	Meaning
Green	Fastboot. firebreak takes it from here.
Orange	It booted normally into Alexa's setup mode. Harmless: unplug and try again.
Solid blue	Usually a charge-only cable or a bad port. Try another cable, then another port.

firebreak unlocks the bootloader, roots the Fire OS 6 already on the Dot (your setup is kept), and blocks Amazon's update servers on both system slots. Ours took 7 minutes 44 seconds and ended with:

The Echo Dot (2nd Generation) is rooted: Fire OS 6574.1 (NS65741/8149), with root adb.

Back up ~/.android/adbkey. The rooted Dot now trusts only that key.

TECHO5 replaces Fire OS with a small Alpine Linux that turns the Dot into an ESPHome voice satellite. It needs the Dot on Fire OS 6574.1 with root adb, which is exactly where firebreak leaves it.

git clone https://github.com/HuskerMinion/techo5-dot
cd techo5-dot
python3 tools/install-dot.py --dry-run

Dry run first. It checks the device, backs up every partition the Dot boots from (13 of them, each verified by md5), downloads and checks the signed release, and builds this Dot's boot image. It writes nothing to the Dot. It asks for your Wi-Fi network name, the password (typed without echo) and a name for the Dot in Home Assistant, and saves a Home Assistant encryption key in backups/<serial>/home-assistant.key.

Then run the real install, pinned to the release the dry run checked:

python3 tools/install-dot.py --release dot-v1.0.61 --name "Kitchen" --wifi "YourNetworkName"

It writes the boot image to the recovery partition and reads it back, unpacks the Linux root filesystem into slot A, sets the Dot to boot Linux, reboots, and waits on the Dot's USB console for a healthy boot. Ours came up healthy on the first try.

Keep backups/. It's your way back to stock Fire OS, and it holds the Home Assistant key.

Safety nets: Fire OS and TWRP stay on the device. After five bad boots, the Dot drops into a rescue mode with a USB console and SSH. Updates are signed, install to a spare slot, and roll back on their own if they don't come up healthy.

If the ring flashes yellow, the Dot is running TECHO5 and healthy, but it isn't on Wi-Fi yet.

Check the network name before you blame the password. Wi-Fi names are case-sensitive. Ours had a one-letter typo: the Dot's own scan, read over its USB console, showed it stuck on wpa_state=SCANNING and listed the real name.

Re-send the network over USB; there's no need to reinstall:

python3 tools/set-wifi.py --serial <your-dot-serial> --ssid "ExactNetworkName"

It asks for the password again and reports Joined: … is at <address> once the Dot is on the network. Your router also needs to allow WPA2: a WPA3-only network won't work. As a fallback, hold the mute button for 5 seconds to send the Dot its network over Bluetooth (Improv).

1. Home Assistant should discover the Dot as an ESPHome device under Settings → Devices & services. If it doesn't, add the ESPHome integration by hand with the Dot's IP address and port 6053.
2. When it asks for the encryption key, paste the contents of backups/<serial>/home-assistant.key.
3. Under Settings → Voice assistants, build or pick an Assist pipeline:
   • Speech-to-text: for example Whisper. Fast on a GPU machine; on small hardware, Speech-to-Phrase is quicker for fixed commands.
   • Conversation agent: for example a local LLM through the Ollama integration.
   • Text-to-speech: for example Piper.
4. On the Dot's device page, choose its pipeline and its wake word: Okay Nabu, Hey Jarvis, Hey Mycroft, Alexa, Computer, Jarvis, Hey Friday, GLaDOS, HAL, Terminator, Marvin or Home Assistant.

Say the wake word and ask it something.

Beyond voice, the Dot can be a multiroom speaker through Music Assistant, an AirPlay or Spotify Connect speaker, a Bluetooth speaker, and a Bluetooth proxy for Home Assistant. The light ring can show wake, listening, thinking, replying and error, and it can pulse with music.

The Dot only listens and speaks; the conversation agent decides what to do. A bad one plays the wrong song, claims it did something it didn't, or makes up the news. So rather than pick on reputation, we tested.

The test. 18 real requests from Marvin's log (music, news and facts, the house, small talk), each run twice. Every model got Home Assistant's real system prompt and all 29 of its tools, and the harness answered tool calls with recordings, so nothing in the house moved. A run counted as right when the model called the right tool and its spoken answer passed the checks. One request ("answer in Thai") was dropped: every model obeys it, so the language rule is enforced live instead.

Model (Ollama)	Where	Right of 34	Median answer
nemotron-3-super	cloud	33	7.89 s
gpt-oss 20B	cloud	32	1.40 s
kimi-k2.6	cloud	32	3.74 s
gemma4 31B	cloud	30	0.82 s
gpt-oss 20B	local, 13 GB	29	0.44 s
glm-4.7-flash	local, 19 GB	28	0.26 s
Qwen 2.5 14B (our old setup)	local, 9 GB	28	0.44 s
deepseek-v4-pro	cloud	27	1.69 s
minimax-m3	cloud	27	3.08 s
Qwen3 14B	local, 9.3 GB	26	0.44 s
gpt-oss 120B	cloud	23	0.93 s
Llama 3.1 8B	local, 4.9 GB	21	0.84 s
glm-5.3	cloud	18	1.40 s

The pick: gpt-oss 20B, on the local GPU. 29 of 34 at 0.44 seconds, and it uses its tools instead of describing them. The old model, Qwen 2.5 14B, would say "playing" with no tool call, which is why we switched. When I need the GPU for other work, its cloud twin (32 of 34, 1.4 s) takes over.

Which brain for your hardware:

You have	Use	What to expect
A GPU with 16 GB or more	gpt-oss 20B	The best local balance of right and fast
About 10 GB	Qwen3 14B	26 of 34, same speed
About 6 GB	Llama 3.1 8B	Every home-control test right, all news searches right, weak at small talk (it calls tools when it should just chat)
No GPU, or no AI wanted	Home Assistant's built-in commands	Lights, timers and music controls in about 0.1 s; no open-ended questions

Our settings. In Home Assistant's Ollama integration: model gpt-oss:20b, context window 64K tokens, thinking off, 20 messages of history, and control of Home Assistant on (the Assist API). We run three pipelines and switch the Dot between them:

• Local: Whisper (turbo) on the GPU, gpt-oss 20B, Piper (en_US-lessac-medium). "Prefer handling commands locally" is on, so fixed commands skip the model.
• Quick: Home Assistant's built-in commands only, with Whisper (base.en) on the CPU.
• Work: Whisper on the CPU, and the same model on Ollama's cloud, so the GPU is free.

A checker anyway. Claude Code and I built a small proxy that sits where Home Assistant expects Ollama. It passes everything through, but reads each final answer before it's spoken: wrong language, empty or runaway answers, news quoted without a search, a news question answered without searching. On a miss it retries once with a correction, then falls back to a plain line. It also keeps a week of notes as memory, has a guest mode that reads and keeps nothing, and forgets on request.

What leaves the house: audio never does. Text leaves only for web searches and, in work mode, answers from the cloud model.

Music Assistant is what makes "play something" work across speakers. It runs beside Home Assistant, pulls from your music services, and plays to almost anything: TECHO5 Dots show up as Sendspin players, Google speakers as Chromecast, and you can group speakers to play in sync.

1. Run the server. We use the container ghcr.io/music-assistant/server with host networking, so it can discover speakers.
2. Add a music provider in Music Assistant's settings. We use Spotify, which needs a Premium account. Add the player providers for your speakers too (Sendspin, Chromecast, AirPlay, DLNA), and a sync group if you want "everywhere".
3. Add the Music Assistant integration in Home Assistant. Each player becomes a media_player, and you get the music_assistant.play_media action, which takes a name ("Bon Iver", a song, a playlist) rather than an exact ID.

Give the model a play tool. Expose a script to Assist and the model can call it. Ours is a cleaned-up version of this:

play_music:
  alias: Play music
  description: >-
    Start music through Music Assistant on the Dot (the default) or another speaker.
    Use it only to start music: a song, artist, album, playlist or station. For a mood
    or genre, pick a fitting real artist and pass that. Call it once per request.
  mode: single
  fields:
    query:
      description: Song, artist, album, playlist or station name
      required: true
      selector:
        text:
  sequence:
    - action: music_assistant.play_media
      target:
        entity_id: media_player.your_dot
      data:
        media_id: "{{ query }}"

The description is the part that matters: it's the instructions the model reads. "Pick a fitting real artist" came straight out of the test, where every model searched the words "cooking dinner" instead.

Keep the quick controls out of the model. Pause, skip, resume and "what's playing" don't need an LLM. Custom sentences in config/custom_sentences/en/ map them to an intent script that acts on whichever speaker is playing, so they answer in about a tenth of a second. Anything starting with "play" goes to the model, which can ask when a request is unclear.

Beyond play. We also gave Marvin tools to list the household's playlists, say what's on one, add or remove the song that's playing, and build a new playlist on a theme. The builder grounds itself with a web search, fills out from Music Assistant's similar tracks, and saves to Music Assistant only, until someone asks in words to save it to Spotify.

A local model knows nothing after its training date, and it will happily guess. SearXNG is a self-hosted metasearch engine: it asks the big engines for you and returns the results without tracking, and its JSON output is easy for Home Assistant to read.

1. Run it, with JSON turned on. In SearXNG's settings.yml:

use_default_settings: true
server:
  limiter: false      # it's only answering your own house
search:
  formats:
    - html
    - json            # off by default; Home Assistant needs it

docker run -d --name searxng --restart unless-stopped \
  -p 8888:8080 -v ./searxng:/etc/searxng searxng/searxng:latest

2. Let Home Assistant call it, in configuration.yaml:

rest_command:
  searxng_search:
    url: "http://127.0.0.1:8888/search?q={{ query | urlencode }}&format=json&categories={{ category | default('general') }}&language=en-US"
    method: GET
    timeout: 20

3. Wrap it in a script the model can use, and expose the script to Assist:

web_search:
  alias: Search the web
  description: >-
    Search the web through a private SearXNG. Use it for news, current events,
    sports results, prices and facts you are not sure of; never guess at current
    information. Answer in at most three short sentences, name the outlet for each
    result, and never read out URLs. The query leaves the house, so never put
    personal details in it.
  mode: parallel
  fields:
    query:
      description: Short search query, like a person would type
      required: true
      selector:
        text:
  sequence:
    - action: rest_command.searxng_search
      data:
        query: "{{ query }}"
      response_variable: r
    - variables:
        found: >-
          {% set ns = namespace(out=[]) %}
          {% for x in (r.content.results if r.status == 200 else [])[:4] %}
          {% set ns.out = ns.out + [{'title': x.title, 'snippet': (x.content or '')[:260]}] %}
          {% endfor %}{{ {'query': query, 'results': ns.out} }}
    - stop: done
      response_variable: found

Four short results keep the answer fast and the context small. Ours also has a news mode that adds SearXNG's news category, limited to the past week. The search query is the one thing that leaves the house here, which is why the description tells the model never to put personal details in it. The checker from Step 5 catches a news answer given without a search.

Symptom	What it means	Fix
firebreak waits forever	A stock Dot never shows up as adb	Do the button gesture
Ring orange	The Dot booted normally	Unplug and retry the gesture
Ring stays solid blue	The cable or port isn't carrying data	Another cable, or a port on the computer itself
Preloader appears and vanishes in the kernel log	Normal; confirms the cable carries data	Nothing to fix
Ring flashes yellow after the install	Healthy, but no Wi-Fi	Check the SSID spelling, then run set-wifi.py
Dot stuck after a failed unlock step	Soft-bricked	See the XDA recovery thread for the 2nd-gen Dot

This is other people's excellent work, put together in one place:

• firebreak by bboe, built on R0rt1z2's amonet
• TECHO5 by HuskerMinion
• Home Assistant and its Assist voice pipeline
• Ollama, Whisper and Piper for the brain, ears and voice
• Music Assistant and SearXNG
• Related projects: echo-dot-assist (Dot 3 2018, Dot 2, Echo 2) and EchoMuse

Go check your drawers.