Siri, answered by your own machine

Step 1Pick the computer

Something in your house that is switched on when you want to ask Siri something and is on the same Wi-Fi as the iPad. Not the guest network. That is the whole requirement.

Only one part of this is heavy: the language model. It wants about 3.5 GB of memory to itself and does the actual thinking. Everything else is small. So judge a machine on whether it can hold and run the model.

WHISPER_MODEL=deepdml/faster-whisper-large-v3-turbo-ct2

The longer version of this comparison is further down, with timings and the trade-offs spelled out.

Step 2Put an AI assistant on it

This is the assistant that will do the installing. It runs on the computer from step 1 — not on your phone, and not in a browser tab, because it needs to actually run commands there.

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

hermes setup --portal

# macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# or with Homebrew
brew install --cask claude-code

npm install -g @openai/codex

# or with Homebrew
brew install --cask codex

Any assistant that can run commands on the machine will do — Gemini CLI, Copilot in a terminal, Cline, Aider. The prompt does not assume any particular one.

Step 3Set up the iPad

Two things on the device itself. The first is required — it is the tweak. The second is optional and lets the assistant finish the job instead of handing you homework.

Install the tweak — required

1. Open Cydia → Sources → Edit → Add, and enter:

   http://repo.theadipost.com/

   Plain http://, not https — an iOS 6 device cannot validate a modern certificate, so an https source refuses to load on exactly the devices this is for.
2. Install “iOS 6 Siri Remastered” from that source. It sits quietly until you give it a server address, so nothing changes yet.

Let the assistant in — optional

With SSH access the assistant checks the tweak installed properly, writes the server address into Settings for you, and reads the tweak's own log if something misbehaves. Without it, you type the address in by hand at the end — about six taps.

1. In Cydia, search for OpenSSH — the one from Telesphoreo — and install it. No respring needed; it starts listening straight away.
2. Find the iPad's address: Settings → Wi-Fi → the blue ⓘ beside your network. It looks like 192.168.1.42.
3. Log in from your computer. The username is root and the password is alpine — the same on every jailbroken device, so anyone else on your Wi-Fi could use it too. On a home network that is usually fine; on a shared or public one, passwd changes it.

   ssh root@YOUR-IPAD-IP

4. If ssh refuses to connect at all, that is expected on a modern computer: iOS 6's SSH is old enough that current clients reject its encryption by default. Add these and it connects:

   ssh -o HostKeyAlgorithms=+ssh-rsa     -o PubkeyAcceptedAlgorithms=+ssh-rsa     -o KexAlgorithms=+diffie-hellman-group14-sha1     root@YOUR-IPAD-IP

5. Tell the assistant the iPad's address when it asks. The prompt in step 4 knows what to do with it: it checks the tweak is installed, writes the server address into Settings for you, and reads the tweak's log if anything misbehaves.

Step 4Give it the prompt

Paste this into the assistant and answer its questions. It explains what it is about to do before each step, and asks before installing anything.

You are not expected to read this. It is written for the machine, and it is long because it spells out the things that otherwise go wrong. Press Copy and paste it — understanding it is optional.

I want you to set up the self-hosted server for "iOS 6 Siri Remastered", a
jailbreak tweak that brings Siri back to life on a jailbroken iPhone or iPad
running iOS 6. Apple retired the service the original Siri called, so the
tweak points the device at a small server on the local network instead. You
are running on the machine that is going to BE that server.

HOW I WANT YOU TO WORK
- Tell me in plain language what you are about to do before each step.
- Ask me before you install anything or change a system setting.
- Never run sudo without first telling me what the command does.
- If something fails, show me the real error instead of guessing at it.
- Assume I have never used Docker or Ollama. Explain, briefly, as you go.

WHAT WILL BE INSTALLED
1. Docker, including Docker Compose. macOS and Windows: Docker Desktop.
   Linux: docker.io plus docker-compose-plugin, or Docker's own repository.
2. Ollama, from https://ollama.com. It runs the language model, directly on
   this machine, NOT inside a container.
3. One language model: gemma4:e4b. That is what this software was built and
   tested on, so use it unless I ask for something else.
4. The server stack itself: two Docker containers. "mynah-server" is a
   FastAPI service on port 8807, which is the one the iPad talks to.
   "mynah-stt" is faster-whisper, which turns recorded speech into text.

STEPS

1. Look at this machine and tell me what it is: operating system, processor,
   how much RAM, and whether there is an NVIDIA GPU. Say plainly whether it
   is comfortable for this. If it is a Raspberry Pi or has under 8 GB of
   RAM, warn me that answers will take tens of seconds, and only then offer
   a smaller Ollama model as an alternative -- otherwise do not raise the
   subject of model choice at all.

2. Install Docker if it is not already there. Verify with:
       docker --version
       docker compose version
   Both must print a version. If the second one fails, Compose is missing;
   install docker-compose-plugin (some distributions call it
   docker-compose-v2). On Linux, add me to the docker group so I am not
   typing sudo constantly, and tell me I must log out and back in for that
   to take effect.

3. Install Ollama if it is not already there, then pull the model and list
   what is installed:
       ollama pull gemma4:e4b
       (this model needs Ollama 0.20.0 or newer -- if the pull is refused,
        update Ollama first, then retry)
       ollama list

   ON LINUX ONLY, and this catches everyone: Ollama listens on 127.0.0.1 by
   default, and a Docker container counts as a different machine, so the
   server will start, look healthy, and fail every single question. Fix it
   one of two ways -- either set Environment="OLLAMA_HOST=0.0.0.0" under
   [Service] via `sudo systemctl edit ollama`, then daemon-reload and
   restart it; or leave Ollama alone and put this machine's own LAN address
   into OLLAMA_URL in step 5. On macOS and Windows nothing needs changing,
   because Docker Desktop already resolves host.docker.internal.

4. Download and unpack the server bundle, into a folder I choose (ask me
   where; somewhere durable, not /tmp):
       curl -O http://repo.theadipost.com/ios6-siri-server.tar.gz
       tar xzf ios6-siri-server.tar.gz
       cd ios6-siri-server
       cp .env.example .env
   That copy is required, not optional: docker compose reads .env and stops
   with an error if it is missing. Read .env.example before editing -- every
   setting is documented inline in it -- and note the docs/ folder beside it,
   which has the long-form SERVER-SETUP, API-KEYS and TROUBLESHOOTING
   documents if you need more detail than I have given you here.

5. Edit .env. It is plain NAME=value lines, no quotes, no spaces around the
   equals sign. This file is the ONLY place the model and the API keys are
   configured -- there are no such fields on the iPad. The settings that
   matter:
       MYNAH_MODEL=      gemma4:e4b, exactly as `ollama list` prints it
       OLLAMA_URL=       http://host.docker.internal:11434 is correct on
                         macOS and Windows; on Linux see step 3
       SERVER_PORT=      8807. Change it only if something already has that
                         port, and then tell me the new number
       DEFAULT_UNITS=    imperial (Fahrenheit, miles) or metric
       DEFAULT_CITY=     ask me for my town. This is what "the weather" and
                         "restaurants near me" mean when I do not name a
                         place. DEFAULT_LAT and DEFAULT_LON are optional and
                         make "near me" sharper
       PUBLIC_BASE=      LEAVE THIS BLANK. The server works out its own
                         address from the device's request. Setting it wrong
                         is a common way to break images
       TMDB_API_KEY=     optional. Unlocks the movies card
       WOLFRAM_APPID=    optional. Richer answer card, with plots
   Ask me once, in passing, whether I already have a TMDB key or a Wolfram
   AppID, and put them in if I do. Do not send me off to sign up for either:
   both are free but entirely optional, and six of the seven cards need no
   key at all. Either can be added months later with a restart.

6. Start it, and warn me first that the very first start downloads container
   images and a speech model and can take several minutes:
       ./start.sh
       ./start.sh logs

    Use ./start.sh, NOT `docker compose up -d`. On a machine with a usable
    NVIDIA GPU, plain compose silently leaves the GPU out: the speech
    container comes back on the processor, everything gets slower, and
    nothing reports a fault. start.sh detects the GPU and picks the right
    override. If you ever need to check which image is actually running:
       docker inspect -f '{{.Config.Image}}' mynah-stt
    A name ending -cuda is the GPU image; -cpu is not.
   Watch the log until it settles, then stop watching with Ctrl-C.

7. Verify, in this order, and show me the output of each one:
       curl http://localhost:8807/healthz
   Expect JSON containing "status":"ok". If it says "stt":"unreachable", the
   speech container is probably still downloading its model -- wait a few
   minutes and check `./start.sh logs`.

       curl "http://localhost:8807/debug?q=weather+in+London"
       curl "http://localhost:8807/debug?q=AAPL+stock"
   /debug runs the complete pipeline except transcription, from typed text.
   Each should return JSON with a "speak" line and a "cards" array. If these
   two work, then everything except the microphone works, and anything still
   wrong afterwards is between the iPad and this machine. This is the single
   most useful diagnostic on the whole server; remember it.

       curl http://localhost:8807/config
   Reports what the server actually believes it is configured with. Any keys
   come back redacted to their last four characters.

8. Find this machine's address on the local network and tell it to me
   clearly:
       macOS:   ipconfig getifaddr en0
       Linux:   hostname -I
       Windows: ipconfig, and read the IPv4 Address of the active adapter
   A local address starts 192.168., or 10., or 172.16 through 172.31.
   Anything starting 127. is the machine talking to itself and is wrong.
   Then make sure the port is reachable from outside this machine: if
   there is a firewall (ufw on Linux, the application firewall on macOS,
   Defender on Windows), open TCP 8807 and tell me exactly what you changed.

9. Make it survive a reboot, and tell me what you did:
       - Docker starting at boot or login (systemctl enable docker on Linux;
         the "start at login" setting in Docker Desktop otherwise)
       - Ollama the same (systemctl enable ollama on Linux)
       - Tell me to give this machine a static address or a DHCP reservation
         in my router, because if the number changes Siri stops working one
         morning for no visible reason
       - If this is a laptop, tell me how to stop it sleeping. Siri only
         works while this machine is awake.

10. THE DEVICE, IF I GIVE YOU ITS ADDRESS. Ask me whether I installed
    OpenSSH on the iPad and, if I did, ask for its address. With it you can
    finish the job instead of handing it back to me. Without it, skip to 11.

    An iOS 6 device runs an SSH server old enough that a current client
    refuses it outright, so you need these flags every time -- if you get
    "no matching host key type" or "no matching key exchange method", this
    is why, not a wrong password:
       ssh -o HostKeyAlgorithms=+ssh-rsa            -o PubkeyAcceptedAlgorithms=+ssh-rsa            -o KexAlgorithms=+diffie-hellman-group14-sha1            root@
    The login is root / alpine unless I tell you I changed it.

    With that working, do these and tell me the result of each:
    a. Check the tweak is installed:
          ls /Library/MobileSubstrate/DynamicLibraries/SiriRemastered.dylib
       If it is missing, tell me to add http://repo.theadipost.com/ as a
       source in Cydia and install "iOS 6 Siri Remastered", then re-check.
    b. Write the server address into the tweak's settings, so I do not have
       to type it on a touchscreen. The file is
          /var/mobile/Library/Preferences/com.aditya.ios6siriremastered.plist
       Set ServerURL to http://:8807 and Enabled
       to true. It is a binary plist; if the device has no usable plist
       editor, copy it to this computer, edit it here, copy it back, and
       chown it back to mobile:mobile. Do NOT invent other keys.
    c. Reload the daemon so it picks the settings up:
          killall assistantd
    d. Tell me to hold the Home button and ask "what is the weather".
       If something is wrong, create /var/mobile/Media/sr-log.txt on the
       device, have me ask again, then read that file -- the tweak writes
       what it did into it. Delete the file when we are finished.

11. Finish by printing, as the last thing you say, exactly what I have to
    type on the iPad. If you did step 10 there is nothing left to type and
    you should say so. Otherwise there is only one thing, and this is it:
        Settings -> iOS 6 Siri Remastered -> Server URL
        http://<this machine's LAN address>:8807
    Plain http://, never https:// -- an iOS 6 device cannot validate a
    modern certificate and https will never work. Include the port. No
    trailing slash. That is the whole of the device setup: the rest of that
    pane is an Enabled switch, my town and units, and a switch per card,
    all of which have sensible defaults.

THINGS TO KNOW WHILE YOU WORK
- The iPad and this machine must be on the SAME network. Guest Wi-Fi and
  separate VLANs exist specifically to stop devices reaching each other, and
  they will silently break this.
- Do not port-forward 8807 and do not expose this to the internet. The
  server has no authentication by design, and iOS 6 cannot encrypt anything,
  so everything between the device and the server is plain text on the LAN.
- The server has to be running for Siri to answer. There is no cloud
  fallback and no queue; if the machine is off, Siri is silent.
- The first question asked after any restart is slow, because the language
  model and the speech model both load into memory on first use. Ten seconds
  to a minute is normal and is not a fault.

There is a second prompt for when something is wrong later, on the prompts page.

sudo apt update
sudo apt install -y docker.io docker-compose-plugin
sudo systemctl enable --now docker

sudo usermod -aG docker "$USER"

docker --version
docker compose version

ollama pull gemma4:e4b

ollama list

sudo systemctl edit ollama.service

[Service]
Environment="OLLAMA_HOST=0.0.0.0"

sudo systemctl daemon-reload
sudo systemctl restart ollama

curl -O http://repo.theadipost.com/ios6-siri-server.tar.gz
tar xzf ios6-siri-server.tar.gz
cd ios6-siri-server

cp .env.example .env

# The model you pulled in step 2, exactly as `ollama list` prints it.
MYNAH_MODEL=gemma4:e4b

# Where Ollama is. Correct as shipped on macOS and Windows.
# On Linux, this works once Ollama listens on 0.0.0.0 (see step 2);
# if you would rather not touch systemd, put the machine's own LAN
# address here instead: OLLAMA_URL=http://192.168.1.50:11434
OLLAMA_URL=http://host.docker.internal:11434

DEFAULT_CITY=Springfield, IL
DEFAULT_LAT=39.7817
DEFAULT_LON=-89.6501

# imperial (Fahrenheit, miles) or metric (Celsius, kilometres)
DEFAULT_UNITS=imperial

TMDB_API_KEY=4f1c2b9a7d3e5f8091a2b3c4d5e6f7a8
WOLFRAM_APPID=XXXXXX-XXXXXXXXXX
SEARXNG_URL=http://host.docker.internal:8088

./start.sh

docker inspect -f '{.Config.Image}' mynah-stt

docker compose logs -f mynah-server

./start.sh                             # start, or restart after a change
./start.sh ps                          # is everything running?
./start.sh logs                        # follow the log
./start.sh down                        # stop everything
./start.sh --build                     # after editing the server itself

docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d

curl http://localhost:8807/healthz

{"status":"ok","stt":"Systran/faster-distil-whisper-small.en",
 "llm":"gemma4:e4b","search":"duckduckgo","movies":"needs TMDB_API_KEY"}

curl "http://localhost:8807/debug?q=weather+in+London"
curl "http://localhost:8807/debug?q=AAPL+stock"
curl "http://localhost:8807/debug?q=who+won+the+Yankees+game+last+night"

curl http://localhost:8807/config

ipconfig getifaddr en0

hostname -I

ipconfig

curl http://192.168.1.50:8807/healthz

http://repo.theadipost.com/

services:
  searxng:
    image: searxng/searxng:latest
    container_name: searxng
    restart: unless-stopped
    ports:
      - "8088:8080"
    volumes:
      - ./searxng:/etc/searxng

search:
  formats:
    - html
    - json

SEARXNG_URL=http://host.docker.internal:8088

What works with no keys at all

Worth saying plainly, because it is unusual: you can install this, run the server, and never sign up for anything. Six of the seven cards need no key, ever.

General conversation and arithmetic need nothing either: those are the local model and the server’s own calculator.

What it cannot do

Some things are missing because no free source for them exists, and a card that invents them would be worse than a card that omits them.

• No movie showtimes. The movies card knows what is playing, not when or where. Showtime data is licensed and paid, everywhere.
• No live restaurant availability. OpenTable’s API is partner-only, so the button on a restaurant opens an OpenTable search for that place. It does not show open tables and cannot book one.
• No ratings, review counts or price levels on restaurants. OpenStreetMap does not carry them, so the card leaves those rows out rather than making them up.
• Sports coverage is a list, not everything. Basketball, American football, baseball, hockey, MLS and the big European football leagues are covered. Cricket, Formula 1, tennis, rugby and golf deliberately come back as a web answer instead of a score card.
• Statistics that do not exist show a dash. Apple’s own athlete card renders an empty cell as “—”, and so does this one.

Only if you need itReaching the device from your computer

A normal installation needs none of this. Add the source in Cydia, install the tweak, type your server address into Settings. Done.

That is worth saying first, because the rest of this section is a list of developer tools and it would be easy to conclude the tweak is harder to install than it is. It is not. Everything on the device happens in Cydia and in Settings, with your fingers, and most people will never need what follows.

But two things change that. If you want to hand the setup to an AI assistant (step 2), or if something goes wrong and you need to see what the device is actually doing, then the AI — or you — needs a way to reach the iPad from a computer. Sorting that out takes ten minutes and is worth doing before you need it, because the moment you need it is the moment the device is misbehaving.

The two Cydia packages the tweak sits on

Both are declared as dependencies, so Cydia installs them for you without asking and you would normally never think about them. It is still worth knowing what they are:

• MobileSubstrate, which Cydia may list as Cydia Substrate — the code-injection framework every iOS 6 tweak is built on. It is what reads SiriRemastered.plist, sees that the tweak wants assistantd, and loads the tweak's library into that process when it starts. Practically every jailbreak installs it as part of itself.
• PreferenceLoader — the package that lets third-party software put a row in Settings. Without it the tweak is still installed and still loaded, but there is no pane to type a server address into, so it stays inert forever.

SSH, over Wi-Fi

Install OpenSSH from Cydia — it is in the default sources, under Networking. It starts by itself and keeps running. That gives you a command line on the iPad from any computer on the same network.

The login is root / alpine, which is the same on every jailbroken device. passwd changes it if you are on a network you do not control.

To find the device's address, on the iPad open Settings → Wi-Fi, tap the blue arrow beside the network you are on, and read IP Address. It will look like 192.168.1.42. That number can change when the router reboots, exactly as the server's can, so pin it in the router if you find yourself looking it up often.

ssh -o HostKeyAlgorithms=+ssh-rsa     -o PubkeyAcceptedAlgorithms=+ssh-rsa     -o KexAlgorithms=+diffie-hellman-group14-sha1     [email protected]

Host ipad
  HostName 192.168.1.42
  User root
  HostKeyAlgorithms +ssh-rsa
  PubkeyAcceptedAlgorithms +ssh-rsa
  KexAlgorithms +diffie-hellman-group14-sha1

Or over the cable, with libimobiledevice

If you would rather not run an SSH server on the device, or the Wi-Fi is the thing that is broken, the USB cable works. libimobiledevice is a set of open-source tools that speak to an iOS device over USB, and the log reader in particular needs nothing installed on the device at all:

brew install libimobiledevice            # macOS
sudo apt install libimobiledevice-utils  # Debian / Ubuntu

Three of its tools matter here:

• ideviceinfo — prints the device's details. If this works, the cable and the pairing are fine, which rules a lot out in one command.
• idevicesyslog — the device's entire system log, live. This is the single most useful thing in this section: it is how you see what the tweak is doing, and it works whether or not SSH is installed. Filter it to the lines that matter: idevicesyslog | grep SiriRemastered.
• iproxy — forwards a port on your computer to a port on the device, which gives you SSH over the cable: iproxy 2222 22 in one window, then ssh -p 2222 root@localhost in another (with the old-algorithm flags above).

Editing files on the device itself

If you have no computer to hand, install iFile from Cydia. It is a full file manager with a plist and text editor built in, so you can read the tweak's settings file or move a file out of the way without SSH at all. Its modern successor is Filza, which you will see recommended everywhere — but Filza wants a newer iOS than this, so on iOS 6 it is iFile you want.

When something is wrong

Start here. Almost everything that goes wrong is one of the six things below, and five of them are the network. If none of it fits, or you want to hand the problem to an assistant, the technical page documents the whole machine in enough detail to debug it properly.

docker compose exec mynah-server python -c "import httpx,os;print(httpx.get(os.environ.get('OLLAMA_URL','http://host.docker.internal:11434')+'/api/tags',timeout=5).status_code)"

docker compose logs -f mynah-server           # follow live, Ctrl-C to stop
docker compose logs --tail 200 mynah-server   # the last 200 lines
docker compose ps                             # running? or restarting in a loop?

If it is stranger than that

There is a second page, and it is the interesting one. It documents the whole thing properly: what happens between the home button and the card on the screen, every file the package puts on the device, exactly which functions the tweak hooks and why, every endpoint the server answers, how to read the device's log line by line, what each failure mode actually means, and what to do in the worst case — a card that takes SpringBoard down with it.

It is written to be pasted, whole, into an AI assistant, along with whatever went wrong. Give it that page and it can reason about this system instead of guessing at it.

It is also simply worth reading if you want to know how a 2012 assistant gets talked into answering from a machine in your kitchen.

Where your words go

The recording of your voice goes to one place: the server on your own network. It is transcribed there, by a model on your own disk. Nothing is sent to Apple — the tweak never opens a connection to Apple's assistant service at all, which is the reason it still works when that service does not.

What does leave your network is the lookup each answer needs, made by the server, on your behalf, and nothing more:

• Open-Meteo for the forecast.
• ESPN for scores and player statistics.
• Yahoo Finance for quotes.
• OpenStreetMap (Overpass and Nominatim), and Wikidata/Wikimedia for restaurant photos.
• DuckDuckGo, or your own SearXNG instead, for open questions.
• TMDB and WolframAlpha — only if you added those keys, and only for those cards.

The language model is Ollama, on your machine. The speech model is in a container, on your machine. Neither ever contacts anyone.