A ringing public telephone. A hardline. One last connection to the real world.
They took every last one.
ZION-CELLPHONE. THE LAST HARDLINE.
ZION-CELLPHONE is a Solana Seeker rebuilt around a simple idea: the most powerful thing your phone can do is belong to you.
An AI that thinks and works on the device. A real Linux command-line environment. Android system controls. A stripped-down, terminal-first home screen. Privacy changes you can audit and reverse. And a hardware wallet left alone.
No cloud is required to run the local model. No AI account. No API key. No bootloader unlock.
Your phone. Your terminal. Your AI. Your rules.
It runs on a real Solana Seeker, tested on Android 16 (security patch 2026-09-05).
| On-device result | Measured outcome |
|---|---|
| Local AI generation | 5.4 tokens/second with the default model |
| First word from OPERATOR after the model restarts | 1.0 s instead of 43 s (the start of the conversation is kept on disk) |
| Model memory after 10 idle minutes | 3.5 GB → 0.2 GB (it reloads on the next question) |
| Debian 13 inside the terminal, no root | installed in 9 s |
| Tool-call correctness | 30/30 cases on the benchmark |
| Real agent tasks | 4/4 completed in full-device validation |
| Outbound connections from the terminal app during the offline inference test | 0 observed |
| Local model loaded in memory | 3.6 GB RAM |
| Bootloader | Locked |
| Root | Not required |
These are results from the tested device and specific test procedures—not claims that the whole phone is invisible, anonymous, or permanently disconnected. Raw artifacts and scripts live in tests/; every figure, with its date and file, is in docs/RESULTS.md.
Most AI products make your phone a remote control for someone else's computer.
ZION-CELLPHONE puts the model on your phone.
The default is HauhauCS Qwen3.5-4B Uncensored Q4_K_M, running through llama.cpp, compiled for the Seeker's CPU. A standard Qwen3.5-4B Q4_0 build is available as a fallback.
Turn off Wi-Fi. Start the model. Keep working.
The project is more than a local chat interface: the agent can use tools, inspect errors, revise its own code, and carry a task through to completion—subject to explicit permissions.
OPERATOR is the local AI agent, launched from the terminal with ai. It is a project identity, not a separate cloud service.
It can:
- Write, run, debug, and fix Python and shell scripts.
- Create and edit permitted files, inspect output, and retry after ordinary programming errors.
- Use git, including creating a commit.
- Read a CSV, calculate results, and explain what it did.
- Search your permitted folders, and diagnose the phone: battery, heat, memory, storage, crashes, logs.
- Manage apps (stop, disable, enable) with a journal and an undo, never the Seed Vault.
- Work with Android services through Shizuku, within ADB shell-level permissions.
- Save every step: a task cut short by Android can be resumed (
/history,/resume). - Continue saved conversations and use your own system instructions.
On the phone, the full validation included four practical programming tasks. The narrower model comparison benchmark separately recorded 3/3 agent tasks.
The difference: you do not merely ask your phone how to do something. You can ask it to do the work, review the proposed action, and decide whether it may proceed.
The terminal environment is built around Termux:
bash · zsh · Python 3.13 · Node.js 24
git · clang · cmake · OpenSSH
nmap · dig · tmux · ripgrep · jq
Code compiled on the phone runs on the phone. The terminal is not a decorative interface to a cloud service.
Need a package Termux does not have? A full Debian 13 runs inside it through proot-distro, without root: installed in 9 seconds, Python and git added in under two minutes.
Shizuku gives supported tools access through the Android shell identity (uid 2000), similar to ADB—not root.
Depending on the command and Android restrictions, that means access to system diagnostics, settings, package information, notifications, Wi-Fi state, screenshots, APK installation, and more. Battery, volume, and toast utilities are also available from Termux.
This does not bypass Android's security model. Shizuku must be started again after reboot using the documented procedure.
No infinite feed. No attention economy. No unnecessary decoration.
The custom launcher replaces the usual app grid with a black, monospace command-center aesthetic:
ZION ZERO
─────────────────────────────────
> terminal
> ai
> applications
> privacy
> radio
> system
> seed-vault
─────────────────────────────────
The launcher is written in Java and built without Gradle. A black wallpaper option covers the home and lock screens, with the previous wallpaper backed up.
This is a reversible reduction of Google and factory software, not a replacement ROM.
The documented configuration:
- Disables 35 unused Google and factory packages in the first group.
- Disables another 23 Google applications and replaces selected functions with open-source alternatives from F-Droid.
- Revokes 16 permissions from Play Services.
- Uses Private DNS (AdGuard DNS) to block known trackers: the four tested on the phone are blocked.
- Journals changes so they can be reversed with a restore script.
- Checks Android and Seed Vault health after modifications.
Important: Google Play Services, the services framework, and the Play Store remain for security updates and wallet-related functionality. Google may still observe activity. ZION-CELLPHONE does not claim to make the device untraceable.
See debloat/PLAN.md for the actual trade-offs, docs/PRIVACY.md for who still talks to whom, and docs/ROOT.md for what works without root and what would need it.
The Solana Seeker includes a Seed Vault. ZION-CELLPHONE does not flash a new OS, unlock the bootloader, or deliberately modify wallet components.
Before and after changes, scripts check the wallet's package baseline, signatures, secure-hardware services, and verified boot state. Additional manual checks are documented in tests/SEEDVAULT_MANUAL.md.
A passing baseline check is not a guarantee of wallet security. Do not entrust private keys or seed phrases to the local model.
The tool layer separates permissions into four levels:
| Level | Purpose |
|---|---|
| Read | Inspect approved data |
| Write | Change approved files |
| Execute | Run commands and code |
| Sensitive | Perform higher-risk operations |
Actions that execute code, reach the network, or affect the system request approval. A refusal is final for that action: the agent must not retry through a workaround.
Built-in file tools exclude hidden files such as .ssh, .env, and .git. Seed Vault packages are protected even from an approved request.
The model is not a security boundary. Permission enforcement belongs in the surrounding code, and the repository includes tests for attempted bypasses and prompt injection.
The offline comparison uses the same questions and seeds, a 4,096-token context, and the benchmark in tests/bench_agent.py. Results are under tests/phone/.
| Measurement | HauhauCS Qwen3.5-4B Uncensored Q4_K_M (default) | Qwen3.5-4B Q4_0 (fallback) |
|---|---|---|
| Generation speed | 5.4 tok/s | 5.8 tok/s |
| Start → end of repeated-question test | 5.4 → 5.6 tok/s | 5.9 → 6.0 tok/s |
| Prompt processing | 20 tok/s | 24 tok/s |
| Cold first token, 1,153-token agent prompt | 45 s | 35 s |
| Subsequent first-token latency, median | 2.6 s | 2.4 s |
RAM peak, model read into memory as OPERATOR runs it (--load-mode none, round6b.log) |
3.6 GB | 3.5 GB |
| Correct tool calls, 15 cases × 2 seeds | 30/30 | 30/30 |
| Python unit tests, 6 functions × 2 | 12/12 | 11/12 |
| Agent benchmark tasks: script, fix, commit | 3/3 | 3/3 |
| 12-turn conversation | 12/12 | 12/12 |
| Observed outbound terminal-app connections | 0 | 0 |
| Peak CPU / skin temperature | 83.5 / 48.5 °C | 81.5 / 44.6 °C |
The default model was chosen by tests/bench_decide.py using a rule defined before measurements were taken, rather than choosing a favorite result afterward.
A separate on-device validation with tests/termux-round7.sh and tests/termux-round7b.sh checked startup without Wi-Fi, approvals, Seed Vault protections, four programming tasks, long conversation handling, fallback selection, network silence with a positive control, dashboard behavior, and Android health.
Prompts are deliberately in French: this tests how the model performs in the language actually used by its owner.
Newer measurements (first-token cache, Linux, speed by thread layout, other models, battery) are in docs/RESULTS.md.
After running the Termux setup and installing launcher version 1.5.3:
- Open Termux normally: a new interactive session opens OPERATOR, the local AI agent.
/exitreturns to the shell. - Tap
terminalin the ZION ZERO launcher: a regular interactive shell, without the AI. - Tap
ai: OPERATOR. Long pressai: its status,open operatoranddashboard.
The launcher starts these through Termux's own shortcuts (~/.shortcuts/terminal, operator, dashboard, installed by the setup). The first time, Termux shows its list once: tap the same name. The Google Play build of Termux has no RUN_COMMAND, so the launcher needs neither that permission nor allow-external-apps.
To stop OPERATOR from opening on every new session: touch ~/.config/zion/autostart-off (the ai button still opens it; rm the file to turn it back on). To remove the whole shell integration: bash terminal/zion-shell-setup.sh --remove, which gives your ~/.bashrc and ~/.zshrc back as they were.
From Termux, after installation:
| Command | Action |
|---|---|
ai |
Launch the agent with tools |
ai --chat |
Talk to the model without tools |
ai "your task" |
Run a single task |
ai -f file "question" |
Ask about a file |
command | ai "question" |
Pipe terminal output to the model |
ai --start |
Load the local model |
ai --stop |
Stop it and release roughly 3.5 GB RAM |
ai --status |
Inspect agent/model status |
ai --resume |
Resume the saved conversation |
/history, /resume N |
Inside OPERATOR: list the saved chats, reopen one |
seeker-ai --models |
Show available models |
seeker-ai --use NAME |
Change the default |
seeker-ai --model NAME |
Run with a selected model once |
dash |
Open the live system dashboard |
The dashboard displays battery, temperature, memory, agent state, processes, per-app network connections, and sensitive permissions.
Configure tools, allowed folders, history, and other settings in:
~/.config/seeker-ai/config.json~/.local/share/seeker-ai/~/.config/seeker-ai/system.md— your own instructions to the agent
Requirements: a Solana Seeker, a Mac or Linux computer with adb, and USB debugging enabled. Tested on Android 16 with the 2026-09-05 security patch; other versions are not verified here.
1. Identify your device.
adb devices
export SEEKER_SERIAL=<serial-shown-by-adb>2. Audit before changing anything. Run the read-only collector and record the Seed Vault baseline.
bash audit/collect.sh
bash tests/seedvault_check.sh baseline3. Review the changes before debloating. Read debloat/PLAN.md. The documented groups are applied in sequence and stop on failed health checks.
bash debloat/apply.sh G1
bash debloat/g2lite.sh4. Install verified applications as needed. The F-Droid fetcher checks the signed index, pinned key, APK SHA-256, and APK signer.
python3 -I apk-manager/fdroid_fetch.py <package> ...5. Install the terminal environment. In the tested setup, Termux comes from Google Play because Play Protect blocks the F-Droid version. From the repository root on your computer:
adb push . /sdcard/Download/seeker-zeroThen, in Termux on the phone:
bash /sdcard/Download/seeker-zero/terminal/termux-setup.shThis installs the terminal tools, agent, dashboard, and both local model files, with SHA-256 verification. It takes llama.cpp from Termux's llama-cpp package; the measurements here used a native build made on the phone (llama.cpp 9e9619d, tests/termux-round3.sh then tests/termux-round5.sh), set as llama_server in ~/.config/seeker-ai/config.json. If an older llama-server refuses an option, OPERATOR says which one.
6. Connect Shizuku. In the Shizuku app, use “Use Shizuku in terminal apps” → “Export files” into terminal/shizuku/, then run terminal/seeker-setup-shizuku in Termux. After a reboot, use terminal/start-shizuku.sh from the computer as documented.
7. Install the launcher. Build it with launcher/build.sh, install the generated APK, and select it as the home app. The launcher offers system → wallpaper: black, preserving the previous wallpaper.
8. Restore when needed. The project tracks reversible changes:
bash restore/restore.sh
# Or restore one group:
bash restore/restore.sh <group>
bash tests/seedvault_check.shReview the script documentation before executing it. Test the Seed Vault manually as well: tests/SEEDVAULT_MANUAL.md.
On the computer:
python3 ai/test_seeker_ai.py
python3 terminal/test_zion_shell.py
python3 terminal/test_seeker_dash.py
python3 tests/eval_fr.py --selftest
python3 tests/bench_decide.py --test
bash tests/test_bench_dry.sh
bash tests/test_bench_power_dry.sh
bash tests/test_chain_dry.sh
bash tests/test_journal_ops.shThe agent test suite covers 51 scenarios, including approval boundaries, protected folders, quote-based bypass attempts, git reads that reach outside a repository, escape sequences hidden in tool names, a remote model server off the phone, interrupted sessions, command injection through notifications, context handling, stopping during a task, a second session opened while the model loads, and a long job (the benchmarks) holding the memory: OPERATOR then pauses instead of loading a second model. It passed on the phone as well (Termux's Python, 13/11/2026). terminal/test_zion_shell.py opens real bash and zsh terminals to check when OPERATOR may start, and that removing the integration gives your shell files back byte for byte.
On the phone:
bash tests/termux-round6.sh
bash tests/termux-round7.sh
bash tests/termux-chain.sh # unattended: speed sweep, French eval and agent bench of every model, battery test
bash tests/health.sh
bash tests/seedvault_check.shThe tests are provided for inspection and reproduction. They do not establish comprehensive security or general reliability on every device.
| Path | Purpose |
|---|---|
ai/ |
Python agent, local model integration, installer, and tests |
terminal/ |
Termux setup, dashboard, and Shizuku utilities |
launcher/ |
Terminal-first Android launcher |
debloat/ |
Documented, health-checked reduction of unwanted packages |
restore/ |
Reversal of journaled changes |
apk-manager/ |
Verified F-Droid APK acquisition |
audit/ |
Read-only device inspection |
privacy/ |
Network/privacy inspection utilities |
tests/ |
Benchmarks, validation, raw results, wallet health checks |
docs/ |
Measured results, privacy audit, root and no-root |
adb.sh |
ADB helper |
You should be able to open a terminal on your own phone, run an AI on your own hardware, inspect what it can access, deny what it cannot, and shut the whole thing off.
That does not make the device magically private. It makes the architecture visible, local, and controllable.
Use the security and network tools only on devices and systems you own or have permission to assess.
They removed the payphones.
We built our own hardline.
The last hardline. The new way out.
