Termux AI Bootstrap

Termux AI Bootstrap

 Turn your Android phone into a
 production AI agent host.
 Personalized step-by-step setup — no           guesswork, no broken dependencies.
 What It Does

 Ask me to help you set up your phone
 and I'll walk you through:

 - System setup — exact pkg install
 commands for your device
 - Python ML stack — the two-tier
 strategy that actually works on
 aarch64                                        - llama.cpp build — cmake flags tuned
 for your phone's CPU                           - Model download — which model fits
 your RAM and runs at usable speed              - Hermes Agent — install and
 configure for cloud or local                   inference
 - Background hardening — keep your             agents alive when Android wants them           dead
 Sample Interaction
                                                User: I want to run a coding                   assistant on my phone. 6GB RAM, fresh
 Termux.

 Agent:                                         > 6GB RAM means you can run up to ~2B
 parameter models at Q4. For coding, I
 recommend DeepSeek-Coder-1.3B or
 Qwen2.5-Coder-1.5B. Here's your
 setup:
 >
 > Step 1: Base packages
 > bash
 > pkg update && pkg upgrade -y                 > pkg install -y git curl cmake
 binutils make clang python python-pip          nodejs-lts
 >                                              >
 > Step 2: Python ML stack (never pip
 install scipy — use pkg)
 > bash
 > pkg install -y python-numpy                  python-scipy
 > pip install httpx pydantic                   >
 >
 > Step 3: llama.cpp
 > bash
 > cd ~/storage/downloads
 > git clone                                    https://github.com/ggerganov/llama.cp
 p                                              > cd llama.cpp && mkdir build && cd
 build                                          > cmake .. -DCMAKE_BUILD_TYPE=Release
 -DLLAMA_BLAS=OFF                               > make -j$(nproc)
 >                                              >
 > Step 4: Model (DeepSeek-Coder-1.3B,
 ~800MB, ~20 tok/s)
 > bash                                         > mkdir -p ~/models
 > curl -L -o                                   ~/models/deepseek-coder-1.3b-q4.gguf
 \                                              >
 "https://huggingface.co/bartowski/Dee          pSeek-Coder-1.3B-GGUF/resolve/main/De          epSeek-Coder-1.3B-Q4_K_M.gguf"
 >
 What Makes This Different                 
 Every command came from real builds
 on aarch64. Not docs. Not
 StackOverflow. Real devices, real
 failures, real workarounds.

 Things it catches that generic AI
 misses:
 - pip install scipy fails on aarch64           — use pkg install python-scipy                 - Whisper and faster-whisper do not
 compile on Android
 - Play Store Termux is outdated — use
 F-Droid
 - Android kills background agents
 when screen turns off
 - 8GB RAM = ~4-5GB usable for models           Who's It For                                                                                  - Developers who want a local AI               agent without cloud bills
 - Tinkerers who want their phone to
 do more
 - Anyone who followed tutorials and
 hit broken dependencies                        - Privacy-conscious users who want
 everything running locally
 What You'll Have When Done                
 - llama.cpp compiled and running on            your phone                                     - A working local LLM matched to your
 RAM
 - Hermes Agent configured (cloud or
 local)                                         - Background processes that survive
 screen-off                                     - Verified everything works
 Supported                                 
 Any Android phone with Termux
 (F-Droid version). 4GB-12GB devices            tested. Not for iOS or Play Store
 Termux.