Aishe CLI

Voice-first AI assistant for your terminal.

Aishe CLI is a standalone command-line AI assistant with voice input/output, persistent memory, chat threads, tool calling, and live voice conversation — all from your terminal.

curl -fsSL https://raw.githubusercontent.com/astromanish/aishe-cli/main/setup.sh | bash
aishe status
aishe chat "Hello!"
aishe live

Features

• Chat — one-shot and streaming chat with tool calling (calculator, timezone, memory, and more)
• Live Voice — press Enter, speak, get a spoken response back. Full duplex: record → STT → LLM → TTS → play
• REPL — continuous text chat mode with thread management
• Memory — persistent personal memory. The AI remembers facts about you across sessions
• Threads — multiple conversation threads with history
• Voice I/O — transcribe audio files, synthesize text to speech
• Pet — your aishe blob. A real-time terminal pet that reflects your activity (idle, thinking, listening, speaking, error). Grows with usage, unlocks milestones, has persistent state.
• Ollama Integration — list, pull, and manage models
• Intent Lab — log and analyze user intent classification stats
• Diagnostics — comprehensive system check with roundtrip tests
• Export — export memory and threads to Markdown/CSV
• Search — full-text search across all threads and memory
• Shell Completions — bash, zsh, and fish

Quick Install

# One-command setup (auto-clones to ~/.local/share/aishe-cli)
curl -fsSL https://raw.githubusercontent.com/astromanish/aishe-cli/main/setup.sh | bash

# Or clone first, then run setup:
git clone https://github.com/astromanish/aishe-cli.git ~/aishe-cli
cd ~/aishe-cli
bash setup.sh

Requirements

• Python 3.9+
• requests + pyyaml (auto-installed by setup script)
• ffmpeg — for mic recording
  • macOS: brew install ffmpeg
  • Ubuntu/Debian: sudo apt-get install ffmpeg
  • Windows: download from https://ffmpeg.org/download.html
• Running services:
  • Ollama on :11434 (local or cloud models) — auto-installed
  • DeepAgent sidecar on :8765 — auto-installed
  • STT (faster-whisper) on :5093 — opt-in via AISHE_INSTALL_VOICE=1
  • TTS (Kokoro ONNX) on :8766 — opt-in via AISHE_INSTALL_VOICE=1

Voice setup

Voice sidecars (STT + TTS) are opt-in to keep the default install light (~50 MB). Enable them on first install or any time afterwards:

# First install with voice
curl -fsSL https://raw.githubusercontent.com/astromanish/aishe-cli/main/setup.sh | AISHE_INSTALL_VOICE=1 bash

# Or add voice to an existing install (re-run setup)
cd ~/.local/share/aishe-cli && AISHE_INSTALL_VOICE=1 bash setup.sh

This downloads:

• Kokoro TTS model (~350 MB) — high-quality multi-voice TTS, 54 voices, CPU-runnable
• faster-whisper STT model — auto-downloaded on first request, default small (~460 MB)

After install, voice works out of the box:

aishe voice speak "Hello, I'm Aishe" -V F4   # synthesize
aishe voice transcribe recording.wav         # transcribe
aishe live                                   # full voice conversation

The STT uses OpenAI Whisper via faster-whisper (CTranslate2 port, CPU or GPU). The TTS uses Kokoro with Supertonic-style voice names (F1-F10, M1-M10) auto-mapped to Kokoro voices.

Voice model tuning

# Use a smaller STT model (faster, less accurate)
AISHE_STT_MODEL=base AISHE_INSTALL_VOICE=1 bash setup.sh

# Use GPU STT (requires libcublas — usually pre-installed with nvidia-driver-535+)
AISHE_STT_DEVICE=cuda AISHE_INSTALL_VOICE=1 bash setup.sh

Platform Support

Platform	Status	Notes
macOS	✅ Full	Native audio playback (afplay), AVFoundation mic recording
Linux	✅ Full	PulseAudio/ALSA mic recording, aplay/paplay playback
Windows	✅ Core	DirectShow mic recording, PowerShell audio playback

Voice features (STT/TTS) work on all platforms — they're HTTP services. Mic recording uses ffmpeg with the appropriate platform backend. Audio playback uses the native player for each OS.

Usage

Status

aishe status

Shows all service health with beautiful boxed UI, model counts, memory entries, and thread counts.

Chat

aishe chat "What is 15 * 4?"
aishe chat "What time is it in Tokyo?" -v   # verbose: shows tool calls
aishe stream "Tell me a short joke"          # streaming: tokens print live
aishe repl                                   # continuous text chat mode

Live Voice Conversation

aishe live                    # press Enter to record, speak, get spoken response
aishe live --list              # list available microphone devices
aishe live -d 1                # use specific mic device
aishe live --duration 8        # record 8 seconds per turn
aishe live --no-tts            # voice input only, text responses
aishe live -V M1               # use male voice for TTS

Memory

aishe memory add "User prefers Hindi voice interface"
aishe memory search "Hindi"
aishe memory list
aishe memory clear

Threads

aishe threads                  # list all threads
aishe threads --new            # create a new thread
aishe threads --show <id>      # show thread messages
aishe threads --delete <id>    # delete a thread

Voice I/O

aishe voice status
aishe voice transcribe recording.wav
aishe voice speak "Namaste, main Aishe hoon"
aishe voice speak "Hello" -o output.wav --no-play

Pet — your aishe blob

aishe pet                       # foreground animated blob in your terminal
aishe pet status                # one-shot frame
aishe pet status --mood thinking  # render a specific mood
aishe pet inspect               # dump all pet state as JSON (debug-friendly)
aishe pet signal --state listening  # poke the pet's mood from any process
aishe pet milestones            # list all milestones
aishe pet enable / disable      # toggle the pet
aishe pet reset                 # wipe pet state + signals

The pet is a real-time metaball rendered in your terminal. Its mood reflects what's happening: idle (gentle pulse), thinking (denser core while the LLM is working), listening (elongated while recording), speaking (squashed during TTS), error (asymmetric twitch). It grows with usage, tracks your streak, and unlocks milestones like first_chat, ten_chats, streak_7.

State lives in ~/Library/Application Support/aishe/pet/state.json (macOS) or ~/.local/share/aishe/pet/ (Linux). Signals are written to a JSONL log that the foreground pet tails — so any aishe subcommand can poke the pet's mood from a separate process.

Run the pet around the system

The pet can attach as a persistent tmux side-pane so it stays alive in the corner of your terminal while you work in the main pane:

# Inside any tmux session:
aishe pet attach                    # split right, 30% width
aishe pet attach --direction bottom # split below
aishe pet attach --ratio 25         # narrower pet pane
aishe pet detach                    # remove the pet pane

# Outside tmux (macOS only) — opens a separate Terminal.app/iTerm2 window:
aishe pet attach --window

The pet in the side-pane reacts in real time to whatever you run in the main pane (aishe chat, aishe live, etc.) — same signal channel, just a different render target. State writes are serialized with flock so the foreground pet and command handlers don't clobber each other.

Persistent tmux hook

Auto-attach a pet pane in every new tmux window:

aishe pet watch install     # adds a hook to ~/.tmux.conf
aishe pet watch status      # check if the hook is active
aishe pet watch uninstall   # remove the hook
tmux source-file ~/.tmux.conf

Ollama

aishe ollama models
aishe ollama pull deepseek-v4-flash:cloud
aishe ollama whoami
aishe ollama signin

Diagnostics

aishe doctor

Runs comprehensive checks: Python environment, dependencies, external tools, all services, STT/TTS/DeepAgent roundtrip tests.

Search & Export

aishe search "pizza"           # search across memory + threads
aishe export                   # export all data to ~/Downloads

Configuration

aishe config                   # view full config
aishe config get voice.default_voice
aishe config set voice.default_voice M1

Shell Completions

aishe completions bash > ~/.bash_completion.d/aishe
aishe completions zsh > /usr/local/share/zsh/site-functions/_aishe
aishe completions fish > ~/.config/fish/completions/aishe.fish

Architecture

┌──────────┐     ┌──────────┐     ┌──────────┐
│  aishe   │────▶│ DeepAgent│────▶│  Ollama  │
│   CLI    │     │  :8765   │     │ :11434   │
└──────────┘     └──────────┘     └──────────┘
       │               │
       ▼               ▼
┌──────────┐     ┌──────────┐
│faster-   │     │  Kokoro  │
│whisper   │     │   TTS    │
│ STT:5093 │     │  :8766   │
└──────────┘     └──────────┘

The CLI talks to four local services:

• DeepAgent — LangGraph agent with tool calling, streaming, and conversation memory
• Ollama — LLM inference (local or cloud models)
• faster-whisper STT — OpenAI Whisper via CTranslate2 (CPU or GPU)
• Kokoro TTS — ONNX-based, 54 voices, ~350 MB model

Data

All data is local:

• Memory: ~/Library/Application Support/aishe/memory/facts.jsonl
• Threads: ~/Library/Application Support/aishe/threads/*.json
• Intent Logs: ~/Library/Application Support/aishe/intent_lab/intent_*.jsonl
• Config: ~/.config/aishe/config.yaml

Extending

The CLI is a single Python package — easy to fork, modify, and build upon. The architecture is service-oriented: swap in any OpenAI-compatible LLM, any STT/TTS backend, or add new tools to the DeepAgent sidecar.

License

MIT