Interactive Mode

​

Usage

rcli

# With RAG index loaded
rcli --rag ~/Library/RCLI/index

# Custom models directory
rcli --models /path/to/models

​

TUI Layout

• Chip Bar — CPU/GPU/RAM specs (M1/M2/M3/M4)
• System Gauges — CPU, Memory, GPU, Battery, Network
• Models Panel — Active LLM, STT, TTS, RAG status
• Chat History — Conversation with performance metrics
• Voice Panel — Animated dog with audio level meter (during recording)
• Input Bar — Text input field with command prompt
• Status Bar — Keyboard shortcuts

​

Keyboard Shortcuts

​

Voice Input

SPACE        Start recording
SPACE/ENTER  Stop recording & transcribe

​

Text Input

Type + ENTER    Send text command to LLM
ESC             Clear input (when empty: quit)

​

Panel Navigation

M    Models — browse/download/switch LLM/STT/TTS
A    Actions — view and execute macOS actions
B    Benchmarks — run performance tests
R    RAG — ingest documents or manage index
D    Cleanup — delete unused models

​

Utilities

T    Toggle tool call trace (show LLM tool calls inline)
X    Clear conversation history & reset context
Q    Quit RCLI

​

Interactive Commands

help               Show help in chat
models             Show active LLM/STT/TTS
voices             List TTS voices
stt                List STT models
actions            List all actions
actions <name>     Show action details

​

Push-to-Talk Workflow

1. Press SPACE to start recording
2. Speak your command (e.g., “open Safari”)
3. Press SPACE or ENTER to stop
4. RCLI transcribes via STT
5. LLM processes and responds
6. TTS speaks the response (if --no-speak not set)

​

Noise Gate

(no speech detected - speak louder or closer to mic)

​

Tool Call Trace

> open Safari
  ~ [TRACE] Tool call: open_app({"app_name": "Safari"})
  ~ [TRACE] open_app -> OK: {"success": true, "output": "Opened Safari"}
  RCLI: Done! Safari is now open.

​

Context Window Management

• Green — 0-30% usage
• Yellow — 30-70% usage
• Red — 70-100% usage

​

Performance Metrics

• TTFA — Time to first audio (STT + LLM first token + TTS first sentence)
• STT — Transcription latency
• LLM — Tokens generated, tok/s, TTFT
• TTS — Synthesis time, real-time factor

​

Panel Details

​

Models Panel (M)

• Installed — Available locally
• Active — Currently loaded (green checkmark)
• Default — Included in rcli setup
• Recommended — Best for most users

• Download (if not installed) — via curl from Hugging Face
• Switch (if installed) — Hot-swap LLM at runtime, or select STT/TTS for next launch

​

Actions Panel (A)

• SPACE to enable/disable (persists to ~/.rcli/actions.json)
• ENTER to execute directly

​

Benchmarks Panel (B)

• all — All suites (default)
• stt, llm, tts, e2e, tools, rag, memory

​

RAG Panel (R)

• Ingest documents — Enter path, index files
• Show status — Display indexed file count
• Clear index — Remove all indexed documents

​

Cleanup Panel (D)

​

Exit Behavior

• ESC (when idle + empty input) → Quit
• Q → Quit
• Ctrl+C → Graceful shutdown

​

Example Session

$ rcli

# TUI launches, showing chip info and gauges
# Press SPACE, speak "create a note called Shopping List"
# Press ENTER to stop recording
# RCLI transcribes, detects tool call, executes create_note, speaks response
# Type "what's 42 times 17?" and press ENTER
# LLM calculates and responds
# Press M, navigate to Qwen3.5 2B, press ENTER to switch
# Press Q to quit

​

Drag-and-Drop RAG

# Finder drag → Terminal receives path
/Users/you/Documents/project.pdf

# Type: rag ingest /Users/you/Documents/project.pdf
# Or use the RAG panel (R) and paste the path