Skip to content

Guide

Architecture·

One daemon owns the body; four personas share one brain and one tool list; a dashboard, a tunnel and a fleet bridge sit beside them on the CM4. Where each piece lives and how a turn flows.

  • anyone about to change how TINY is wired
  • 1 min
  • 211 words
  • from the code
  • verified 2026-09-21

In 10 seconds

  • Pollen's daemon (:8000) alone touches motors, camera, mics, speaker. TINY is an HTTP/WS client.
  • One tool list (19 reachy_* + 8 shared) and one factory (build_agent(persona)) serve five faces.
  • One brain: a SQLite file every persona reads first and writes last (the brain).

The body·

The MuJoCo twin. Each hotspot opens the tool that drives that part.

No arms, no legs: a 6-DOF head, a rotating base, two antennas, a camera, four mics, a speaker.

Physical layout·

flowchart TB
  subgraph HOST["Reachy Mini Wireless · CM4 (or your laptop, for a Lite)"]
    direction TB
    D["reachy-mini-daemon · FastAPI :8000<br/>motors · camera · mics · speaker · IMU"]
    subgraph TINY["TINY · /venvs/apps_venv · systemd user units"]
      V["tiny-voice<br/>voice_listener.py"]
      T["tiny-telegram<br/>telegram_listener.py"]
      K["tiny-thinker<br/>thinker_loop.py"]
      DB["reachy-dashboard<br/>dashboard.server :8097"]
    end
    B[("mem.db<br/>the shared brain")]
    TTS["tiny-tts · Piper :5002"]
  end
  V & T & K & DB -->|HTTP/WS| D
  V & T & K & DB <--> B
  T & K & DB --> TTS
  TUN["cloudflared tunnel"] --> DB
  FLEET["tiny.technology fleet<br/>TINY_MCP=1 · use_device"] -.-> V & T & K & DB

A Lite is the same picture on a laptop; REACHY_USE_SIM=1 swaps in MuJoCo. Every tool shares one cached get_mini() client that rebuilds itself after a daemon restart.

Never construct a second ReachyMini()

Two clients fight the daemon.

# the SDK surface TINY relies on
goto_target · set_target · wake_up · goto_sleep · get_current_head_pose · get_current_joint_positions
enable_motors · disable_motors · enable_gravity_compensation · look_at_image · play_move · enable_wobbling
media.get_frame · media.play_sound · imu · RecordedMoves.list_moves()

One turn·

flowchart LR
  M(["message · voice · timer"]) --> I["inject live state<br/>head pose · antennas · IMU · other personas' last turns"]
  I --> P["model plans tool calls"]
  P --> T["expression + speech<br/>in the same batch"]
  T --> R["reply · brain write"]
  R --> U(["speaker · chat · log"])

A persona wakes knowing what the others did and where its head is. Gestures go in the same turn as speech.

The tool layers·

layer modules what
Robot reachy_motion · reachy_expression · reachy_state · reachy_camera · head_tracking · turn_to_sound · reachy_audio 19 wrappers, every angle clamped — safety
Brain memory · voice_bridge · telegram · vision · prompts · manage_* 7 shared tools — the brain
Fleet tiny_mcp use_device & co, TINY_MCP=1 only — fleet

tiny.py·

build_tools()            # text personas: everything above
build_voice_tools()      # voice + dashboard Ask: the latency-slim list
build_agent(persona)     # shell · telegram · thinker · dashboard — one factory, per-persona prompt
build_shell_agent()      # the REPL you get from `make run`
build_voice_agent()      # the voice persona: strands.bidi via tools/bidi_compat (openai · nova_sonic · gemini)

Change a tool once, every face gets it. Daemon down → every tool returns an error string the model can read aloud.