reachy-mini-cli 0.18.1

Agent and CLI for operating the Reachy Mini expressive robot — device setup, app management, and runtime ops.

reachy-mini-cli

Agent and CLI for operating the Reachy Mini expressive robot — device setup, app management, and runtime ops.

Install

# Recommended — real mode (local robot: daemon binary + SDK):
uv tool install 'reachy-mini-cli[daemon]'
# or with pip:  pip install 'reachy-mini-cli[daemon]'

# HTTP-only remote profile (no local robot; talk to a daemon elsewhere):
uv tool install reachy-mini-cli

The installed command is reachy-mini-cli (short alias: reachy). Then:

reachy-mini-cli quickstart     # copy-paste install + start-real-mode sequence
reachy-mini-cli daemon start   # bring the local daemon up (wakes the robot)
reachy-mini-cli device status  # verify it answers
reachy-mini-cli listen run     # orient the head toward sound (Ctrl-C to stop)

See Install profiles below for why reachy-mini is an extra.

What you get

• An agent-first CLI cited from teken (afi-cli).
• A mesh identity — culture.yaml (suffix + backend) and the matching prompt file (CLAUDE.md for backend: claude).
• The canonical guildmaster skill kit (11 skills) under .claude/skills/, vendored cite-don't-import. See docs/skill-sources.md.
• A build + deploy baseline — pytest, lint, the agent-first rubric gate, and PyPI Trusted Publishing wired into GitHub Actions.

Developer quickstart

For working on the repo itself (an editable checkout, not an end-user install):

uv sync --extra daemon                # recommended — SDK + the reachy-mini-daemon binary
uv sync                               # bare — numpy only; HTTP remote profile (--transport http)
uv run pytest -n auto                 # run the test suite
uv run reachy whoami                  # identity from culture.yaml
uv run reachy learn                   # self-teaching prompt (add --json)
uv run teken cli doctor . --strict    # the agent-first rubric gate CI runs

CLI

Verb	What it does
whoami	Report this agent's nick, version, backend, and model from culture.yaml.
learn	Print a structured self-teaching prompt.
explain <path>	Markdown docs for any noun/verb path.
overview	Read-only descriptive snapshot of the agent.
doctor	Check the agent-identity invariants (prompt-file-present, backend-consistency).
cli overview	Describe the CLI surface itself.

Every command supports --json. Results go to stdout, errors/diagnostics to stderr (never mixed). Exit codes: 0 success, 1 user error, 2 environment error, 3+ reserved.

Robot operations

The daemon, device, app, and move noun groups operate the Reachy Mini.

Install profiles

listen is SDK-first — the sdk transport is its default, giving direct in-process access to the mic array (real DoA + RMS loudness via reachy_mini). numpy is a base dependency (the RMS detector; a pure wheel). reachy-mini itself stays an extra, not a base dep, because its transitive stack (pycairo / gstreamer / pyaudio) needs system libraries a bare box / CI lack — so the recommended install bundles it via an extra.

• Recommended — with the SDK + daemon: uv tool install 'reachy-mini-cli[daemon]' (or pip install 'reachy-mini-cli[daemon]' / [sdk]). Pulls reachy-mini, so the sdk transport listen defaults to works out of the box and reachy daemon start can bring the daemon up locally.
• Bare — HTTP remote profile: pip install reachy-mini-cli (no extra). numpy-only; use --transport http (or REACHY_TRANSPORT=http) with --base-url / REACHY_BASE_URL to talk to a daemon running elsewhere via its REST API. Running the sdk transport here exits 2 with a hint to install the [sdk] extra.

reachy-cli still works as a transitional alias (pip install reachy-cli) — it just pulls in reachy-mini-cli. Either package installs the same tool, and the command is reachy (or reachy-mini-cli).

Bring the daemon up

device/app/move are clients of a running daemon; daemon is the other half — it manages the local reachy-mini-daemon process.

Verb	What it does
daemon start	Spawn reachy-mini-daemon in the background, then poll its health route until ready. Idempotent.
daemon stop	Stop the daemon this CLI started (SIGTERM, then SIGKILL).
daemon status	Reconcile the tracked process (running/stopped/stale) with the HTTP health check.

reachy-mini-daemon defaults to --wake-up-on-start, so daemon start already wakes the robot. Forward daemon args after --, e.g. reachy daemon start -- --sim --no-wake-up-on-start. State (PID + log) lives under $XDG_STATE_HOME/reachy (~/.local/state/reachy).

Transports

The device, app, move, and listen verbs talk to the robot through a selectable transport flavor:

• http — the Reachy daemon's REST API. Uses only the Python standard library. Point it at a daemon with --base-url or REACHY_BASE_URL (default http://localhost:8000). This is the default for device, app, and move.
• sdk — the in-process reachy_mini client (included in the base install). Covers motion/state and live audio streaming. This is the default for listen: it streams mic audio directly in-process for real DoA + RMS loudness. Daemon and app verbs still require http.

Select per command with --transport {http,sdk} (or REACHY_TRANSPORT). Action verbs also accept --timeout. If no daemon is reachable, the command exits 2 with a clean error:/hint: pair — never a traceback.

Verb	What it does
device status	Daemon status: state, version, wireless/lite, simulation, IP.
device state	Live robot state: head pose, antenna positions, body yaw.
app list	Available apps (installed and installable).
app status	The currently running app, if any.
app start <name>	Start an installed app by name.
app stop	Stop the currently running app.
move goto	Move head/antennas (mm + degrees); see reachy explain move for flags.
move wake	Play the wake-up animation.
move sleep	Play the go-to-sleep animation.

Each noun also exposes overview (e.g. reachy move overview).

uv run reachy daemon start            # bring the local daemon up (and wake the robot)
uv run reachy device status           # now answers instead of exit-2
uv run reachy app list --json
uv run reachy move goto --z 10 --pitch -5 --duration 2
uv run reachy move wake
uv run reachy daemon stop             # put it back down when you're done

Demo mode — make the robot feel alive

demo-mode runs a continuous loop that streams gentle idle motion to the robot — a slow breathing oscillation, the occasional glance to a new gaze target, and a little antenna sway — so an idle robot looks present rather than frozen. It drives the robot through the same transport, so it needs a running daemon. It is built to run always-on and improve over time, so it has three layers.

Process (tracked background loop; PID + log under $REACHY_STATE_DIR / $XDG_STATE_HOME/reachy):

Verb	What it does
demo-mode start	Spawn the feel-alive loop in the background (idempotent; preflights the daemon first).
demo-mode stop	Stop the loop this CLI started (SIGTERM eases the robot back to neutral, then SIGKILL).
demo-mode restart	Apply an update — restart the service if active, else relaunch the background loop.
demo-mode status	Loop process state + systemd unit state + whether the daemon answers.
demo-mode run	Run the loop in the foreground (what start/the service launch); Ctrl-C to stop.

Config — persisted tuning at $XDG_CONFIG_HOME/reachy/demo-mode.json, read by run/start (CLI flags override per-invocation):

uv run reachy demo-mode config                       # show resolved config
uv run reachy demo-mode config --init                # write defaults
uv run reachy demo-mode config --set energy=0.8 interval=3

Keys: transport, base_url, timeout, interval (tempo), energy (liveliness 0–n), interpolation, seed, wake, settle.

Service — run it always-on as a systemd --user unit (auto-restart on crash, start on boot):

uv run reachy demo-mode install      # write the reachy-demo-mode.service unit
uv run reachy demo-mode enable       # enable --now + linger (survives reboot)
uv run reachy demo-mode disable      # stop + disable
uv run reachy demo-mode uninstall    # remove the unit

The full flow:

uv run reachy daemon start                    # something for the loop to drive
uv run reachy demo-mode config --set energy=0.7
uv run reachy demo-mode start                 # robot starts feeling alive
uv run reachy demo-mode restart               # apply config/code updates
uv run reachy demo-mode stop                  # eases back to neutral

As you make the motion richer over time, edit reachy/alive.py (or the config) and demo-mode restart to apply it.

Behaviors — compose motion on a 50 Hz engine

behavior runs a 50 Hz engine that composes named behaviors on a per-channel contention model (head / antennas / body_yaw), with feel-alive as a passive base layer. Push behaviors onto the running engine from separate commands:

uv run reachy behavior engine start            # bring the 50 Hz loop up
uv run reachy behavior run speak --duration 8  # head bobs like speech
uv run reachy behavior status --json
uv run reachy behavior stop all                # keeps the feel-alive base layer

See reachy explain behavior for the full catalog, channels, and contention model.

Listen — two-tier sound orienting (SDK-first)

listen is SDK-first and two-tier: it streams live mic audio from the reachy_mini SDK in-process (the sdk transport is the default), giving it real Direction of Arrival and real RMS loudness without polling the daemon.

Tier 1 — antenna lean: On every tick, the antennas lean toward the live DoA (head holds). This gives the robot a subtle "perked ear" reaction to any live sound.

Tier 2 — head→body turn: On detected speech or a loud RMS "snap" transient (a sudden noise spike detected by SnapDetector), the robot executes a smooth head→body turn. The head turns first; if the source is beyond --head-only-band degrees from center, the body rotates to face it and the head re-centers. A latched-DoA guard prevents stale angles from firing a spurious turn at rest.

Both tiers drive the daemon's smooth minjerk goto planner through a serial motion queue, so turns are soft and never conflict. The HTTP transport is available via --transport http / REACHY_TRANSPORT=http for a control box connecting remotely.

uv run reachy daemon start                          # bring the daemon up
uv run reachy listen run                             # foreground, SDK transport (default)
uv run reachy listen run --transport http            # foreground, HTTP transport
uv run reachy listen start                           # background, tracked process
uv run reachy listen status --json
uv run reachy listen stop                            # eases back to center

Key tuning flags: --gain, --deadband, --hold, --speed, --recenter-after, --speech-only (Tier-2 speech only; Tier-1 still runs), --antenna-max, --body-yaw-max, --head-only-band, --snap-ratio, --snap-floor. See reachy explain listen for the full reference.

Vision — pixel-based motion + light orienting

vision is the camera counterpart to listen: it reads frames from the robot's camera and orients the head toward the strongest visual event. The camera was previously unused by the CLI — this is a net-new perception channel.

Pixel-based; no ML and no GPU. Two detectors run in pure pixel math on every frame:

• Motion (primary cue) — frame differencing: the centroid of pixel change between consecutive frames maps to a head-yaw offset.
• Light (fallback cue) — brightness/centroid: when no motion fires, the weighted brightness centroid of the frame drives a softer look toward the bright region.

Like listen, vision mirrors the serial motion queue: both detectors drive the daemon's smooth minjerk goto planner one move at a time, so turns are soft and never conflict.

Local-profile only. Frames come via the in-process reachy_mini SDK (the sdk transport is the default); the http transport gives camera-metadata-only access (vision specs). The [sdk] / [daemon] extra is required for vision run and the background process.

reachy-mini-cli vision specs         # check camera metadata (remote-safe)
reachy-mini-cli daemon start         # bring the daemon up
reachy-mini-cli vision run           # foreground, SDK transport (default)
reachy-mini-cli vision start         # background, tracked process
reachy-mini-cli vision status --json
reachy-mini-cli vision stop          # eases back to center

Key tuning flags: --gain, --max-yaw, --deadband, --hold, --speed, --motion-threshold. See reachy explain vision for the full reference.

Say — text-to-speech pipe

say is a dumb pipe: text → TTS synthesis → robot speaker playback. No LLM, no senses — deliberately boundary-clean so agents can compose it into pipelines. Pass "-" as the text argument to read from stdin.

reachy-mini-cli say run "Hello from Reachy"
echo "Hello from stdin" | reachy-mini-cli say run -
reachy-mini-cli say run "Remote" --transport http --base-url http://reachy.local:8000

TTS endpoint — Magpie-style HTTP server: REACHY_TTS_URL (default http://localhost:9000), REACHY_TTS_VOICE (voice identifier), --speed (forwarded, a no-op if the server ignores it). Playback transport — REACHY_TRANSPORT (sdk default; http for remote) via the same --transport flag as listen/vision. See reachy explain say.

Think — continuous cognition loop with body expression

think runs a sentence-streamed cognition loop that also moves the robot's body in step with its thoughts. On each turn it snapshots live senses (DoA + mic loudness) into an event buffer, sends the buffer to an LLM (streamed), and plays each synthesized sentence through the speaker while the LLM is still generating the next — so speech starts before the turn completes. An empty buffer (no notable events) is a silent no-op turn (no LLM call, no audio).

Convention — *emoji* / "quoted" output format: the LLM interleaves expression markers and speech in its output stream. Only "quoted" text is spoken aloud; each *emoji* marker (e.g. *🤔*) drives one calm expression move on the robot. Text outside these delimiters is silently discarded.

Expression catalog — reachy/speech/expressions.toml is an editable, emoji-keyed data file mapping each emoji to a target head/antenna/body pose. Tune expressions by editing this file directly; no code change is needed. Starter set: 🤔 😮 🙂 👂 😐 🎉 😔 (+ neutral fallback for unknown emoji). Unknown emoji fall back to neutral silently.

Focused idle: while think run is active it writes a file flag (think_active.flag under $REACHY_STATE_DIR). A co-running listen/idle loop reads this flag and drops to a low-energy "focused breathe" — so the body quiets and stillness becomes the thinking posture even if both listen and think run together.

Like listen, think is SDK-first: the sdk transport (default) streams real DoA + mic RMS in-process; --transport http polls the daemon's DoA route. Like daemon, it has a foreground loop (run) and a tracked background process (start/stop/restart/status) managed by its own supervisor (reachy/speech/supervisor.py).

reachy-mini-cli daemon start                              # bring the daemon up
reachy-mini-cli think run                                 # foreground loop (Ctrl-C to stop)
reachy-mini-cli think run --max-turns 3                   # stop after 3 spoken turns
reachy-mini-cli think start --llm-model mistral-small     # background process
reachy-mini-cli think status --json
reachy-mini-cli think stop
reachy-mini-cli think demo                                # drive a scripted *emoji* "speech" stream (no LLM)

think demo runs a fixed *emoji* "speech" script through the same path the cognition loop uses (marker parser → expression moves + spoken quoted text) so you can watch the body language on a live robot without an LLM. See docs/verification/think-body-expression.md for the manual observer checklist.

Catalog tooling:

reachy-mini-cli think expressions           # list catalog emojis + pose descriptors
reachy-mini-cli think expressions list      # same (explicit verb)
reachy-mini-cli think expressions check     # flag poses too similar to tell apart
reachy-mini-cli think expressions check --json

expressions check exits 0 always — a flagged pair is a warning, not an error. The --json ok field is the machine-readable signal.

LLM endpoint — REACHY_LLM_BASE_URL, REACHY_LLM_API_KEY, REACHY_LLM_MODEL; pure urllib streaming (no OpenAI SDK, no new base dep). TTS endpoint — REACHY_TTS_URL, REACHY_TTS_VOICE (same as say). Playback — REACHY_TRANSPORT (sdk/http); same transport flag as say. Key pacing flags: --turn-interval (seconds between turns), --mute-after-speak (self-mute window after speaking, default 2.5 s). See reachy explain think for the full reference.

Pat — proprioceptive touch + snuggle

pat makes the robot feel a head pat and lean into it — with no touch sensor. It holds a baseline head pose, reads the actual head pose back from the robot, and detects a pat from the commanded-vs-actual deviation: a downward pitch press is a scratch (head/neck pat) and a sideways yaw nudge is a side-nudge (side_pat). On a detected pat it leans in — a calm lean→nuzzle→settle snuggle (a pitch-down dip for a scratch; a yaw-toward + soft body-yaw for a side_pat) — enqueued onto the same shared serial MotionQueue the executor drains one move at a time, so the reaction is smooth and never fights other motion.

Like listen/think, pat is SDK-first: the head_pose read-back is an SDK-only capability, so the sdk transport (default) is required to feel a pat; --transport http is available for non-pose ops but cannot read head_pose. The demo verb needs no robot and no [sdk] extra — it synthesizes pat events and runs them through the real lean planner so you can verify the wiring in CI or by eye.

reachy-mini-cli pat run                                   # foreground loop (Ctrl-C to stop)
reachy-mini-cli pat run --ticks 200                       # bounded run (testing/ops)
reachy-mini-cli pat run --press-threshold 1.5 --min-presses 3   # retune sensitivity
reachy-mini-cli pat demo                                  # synthesize pats, no robot
reachy-mini-cli pat demo --json --count 1                 # one scripted reaction, JSON

Breaks idle stillness: while a pat reaction is enqueued, pat writes a file flag (pat_active.flag under $REACHY_STATE_DIR, the counterpart to think's think_active.flag). A co-running listen/idle loop reads this flag and pauses its idle wander entirely so the lean/snuggle owns the motion — a scratch fully breaks stillness, where a think turn only quiets it to a focused breathe. See reachy explain pat for the full reference.

Sleep — decay-to-sleep + wake

sleep makes the robot fall asleep when left alone and wake up when addressed. It tracks idle time against configurable thresholds and decays through three states — ALERT → DROWSY → ASLEEP — using a purely injected-clock idle timer (no wall-clock dependency). Motion follows the state: a drowsy fade on the way down, a quiet sleep-breathe cycle while asleep, and a wake gesture on resumption.

Idle timer and state machine — reachy/sleep/state.py holds SleepState (ALERT/DROWSY/ASLEEP) and the injected-clock timer. Any qualifying stimulus resets the timer; without one, the robot decays through the thresholds automatically. reachy/sleep/stimulus.py classifies incoming sense events as qualifying stimulation, including a self-mute exclusion so the robot does not wake itself from its own speaker output.

Two-tier wake — reachy/sleep/wake.py: Tier 1 (default) wakes on detected speech or a loud RMS snap transient (same signals as listen Tier 2). Tier 2 adds optional wake-word detection via a pluggable backend (see below); the wake-word path is lazy-loaded and degrades gracefully if the extra is absent.

Pat-only wake (--no-audio-wake / --wake pat) — pass this flag to disable all audio wake so only a physical head pat rouses the sleeping robot. Useful when the robot is in a noisy environment where speech/snap signals would cause false wakes. Requires the sdk transport (pat detection reads live head_pose); on --transport http it exits 2 with a clean CliError.

Pluggable wake-word backend — opt in with --wake-word (+ optional --wake-word-kind {http,openwakeword} and --wake-phrase). Two backends:

• http (default) — an external OpenAI-compatible STT service, targeting the model-gear / NVIDIA Parakeet endpoint (POST /v1/audio/transcriptions, a multipart WAV upload; stdlib urllib, no extra required). It buffers a rolling ~1.5 s audio window (one tick's mic chunk is too short to transcribe a phrase) and matches the wake phrase against the response text. Configure with REACHY_STT_URL (default http://localhost:9002), REACHY_STT_PHRASE, REACHY_STT_LANGUAGE, REACHY_STT_TIMEOUT. The service is externally managed.
• openwakeword — on-box wake-word engine under the [cpu] extra (lazy-loaded; the dep list is empty until openwakeword gains a cp312 wheel).

The [gpu] extra is a generic GPU compute-class pin for future GPU-accelerated features — it does not bundle an on-box STT model.

Idle interrupt — strongest suppressor: sleep writes sleep_active.flag (via reachy/motion/sleep_signal.py) while the robot is in DROWSY or ASLEEP state. A co-running listen/idle loop reads this flag as the highest-priority idle interrupt — above pat_active.flag and think_active.flag — and yields the motion channel entirely to sleep's SleepProducer (reachy/motion/sleep.py).

Like pat/think, sleep is SDK-first: the sdk transport (default) is the full path; --transport http is available for non-pose ops. A missing [sdk] extra raises a clean exit-2 CliError. sleep has its own supervisor (reachy/sleep/supervisor.py, tracking sleep.pid/sleep.log under $REACHY_STATE_DIR) separate from the listen and think supervisors.

reachy-mini-cli sleep run                                 # foreground decay loop (Ctrl-C to stop)
reachy-mini-cli sleep run --ticks 300                     # bounded run (testing/ops)
reachy-mini-cli sleep run --idle-timeout 30               # fall asleep after 30 s of quiet
reachy-mini-cli sleep run --no-audio-wake                 # pat-only wake (ignore all sound)
reachy-mini-cli sleep run --wake pat                      # same as --no-audio-wake
reachy-mini-cli sleep run --wake-word                     # opt-in wake-word (HTTP STT default)
reachy-mini-cli sleep run --wake-word --wake-word-kind openwakeword --wake-phrase "hey reachy"
reachy-mini-cli sleep start                               # background tracked process
reachy-mini-cli sleep status --json                       # state + idle timer + health
reachy-mini-cli sleep stop
reachy-mini-cli sleep restart
reachy-mini-cli sleep demo                                # walk ALERT→DROWSY→ASLEEP→wake, no robot
reachy-mini-cli sleep demo --json                         # machine-readable event trace

sleep demo uses injected senses and a fake clock to walk the full state machine deterministically — no robot and no [sdk] extra needed. See reachy explain sleep for the full flag reference.

Make it your own

1. Rename the package reachy/ and the reachy-mini-cli CLI/dist name throughout pyproject.toml, the package, tests/, and sonar-project.properties.
2. Edit culture.yaml with your suffix and backend.
3. Rewrite CLAUDE.md for your agent and run /init.
4. Re-vendor only the skills you need from guildmaster (see docs/skill-sources.md).

See CLAUDE.md for the full conventions (version-bump-every-PR, the cicd PR lane, deploy setup).

License

MIT — see LICENSE.