Intern Developer Guide

This guide walks a Developer Edition owner through the workflows they need on day one: SSH into the device, ship code changes, watch logs, and hit the HAL and OS Server APIs. Everything below assumes the Intern v2 Pro (OrangePi sun60iw2) shipped with SSH open and Codex / Claude Code / OpenCode pre-installed.

Developer Edition only. You can tell the editions apart by the case colour of the unit you bought:

Black case → Developer Edition. SSH is open, the on-device toolchain (Codex / Claude Code / OpenCode) is pre-installed, and every workflow in this guide applies.

Yellow case → OpenClaw edition. Consumer image, SSH is closed, no developer toolchain. This guide does not apply.

Blue case → Hermes edition. Consumer image, SSH is closed, no developer toolchain. This guide does not apply.

If your unit is yellow or blue and you need the developer workflows below, you'll need a Developer Edition (black) unit — the consumer images cannot be unlocked in the field.

The source of truth is the autonomous-os repo. Every file path in this guide is relative to that repo unless noted otherwise.

1. Device access

1.1 SSH into the device

SSH is exposed only on the Developer Edition (black case) — the yellow OpenClaw and blue Hermes consumer units keep sshd disabled and cannot be reached this way. On the Developer Edition, the SD-card image creates a single Linux user with sudo:

	Value
SSH user	`orangepi`
SSH password	`orangepi`
Sudo password	`orangepi` (same as SSH)

Your device IP is assigned by your router's DHCP. Find it in the router admin page or on the web dashboard.

ssh orangepi@<device-ip>
# or
sshpass -p 'orangepi' ssh orangepi@<device-ip>

To become root:

sudo -i
# password: orangepi

1.2 On-device filesystem layout

Path	What lives there
`/usr/local/bin/os-server`	Go binary — main HTTP API server (Gin, :5000)
`/usr/local/bin/bootstrap`	Go binary — OTA worker
`/opt/hal/`	Python HAL (FastAPI, :5001) — full source lives here
`/opt/hal/.env`	HAL environment variables (audio devices, VAD, model URLs, …)
`/root/config/config.json`	OS Server config — LLM base URL, API key, channel tokens, Wi-Fi, admin hash
`/root/config/bootstrap.json`	Bootstrap OTA metadata URL
`/root/.openclaw/`	OpenClaw runtime state — `openclaw.json`, skills, workspace
`/etc/systemd/system/os-server.service`	Service unit for os-server
`/etc/systemd/system/hal.service`	Service unit for HAL
`/etc/systemd/system/bootstrap.service`	Service unit for OTA worker
`/etc/nginx/sites-enabled/default`	Nginx — proxies `/api/` to os-server, `/api/hardware/` to HAL, serves setup Web UI

Web (nginx) sits in front on port 80. The /api/* prefix is os-server, the /api/hardware/* prefix is HAL.

1.3 Web CLI — browser terminal from the Admin panel

If SSH is inconvenient (locked-down laptop, guest network, no ssh client), the Admin Web UI ships an interactive terminal at:

http://<device-ip>/monitor#cli

Admin password (how to sign in). The Admin panel is gated by the device admin password — different from the SSH orangepi password above. On a factory-fresh unit the default is the 4-character suffix printed on the sticker on the underside of the device: the sticker reads Intern-XXXX, and XXXX (case-sensitive) is your default admin password. Enter it on the Admin login page at http://<device-ip>/login.

To rotate the password, sign in and go to Settings → General (http://<device-ip>/setting#general); the new value is written to /root/config/config.json and used from the next login. If you forget the password later, either read it back from /root/config/config.json over SSH or do a soft-reset from the Admin panel to fall back to the Intern-XXXX default.

Once you're signed in, click Device → CLI in the sidebar. Each tab opens a fresh interactive bash -il running as root on the device, exactly like SSH — history, aliases, $PATH, colours, everything. Up to 6 tabs can be open at once and backgrounded tabs keep their shell alive, so a build in tab 1 keeps running while you tail a log in tab 2.

Transport: WebSocket at /api/system/shell, admin-auth gated.
The Web CLI also pre-injects the active agent runtime's env file (e.g. /root/.claudecode/.env for the Claude Code runtime), so claude, codex, opencode etc. inherit the campaign API key and don't prompt for login the way a raw SSH shell would.
Source: system/server/system/shell.go (backend PTY + WebSocket) and system/web/src/pages/monitor/CliSection.tsx (xterm.js frontend).

Use SSH for scripted / long-running work (rsync, tmux sessions that must survive a browser close). Use the Web CLI for quick "poke at the device" sessions or when you're on a machine without a proper terminal.

2. Logs

Every long-running component is a systemd unit — use journalctl -u <name>.

# Live tail — most common
sudo journalctl -u os-server -f
sudo journalctl -u hal -f
sudo journalctl -u openclaw -f
sudo journalctl -u bootstrap -f

# Last 100 lines
sudo journalctl -u os-server --no-pager -n 100

# Errors only
sudo journalctl -u hal -p warning --no-pager -n 100

# By time window
sudo journalctl -u os-server --since "10 minutes ago"
sudo journalctl -u hal --since "2026-07-08 10:00" --until "2026-07-08 11:00"

Grep tip: the log lines carry a component=<name> tag. Filter by that when tailing os-server:

sudo journalctl -u os-server -f | grep component=mqtt

3. Configuration

3.1 LLM / STT / TTS (OS Server — `/root/config/config.json`)

The main config for the LLM brain plus messaging channels lives here. Edit directly, then restart os-server for the change to apply.

Key fields:

{
  "llm_base_url": "https://api.openai.com/v1",
  "llm_api_key": "sk-…",
  "llm_model": "gpt-4o-mini",

  "stt_base_url": "https://api.openai.com/v1",
  "stt_api_key": "sk-…",
  "stt_language": "en",

  "tts_base_url": "https://api.openai.com/v1",
  "tts_api_key": "sk-…",
  "tts_provider": "openai",
  "tts_voice": "alloy",

  "channel": "telegram",
  "telegram_bot_token": "…",
  "telegram_user_id": "…"
}

To use a self-hosted OpenAI-compatible endpoint (Ollama / vLLM / LM Studio), set llm_base_url to http://<host>:<port>/v1. Any API key placeholder works if the endpoint doesn't check auth.

sudo nano /root/config/config.json
sudo systemctl restart os-server

3.2 HAL runtime env (`/opt/hal/.env`)

HAL reads its runtime tuning (audio devices, VAD thresholds, realtime voice config, camera settings) from /opt/hal/.env. A curated set of presets lives in the repo under hal/env-presets/.

sudo nano /opt/hal/.env
sudo systemctl restart hal

Common knobs:

Var	What it does
`HAL_AUDIO_INPUT_ALSA`	Mic ALSA device (`plug:device_micro2` on Intern v2 Pro)
`HAL_AUDIO_OUTPUT_ALSA`	Speaker ALSA device
`HAL_VAD_THRESHOLD`	Voice-activity RMS floor (lower = more sensitive)
`HAL_SILENCE_TIMEOUT`	Seconds of silence before committing a turn
`HAL_REALTIME_TURN_DETECTION`	`on` = server-side VAD (Gemini/OpenAI), `off` = client VAD
`HAL_GEMINI_LIVE_MODEL`	Gemini realtime model id
`HAL_OPENAI_REALTIME_MODEL`	OpenAI realtime model id
`DEVICE_TYPE`	`intern-v2` — hardware kit gate for drivers like the mic mute switch

4. Building + shipping code

Do the build on your Mac / dev machine — the device only runs the artifacts. The Go binary cross-compiles to linux/arm64; the HAL is plain Python, no build step needed.

Disable the OTA bootstrap first. The bootstrap service polls the Intern team's OTA channel and, if it finds a newer official build, will reinstall os-server / HAL / web on top of whatever you just pushed — silently wiping your custom binary at the next tick. Before you push a single custom build to a device, turn bootstrap off so your changes stick:
sudo systemctl disable --now bootstrap
See §5 for the full enable / disable / re-enable flow. Only turn bootstrap back on when you want to go back to tracking the official fleet builds — at that point the next tick will overwrite your custom changes with the current OTA version.

4.1 Go — os-server + bootstrap

# From the repo root
make os-build            # → system/os-server
make os-build-bootstrap  # → system/bootstrap

Push the binary and restart:

sshpass -p 'orangepi' scp system/os-server orangepi@<ip>:/tmp/os-server-new
sshpass -p 'orangepi' ssh orangepi@<ip> "echo orangepi | sudo -S bash -c '
  cp /usr/local/bin/os-server /usr/local/bin/os-server.bak.\$(date +%s)
  mv /tmp/os-server-new /usr/local/bin/os-server
  chmod 755 /usr/local/bin/os-server
  systemctl restart os-server
'"

4.2 Python — HAL

Sync individual files, or the whole directory:

# One file
sshpass -p 'orangepi' scp hal/drivers/mic_button.py orangepi@<ip>:/tmp/
sshpass -p 'orangepi' ssh orangepi@<ip> "echo orangepi | sudo -S bash -c '
  mv /tmp/mic_button.py /opt/hal/drivers/mic_button.py
  chown root:root /opt/hal/drivers/mic_button.py
  systemctl restart hal
'"

# Whole HAL tree (destructive — replaces /opt/hal contents)
sshpass -p 'orangepi' rsync -avz --delete hal/ orangepi@<ip>:/tmp/hal/
sshpass -p 'orangepi' ssh orangepi@<ip> "echo orangepi | sudo -S bash -c '
  rsync -avz --delete /tmp/hal/ /opt/hal/
  systemctl restart hal
'"

4.3 Web (Setup + Admin UI)

The React SPA lives in system/web/. Nginx serves the built assets from /usr/share/nginx/html/setup/.

# Build
cd system/web && npm install && npm run build   # → dist/

# Push (from repo root)
cd system/web/dist && zip -qr /tmp/setup-web.zip .
sshpass -p 'orangepi' scp /tmp/setup-web.zip orangepi@<ip>:/tmp/
sshpass -p 'orangepi' ssh orangepi@<ip> "echo orangepi | sudo -S bash -c '
  find /usr/share/nginx/html/setup -mindepth 1 -not -name VERSION -exec rm -rf {} + 2>/dev/null || true
  unzip -o -q /tmp/setup-web.zip -d /usr/share/nginx/html/setup
  chown -R root:root /usr/share/nginx/html/setup
'"

Nginx serves the SPA — no restart needed after replacing files.

5. Controlling the OTA bootstrap

The bootstrap service polls https://cdn.autonomous.ai/os/ota/metadata.json and applies any component (os-server, HAL, web, bootstrap itself) whose official version is newer than what's on disk. Left running, it will overwrite any custom os-server / HAL / web build you pushed with scp, usually within a minute — the Intern team ships OTAs on their own cadence and there is no way to opt-out per-file. If you're doing custom development, disable bootstrap first (§4 callout) and only re-enable it when you want your device to snap back to the official fleet build.

# Stop for one session (survives until next boot)
sudo systemctl stop bootstrap

# Persist across reboots — nothing OTA will hit this device
sudo systemctl disable --now bootstrap

# Re-enable when you're done developing
sudo systemctl enable --now bootstrap

To confirm nothing else is overwriting your binary, tail the bootstrap log before and after your push:

sudo journalctl -u bootstrap -f

6. HAL API — TTS / STT / Mic / Speaker / LED

Full HAL routes live in hal/routes/. Reach them from the device with curl http://127.0.0.1:5001/…, or from your laptop / web dashboard via nginx at http://<device-ip>/api/hardware/….

Every endpoint returns {"status": 1, "data": <payload>, "message": null} on success, {"status": 0, "data": null, "message": "…"} on failure.

Two ways to call anything below:

Over HTTP (from anywhere — a Python script, a curl, another service). Every subsection ships a curl you can paste.
In-process from HAL Python code — import the route function directly and call it with the pydantic request. Useful when you're writing a new HAL driver or route.

Line numbers below are anchored to the current main — if they drift, grep for the decorator (e.g. @router.post("/voice/speak") in the same file.

6.1 TTS — make the speaker talk

What	Where
Route module	`hal/routes/voice.py`
`POST /voice/speak` — one-shot say (interrupts current speech)	`voice.py:189` — `def speak_text(req: SpeakRequest)`
`POST /voice/speak-queue` — say after current speech finishes	`voice.py:284` — `def speak_queue_text(req: SpeakRequest)`
`POST /tts/stop` — cut off the currently-playing sentence	`voice.py:328` — `def stop_tts()`
`SpeakRequest` schema (text, voice, provider, interruptible, cached…)	`hal/models.py:228`

# Fire-and-forget say
curl -X POST http://127.0.0.1:5001/voice/speak \
  -H 'Content-Type: application/json' \
  -d '{"text": "Hello from the device"}'

# Queue behind any in-flight speech (e.g. multi-turn narration)
curl -X POST http://127.0.0.1:5001/voice/speak-queue \
  -H 'Content-Type: application/json' \
  -d '{"text": "…and that concludes the update."}'

# Interrupt whatever is playing right now
curl -X POST http://127.0.0.1:5001/tts/stop

Call it from inside HAL Python:

from hal.models import SpeakRequest
from hal.routes.voice import speak_text

speak_text(SpeakRequest(text="Hello", voice="Rachel", interruptible=False))

6.2 STT — read what the mic heard

STT runs continuously as part of the voice pipeline — you generally don't call it to transcribe an ad-hoc WAV. Instead:

Live turn stream (VAD → STT → LLM → TTS): subscribe to the OS Server flow monitor SSE endpoint — see docs/flow-monitor.md.
Current voice-pipeline state (STT provider, language, VAD, is speaking): GET /voice/status — hal/routes/voice.py:443 (def voice_status()).
Live mic RMS level (SSE, useful for a "am I hearing you?" indicator): GET /voice/mic-level — hal/routes/voice.py:376 (async def mic_level_stream(...)).

curl http://127.0.0.1:5001/voice/status
curl -N http://127.0.0.1:5001/voice/mic-level   # SSE stream

The full STT drivers (OpenAI, Whisper local, Google, …) live under hal/drivers/stt/ if you need to plug a new one in.

6.3 Mic — mute / unmute mid-conversation

What	Where
Route module	`hal/routes/voice.py`
`POST /voice/mute` — stops feeding audio into VAD/STT, ignores mic	`voice.py:336` — `def mute_mic()`
`POST /voice/unmute` — re-arms the mic	`voice.py:349` — `def unmute_mic()`
Slide-switch driver (physical mic mute on PD1, Intern v2 Pro)	`hal/drivers/mic_button.py` — calls `mute_mic()` / `unmute_mic()` on GPIO edge

curl -X POST http://127.0.0.1:5001/voice/mute
curl -X POST http://127.0.0.1:5001/voice/unmute

The physical slide switch on the top of the Intern v2 Pro just calls those two endpoints internally — that's the whole pattern for any hardware button you want to bolt on.

6.4 Speaker — mute / play WAV / stop

What	Where
Route module	`hal/routes/music.py`
`POST /audio/play` — play a music track or WAV path	`music.py:219` — `def audio_play(req: MusicPlayRequest)`
`POST /audio/stop` — stop current playback	`music.py:271` — `def audio_stop()`
`POST /speaker/mute` — mute the speaker at the ALSA level (TTS still runs but silent)	`music.py:280` — `def mute_speaker()`
`POST /speaker/unmute` — restore	`music.py:294` — `def unmute_speaker()`
`GET /audio/status` — what's playing right now	`music.py:304` — `def audio_status()`

# Play a local WAV (playing music via the LLM's music skill goes here too)
curl -X POST http://127.0.0.1:5001/audio/play \
  -H 'Content-Type: application/json' \
  -d '{"path": "/opt/hal/assets/ok.wav"}'

# Stop
curl -X POST http://127.0.0.1:5001/audio/stop

# Speaker mute / unmute — TTS keeps running, output is silenced
curl -X POST http://127.0.0.1:5001/speaker/mute
curl -X POST http://127.0.0.1:5001/speaker/unmute

/speaker/mute is the right knob if you want the device to still "think out loud" internally but not disturb the room — TTS advances, LEDs still pulse, only the audio out is muted. Use /tts/stop if you want it to actually shut up.

6.5 LED

What	Where
Route module	`hal/routes/led.py`
`POST /led/solid` — solid RGB colour	`led.py:73` — `def set_led_solid(req: LEDSolidRequest)`
`POST /led/paint` — per-pixel colours	`led.py:91` — `def set_led_paint(req: LEDPaintRequest)`
`POST /led/effect` — named animation (breathing, blink, pulse, rainbow…)	`led.py:119` — `def start_led_effect(req: LEDEffectRequest)`
`POST /led/status` — status-cue overlay (booting, listening, error)	`led.py:183` — `def set_led_status(req: LEDStatusRequest)`
`POST /led/off`	`led.py:101` — `def turn_off_leds(...)`
`POST /led/effect/stop` — stop the running animation, keep last frame	`led.py:229` — `def stop_led_effect()`
`POST /led/restore` — hand strip back to user's saved state (after a transient overlay)	`led.py:206` — `def restore_led()`
`GET /led` — current state	`led.py:35` — `def get_led_state()`
`GET /led/color` — current RGB colour of the whole strip	`led.py:43` — `def get_led_color()`
Named status → colour/effect table (STATUS_LED_PRESETS)	`hal/presets.py:188`

Concrete calls, all copy-pasteable:

# 1. Solid colour — cold blue
curl -X POST http://127.0.0.1:5001/led/solid \
  -H 'Content-Type: application/json' \
  -d '{"color": [0, 135, 255]}'

# 2. Per-pixel paint — colour each of the N LEDs individually
#    (useful for progress bars, level meters, tests)
curl -X POST http://127.0.0.1:5001/led/paint \
  -H 'Content-Type: application/json' \
  -d '{"colors": [[255,0,0],[0,255,0],[0,0,255],[255,255,0]]}'

# 3. Named animation with tint + speed
curl -X POST http://127.0.0.1:5001/led/effect \
  -H 'Content-Type: application/json' \
  -d '{"effect": "breathing", "color": [0, 255, 0], "speed": 1.0}'

# Other effects: blink, pulse, rainbow, wave, chase, sparkle, …
curl -X POST http://127.0.0.1:5001/led/effect \
  -H 'Content-Type: application/json' \
  -d '{"effect": "rainbow", "speed": 0.5}'

# 4. Status overlay — HAL picks the colour/effect from STATUS_LED_PRESETS
#    ("booting", "listening", "wifi_connecting", "hardware", "agent_down", …)
curl -X POST http://127.0.0.1:5001/led/status \
  -H 'Content-Type: application/json' \
  -d '{"state": "listening"}'

# 5. Stop the animation but keep the last frame lit
curl -X POST http://127.0.0.1:5001/led/effect/stop

# 6. Restore the user's saved LED after a transient overlay
curl -X POST http://127.0.0.1:5001/led/restore

# 7. Off
curl -X POST http://127.0.0.1:5001/led/off

# 8. Read current state
curl http://127.0.0.1:5001/led
curl http://127.0.0.1:5001/led/color

Where the LED is actually driven from in this repo (good places to copy patterns from):

Caller	File	What it does
Status-cue overlay service	`system/statusled/service.go:71` — `func (s *Service) Set(state State)`	Priority-stacked overlay: booting / ota / error / connectivity / hal_down / agent_down / hardware / wifi_connecting. Highest priority wins; on `Clear` the strip is restored.
State constants (booting, wifi_connecting, agent_down, …)	`system/statusled/service.go:19-26`	The exact set of states you can pass to `/led/status`.
Wi-Fi connecting blue-blink	`system/device/setup.go:80` — `s.statusLED.Set(statusled.StateWifiConnecting)`	Fires the blue blink while the STA join is happening during setup.
HAL health watchdog	`system/healthwatch/service.go:94,107,136`	Sets `StateHALDown` / `StateHardware` when HAL or a driver stops responding.
Agent-runtime health	`runtimes/openclaw/service_ws.go:43`, `runtimes/claudecode/client.go:57`, `runtimes/codex/client.go:60`, `runtimes/opencode/client.go:60`, `runtimes/picoclaw/client.go:50`, `runtimes/hermes/health.go:126`	Every agent runtime sets `StateAgentDown` when its socket drops so the user sees the LED go red without having to check logs.
Ambient "breathing" idle behaviour	`system/ambient/service.go`	Drives soft colour drift + breathing while there's no interaction. Uses `/led/effect` with `transient=true` so it doesn't clobber the user's saved state.
LLM skill / tool-call bridge (`[HW:/led/…:{...}]`)	`system/server/agent/delivery/http/handler_hw.go`	When an OpenClaw / Hermes / Claude Code skill emits e.g. `[HW:/led/solid:{"color":[255,0,0]}]`, this handler forwards it to HAL. This is how any skill turns the LED red without wiring plumbing itself.
Go HAL client (helpers you'd call from an os-server service)	`system/lib/hal/client.go:76-131` — `StartEffect`, `StopEffect`, `SetLEDStatus`, `RestoreLED`, `GetColor`	Fire-and-forget wrappers around the endpoints above. New Go services should use these instead of hand-rolling HTTP.

Full LED preset catalogue and rendering rules (colours, priorities, per-preset behaviour): docs/led-control.md.

6.6 Other useful HAL surfaces

Domain	Source file	Highlights
Servo (Lamp only)	`hal/routes/servo.py`	`/servo/aim`, `/servo/track`, `/servo/play`
Emotion	`hal/routes/emotion.py`	`/emotion/express` — coordinated servo + LED + display
Scene	`hal/routes/scene.py`	`/scene/reading`, `/scene/focus`, …
Camera	`hal/routes/camera.py`	`/camera/snap`, `/camera/enable`, `/camera/disable`
Display (Lamp)	`hal/routes/display.py`	`/display/text`, `/display/eye`
Face recognition	`hal/routes/face.py`	Enroll / list / remove

7. MQTT — add a new event kind

The device speaks MQTT for admin commands (fired from the web dashboard) and status reports. Every command is an envelope shaped like {"cmd": "data", "kind": "<kind>", "data": {…}}.

7.1 Sending an MQTT command from the backend

The web admin fires commands via the BFF endpoint POST /device/{id}/send-message?listen=true, which relays over MQTT. Web-side client wrapper (in the autonomous.ai frontend repo) is StandToEarnApi.sendDeviceMessage.

Example payload:

{ "cmd": "data", "kind": "device.soft_reset", "data": {} }

7.2 Receiving on the device — add a new handler

Adding a new MQTT command is three edits inside system/:

Declare the kind constant in system/domain/device.go:

const (
    KindDeviceRename    = "device.rename"
    KindDeviceSoftReset = "device.soft_reset"   // ← your new one
    // …
)

Write the handler — one file under system/server/device/delivery/mqtt/. See device_soft_reset_handler.go and device_rename_handler.go for reference patterns (envelope unmarshalling, ack via publishDataResult, goroutine-off-callback for long work).
Wire it into the dispatcher — one case in the switch in system/server/device/delivery/mqtt/handler.go dispatchData():
```
case domain.KindDeviceSoftReset:
    return h.handleDeviceSoftReset(env)
```

Rebuild os-server (§4.1) and the new command is live.

The full MQTT protocol reference (topics, envelope shape, privacy vs inline delivery, ack model) lives in docs/mqtt.md.

8. Web UI — Setup and Admin

The React SPA in system/web/ covers both flows:

Route	What it is	Key files
`/setup`	First-boot provisioning — Wi-Fi + admin password + LLM/channel prefill	`src/pages/Setup.tsx`, `src/components/setup/*`
`/login`	Admin login (bcrypt-checked against `config.admin_password_hash`)	`src/pages/Login.tsx`
`/monitor`	Post-login dashboard — flow monitor, live device state	`src/pages/monitor/*`
`/setting`	Admin settings — rotate password, edit config, channel management	`src/pages/settings/*`

Setup form deep-links: the companion app pushes ?llm_api_key=…&tele_token=… into the Setup page so operators don't retype secrets. Params are captured at module load and stripped from the URL for privacy (see hooks/setup/useSetupUrlParams.ts).

9. Repo map — what lives where

Path	Owner
`system/cmd/os-server/main.go`	OS Server entry point
`system/cmd/bootstrap/main.go`	OTA worker entry point
`system/server/`	HTTP handlers (Gin), organised by domain
`system/`	Business services — agent, device, network, openclaw, hermes, mqtt, statusled, healthwatch, …
`system/domain/`	Shared Go types
`hal/`	Python HAL — drivers, routes, board profiles
`hal/drivers/`	Hardware drivers (rgb, motors, voice, sensing, gpio_button, mic_button, …)
`hal/routes/`	FastAPI routes (voice, led, camera, emotion, scene, music, servo, …)
`robots/contract/`	Frozen HAL capability ABI (`capabilities.md`, `ROBOT-SPEC.md`)
`skills/`	25 built-in skills — agents auto-discover these, including `skill-creator` for authoring and evaluating new skills
`robots/`	Per-device declarations (`intern-v2/`, `lamp/`, `unitree-go2w/`) with `ROBOT.md` + `SOUL.md` + `SAFETY.md`
`scripts/provision/`	Image build + on-device setup scripts
`scripts/release/`	OTA upload scripts (`upload-os-server.sh`, `upload-hal.sh`, `upload-web.sh`)
`docs/`	Docs (this file lives here)

10. Further reading

docs/overview.md — the layered stack, top to bottom
docs/os-server.md — routes, wire graph, run flags
docs/setup-flow.md — provisioning, AP→STA handoff, admin password default
docs/mqtt.md — full MQTT command reference
docs/realtime-voice.md — Gemini Live / OpenAI Realtime / Qwen Omni pipelines
docs/bootstrap-ota.md — OTA metadata format + promotion model
docs/agentic/hermes.md, docs/agentic/picoclaw.md, docs/agentic/codex.md, docs/agentic/opencode.md — per-runtime protocol

11. Getting help & buying a Developer Edition device

Where to buy. The Developer Edition (black case, SSH open, on-device toolchain pre-installed — the unit this whole guide targets) is sold on the official store:

https://www.autonomous.ai/intern?product_url_sequence_code=7496_8

Make sure you pick the Developer Edition SKU at checkout — the yellow OpenClaw and blue Hermes editions on the same product page ship with SSH closed and cannot be turned into a developer unit later.

Support:

GitHub: autonomous-ai/autonomous-os — issues + discussions

Navigation Menu

Uh oh!

Intern Developer Guide

1. Device access

1.1 SSH into the device

1.2 On-device filesystem layout

1.3 Web CLI — browser terminal from the Admin panel

2. Logs

3. Configuration

3.1 LLM / STT / TTS (OS Server — `/root/config/config.json`)

3.2 HAL runtime env (`/opt/hal/.env`)

4. Building + shipping code

4.1 Go — os-server + bootstrap

4.2 Python — HAL

4.3 Web (Setup + Admin UI)

5. Controlling the OTA bootstrap

6. HAL API — TTS / STT / Mic / Speaker / LED

6.1 TTS — make the speaker talk

6.2 STT — read what the mic heard

6.3 Mic — mute / unmute mid-conversation

6.4 Speaker — mute / play WAV / stop

6.5 LED

6.6 Other useful HAL surfaces

7. MQTT — add a new event kind

7.1 Sending an MQTT command from the backend

7.2 Receiving on the device — add a new handler

8. Web UI — Setup and Admin

9. Repo map — what lives where

10. Further reading

11. Getting help & buying a Developer Edition device

Uh oh!

FilesExpand file tree

developer-guide.md

Latest commit

History

developer-guide.md

File metadata and controls

Intern Developer Guide

1. Device access

1.1 SSH into the device

1.2 On-device filesystem layout

1.3 Web CLI — browser terminal from the Admin panel

2. Logs

3. Configuration

3.1 LLM / STT / TTS (OS Server — /root/config/config.json)

3.2 HAL runtime env (/opt/hal/.env)

4. Building + shipping code

4.1 Go — os-server + bootstrap

4.2 Python — HAL

4.3 Web (Setup + Admin UI)

5. Controlling the OTA bootstrap

6. HAL API — TTS / STT / Mic / Speaker / LED

6.1 TTS — make the speaker talk

6.2 STT — read what the mic heard

6.3 Mic — mute / unmute mid-conversation

6.4 Speaker — mute / play WAV / stop

6.5 LED

6.6 Other useful HAL surfaces

7. MQTT — add a new event kind

7.1 Sending an MQTT command from the backend

7.2 Receiving on the device — add a new handler

8. Web UI — Setup and Admin

9. Repo map — what lives where

10. Further reading

11. Getting help & buying a Developer Edition device

3.1 LLM / STT / TTS (OS Server — `/root/config/config.json`)

3.2 HAL runtime env (`/opt/hal/.env`)