Customize Your Intern

The Developer Edition lets you change which LLM thinks, which agent runtime runs the conversation, how the Intern sounds and what it hears. This section covers each piece:

Page	What you change
LLM	Base URL, API key and model of the brain; bring your own provider
Agent Runtime	OpenClaw, Hermes, Claude Code, Codex, OpenCode or PicoClaw
TTS	Voice provider (ElevenLabs, OpenAI, Gemini, offline Piper), voice, speed
STT & Language	Speech-to-text provider and conversation language
Listening & Wake Word	VAD sensitivity, turn-taking, wake word, button, audio devices, volume
Realtime Voice	Speech-to-speech mode: Gemini Live, OpenAI Realtime, GPT-Live

Where settings live

File (on device)	Owned by	Holds	Applied
/root/config/config.json	os-server	LLM, agent runtime, TTS/STT, realtime, wake word	Through the API/UI: live or with an automatic restart. Hand edits: sudo systemctl restart os-server
/opt/hal/.env	HAL	Low-level HAL_* tuning — audio devices, VAD, echo, realtime overrides	Read once at start: sudo systemctl restart hal

HAL reads config.json too — TTS, STT, wake word and the realtime block are stored there by os-server and read by HAL at start. When a HAL_* variable in /opt/hal/.env and a config.json value cover the same knob, the env var wins.

Edit config.json through the API, not by hand

The API validates input, normalises URLs, applies most changes live and restarts the right service for you. Hand-editing works, but some changes (such as the agent runtime switch) only take effect through the API — see each page.

Three ways to change a setting

1. Admin Web UI

Open http://<device-ip>/setting and sign in with the admin password (the XXXX from the Intern-XXXX sticker — see Device Access).

The public menu shows General, Wi-Fi, My Voice, MCP Tools, Plugins, Timezone, Scheduled and Connectors. The developer tabs — AI Brain, Runtime, Language, Voice, Realtime (plus Channels and MQTT) — only appear in debug mode. Add ?debug=true before the #:

http://<device-ip>/setting?debug=true#llm        AI Brain
http://<device-ip>/setting?debug=true#runtime    Runtime
http://<device-ip>/setting?debug=true#stt        Language (STT)
http://<device-ip>/setting?debug=true#tts        Voice (TTS)
http://<device-ip>/setting?debug=true#realtime   Realtime
http://<device-ip>/setting#general               General (wake word, admin password)

Without ?debug=true, a deep link to a developer tab silently falls back to General.

2. HTTP API

The os-server API is served by nginx at http://<device-ip>/api/…. Every settings endpoint is admin-gated. Log in once with the admin password and reuse the session token as a Bearer token:

DEVICE=http://<device-ip>
TOKEN=$(curl -s -X POST $DEVICE/api/login \
  -H 'Content-Type: application/json' \
  -d '{"password":"XXXX"}' | jq -r .data.token)

# Read the current config (secrets are redacted to has_* flags)
curl -s -H "Authorization: Bearer $TOKEN" $DEVICE/api/device/config | jq .data

os-server also accepts the device's current llm_api_key as a Bearer token. That means changing llm_api_key also changes that credential — prefer the login token in scripts.

PUT /api/device/config has PATCH semantics: send only the fields you want to change; an empty string means "leave unchanged".

curl -s -X PUT $DEVICE/api/device/config \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"tts_voice":"Brian","tts_speed":1.1}'

os-server responses use the envelope {"status": 1, "data": …} on success and {"status": 0, "message": "…"} on failure.

The HAL API is also reachable from your laptop through os-server at http://<device-ip>/api/hardware/<hal-path> with the same Bearer token (on the device itself, call HAL directly at http://127.0.0.1:5001).

3. Edit files over SSH

sudo nano /root/config/config.json && sudo systemctl restart os-server
sudo nano /opt/hal/.env            && sudo systemctl restart hal

config.json is plain JSON — a syntax error stops os-server from starting. Check with sudo jq . /root/config/config.json before restarting.

What applies live and what restarts

Change (via API/UI)	Effect
llm_model	Live — the runtime's primary model is swapped
llm_api_key, llm_base_url, llm_disable_thinking	Runtime config is regenerated and the gateway restarts; HAL restarts for key/URL changes
tts_provider, tts_voice, tts_speed, tts_api_key, tts_base_url	Live — pushed to HAL without a restart
stt_language, stt_*, deepgram_api_key, wakeword, realtime	HAL restarts automatically (a few seconds of no voice)
Agent runtime switch	Runs the switch procedure, then restarts os-server
Anything in /opt/hal/.env	Needs sudo systemctl restart hal

You can also restart services from the API: POST /api/system/restart/hal, POST /api/system/restart/os-server, or POST /api/agent/restart for the active agent runtime.

Going back to the Autonomous defaults

The device ships pointed at the Autonomous AI proxy. The first time you change the LLM, voice or realtime credentials, os-server snapshots the shipped values into autonomous_defaults in config.json. To go back:

curl -s -X POST $DEVICE/api/device/restore-defaults \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"section":"llm"}'        # or "voice" / "realtime"

A factory reset (hold the button 10 s, or soft-reset from the Admin panel) wipes config.json entirely.

What survives an OTA update

The bootstrap service installs official updates automatically (Controlling the OTA Bootstrap). Know where to put your changes so they aren't overwritten:

Item	Path on device	After an OTA
os-server config	/root/config/config.json	Kept
HAL env	/opt/hal/.env	Kept by HAL updates, replaced by device-profile updates
Web UI	/usr/share/nginx/html/setup/	Replaced

If you customise anything in the "Replaced" rows, keep a copy and re-apply it after updates — or disable the bootstrap service while you develop.