Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

11.5 KB
Raw Blame
# ============================================================
#  Dodaios — configuration TEMPLATE
# ------------------------------------------------------------
#  This file is OPTIONAL. Dodaios boots with no config file at all
#  (every value below has a built-in default in dodaios/core/config.py).
#  Copy this to `dodaios.toml` only if you want to override defaults
#  (an install that already has a `jarvis.toml` keeps using that file):
#      cp dodaios.toml.example dodaios.toml
#
#  SECRETS DO NOT GO HERE. API keys live in your OS credential manager
#  (run `python -m dodaios --wizard`) or in `.env` (see .env.example).
#  Never commit a real jarvis.toml — it is gitignored.
#
#  Overrides cascade: jarvis.toml  →  ENV (JARVIS__SECTION__KEY=...)
#  DODAIOS__SECTION__KEY=... is the same override; if both are set, it wins.
#  Any section you delete falls back to the code defaults.
#  This template ships cloud-first defaults (no GPU, no local model,
#  no microphone required) so it runs on a 1 vCPU / 1 GB VPS.
# ============================================================

[profile]
name = "default"
language = "auto"            # "de" | "en" | "auto" (bilingual auto-detect)

[persona]
# The assistant's name. Empty -> derived from your chosen wake-word phrase
# below (e.g. a phrase like "Hey Nova" gives the name "Nova"). Leave blank
# to keep the default. The wake word is always YOUR choice — nothing is
# preset for you.
name = ""

# ── Voice trigger (optional; irrelevant on a headless/browser-only host) ──
[trigger]
wake_word_enabled = true     # always-on wake listener (needs a mic)
single_turn_mode = false     # false = conversation stays open until "hang up"/idle
hotkey_pet_toggle = "alt+win+p"  # hide / show the desktop pet (macOS: cmd+alt+p); "" = off

[trigger.wake_word]
# Type any phrase you like — no wake word ships with the product (design
# 2026-07-07). "auto" resolves the best generic engine for your phrase:
# your own trained .onnx -> any-word Vosk keyword spotting -> local-Whisper
# transcript match -> honest hotkey-only mode with a clear message.
# Leave blank to pick your word during onboarding.
phrase = ""
engine = "auto"              # auto | openwakeword | vosk_kws | stt_match | custom_onnx
sensitivity = 0.5            # 0.5..1 (values below 0.5 are lifted — the wake would be deaf)

# ── Speech-to-Text ──
[stt]
provider = "groq-api"        # groq-api | openai-api | openrouter-stt | gemini-api | vertex-stt | faster-whisper (local)
device = "cpu"               # cpu | cuda — keep "cpu" unless you have a local GPU
language = "auto"            # auto, or a weak ISO-639 preference such as "fr"
bias_prompt = ""             # optional comma-separated vocab to stabilize proper nouns
temperature = 0.0            # deterministic when the selected model supports it

# Cloud model pins are provider-specific; an omitted pin uses that plugin's
# default. Configure these in the app's model picker in normal use.
[stt.models]
# openai-api = "gpt-4o-transcribe"
# openrouter-stt = "openai/gpt-4o-transcribe"

# ── Dictation quality ──
[dictation]
# Short segments update the live preview. Delivery comes from a second pass
# over the complete recording in long overlapping windows.
final_quality_pass = true
final_window_seconds = 25.0
final_overlap_seconds = 1.5
# Auto-detect each final window so language changes remain in the spoken
# language. Turn this off only to apply the preference as a hard monolingual pin.
code_switching = true

# ── Text-to-Speech ──
# Cloud default = Gemini Flash TTS (reuses your GEMINI_API_KEY). Alternatives:
#   provider = "cartesia"    (Sonic 3.5, 42 languages — needs CARTESIA_API_KEY + [tts.cartesia] voice ids)
#   provider = "grok-voice"  (reuses GROK_API_KEY)
#   provider = "elevenlabs"  (needs ELEVENLABS_API_KEY)
[tts]
provider = "gemini-flash-tts"
fallback = "grok-voice"
model = "gemini-3.1-flash-tts-preview"
voice_de = "Charon"
voice_en = "Charon"
language_code = "auto"
streaming = true
# Legacy TTS-only Vertex route (predates the [google] section below). Kept for
# existing installs; new setups should pick provider = "vertex-tts" instead, which
# shares ONE Vertex credential with the brain, STT and realtime tiers.
use_vertex = false
vertex_project = ""
service_account_path = "~/.config/jarvis/vertex-sa.json"

# ── Google endpoint routing (AI Studio vs Google Cloud Vertex AI) ──
# vertex_mode decides where an AMBIGUOUS key stored in a Gemini slot goes:
#   "auto"   probe once per process and remember (default)
#   "always" force Vertex   |   "never" force AI Studio
# The vertex_* fields below configure the dedicated Vertex provider family
# (brain "vertex", stt "vertex-stt", tts "vertex-tts", realtime "vertex-live",
# and the same slug as a subagent). Leave vertex_project empty to authenticate
# with a Vertex API key; set it to use a Google Cloud project with Application
# Default Credentials and no key at all.
# Note that Vertex takes an API key ONLY in express mode (AQ.). A normal Cloud
# project authenticates via Application Default Credentials — no key at all:
#   gcloud auth application-default login
[google]
vertex_mode = "auto"
vertex_project = ""
# "global" carries the current Gemini generation; a named region (europe-west4,
# us-central1, …) is what data residency needs — and regions lag on new models.
vertex_location = "global"
# The Live/realtime socket is a DIFFERENT endpoint: "global" serves no Live
# session at all, so realtime needs a region. Empty = follow vertex_location
# when it names a region, else fall back to a region and log which one.
vertex_realtime_location = ""
service_account_path = ""     # only when this machine is not signed in with gcloud

# ── Brain (multi-provider; never hardcode a single vendor) ──
# Cloud-reachable default = Gemini (needs only GEMINI_API_KEY). Switch the
# whole stack by changing these four to "claude-api" | "openai" | "grok" |
# "openrouter" | "vertex" and supplying that provider's key.
[brain]
primary = "gemini"
fallback = "gemini"
deep_brain = "gemini"
routing_provider = "gemini"
routing_model = "gemini-3.5-flash"

[brain.providers.gemini]
model = "gemini-3.5-flash"          # fast router/voice tier
deep_model = "gemini-3.1-pro-preview"  # heavy worker/critic tier

[brain.providers.claude-api]
auth_mode = "api_key"               # uses ANTHROPIC_API_KEY
model = "claude-haiku-4-5-20251001"
deep_model = "claude-opus-4-8"

[brain.providers.openai]
model = "gpt-5.5"
deep_model = "gpt-5.5-pro"

[brain.providers.grok]
model = "grok-4.3"
deep_model = "grok-4.3"

[brain.providers.openrouter]
model = "anthropic/claude-opus-4.8"
base_url = "https://openrouter.ai/api/v1"

# Optional current local preset. Keep the cloud-first [brain] defaults above
# unless you intentionally run this server. Start llama.cpp with the matching
# stable alias (cross-platform commands are in the local-AI guide), then make
# "local-openai" the active Brain in-app. Keep a ready provider from another
# family in the fallback chain (AP-22).
[brain.providers.local-openai]
model = "gemma-4-12b-it-qat-q4_0"
base_url = "http://127.0.0.1:8080"

# ── Safety: anti-confirmation-fatigue whitelist + hard blacklist ──
[safety]
approval_mode = "bypass"
default_tier = "safe"
always_confirm_tiers = ["ask"]
always_block_tiers = ["block"]

[safety.whitelist]
# fnmatch patterns against "<tool_name> <args>" or the bare "<args>"; a match
# downgrades to "safe". Do NOT add "run_shell *": shell commands come with a
# tiered model out of the box (Claude-Code permission style) — commands that
# only read or modify run without confirmation, destructive commands (rm,
# del, format, Remove-Item, ...) ask first UNLESS the user's own request
# already named the destruction ("delete the folder X"). A blanket
# "run_shell *" entry would switch all of that off.
commands = [
    "open_app *", "type_text *", "search_web *", "remember *",
    "dispatch_to_harness *", "screenshot *",
    "git status", "git log *", "git diff *",
]

[safety.blacklist]
commands = ["format *", "rm -rf /", "del /f /s /q C:\\", "shutdown /p", "reg delete HKLM\\*"]

# ── Audio I/O ──
[audio]
input_device = "auto"        # "auto" | shortest-unique device-name token (e.g. "PRO X")
output_device = "auto"
echo_cancellation = true       # requested; raw PortAudio reports unavailable honestly
sample_rate = 16000          # mandatory 16 kHz (AP-L-16)

# ── Browser voice (hosts without a microphone: a VPS, `jarvis serve`) ──
# The browser that presses Start holds the call over /ws/audio. Realtime mode
# always serves it; this switch covers the classic speech-to-text -> brain ->
# text-to-speech bridge in pipeline mode. See docs/headless-vps-deployment.md.
[browser_voice]
enabled = true

# ── UI / tray ──
[ui]
tray_enabled = true
admin_api_port = 47821
startup_chime = true
# On-screen overlay shown by default on a desktop install: the slim
# "jarvis_bar" (the Dodaios Bar). Alternatives: "mascot" (the hooded figure
# of the logo, as a desktop pet), "voice_orb" (the procedural voice orb),
# "pet" (the desktop pet chosen under My Pets) or "none".
# This mirrors the built-in default — listed here so the bar is the obvious,
# self-documented out-of-the-box choice and never silently disappears.
orb_style = "jarvis_bar"
bar_persistent = true        # keep the bar visible at all times (not only during a voice session)
pet_id = "gigi"              # desktop pet (orb_style = "pet"): built-in id, your own u… id, or "none"
pet_scale = 1.0              # pet size multiplier, 0.5–2.0
pet_bubble = true            # show the pet's status bubble (what Dodaios is doing)
pet_strip_always = false     # true: the pet's buttons always show; false: only while active or hovered

# ── Memory ──
[memory]
recall_store = "sqlite"
archival_store = "sqlite"   # chromadb was removed; sqlite is the supported backend
data_dir = "./data"          # per-user runtime data (gitignored)

# ── Knowledge wiki (Obsidian vault) — self-creates an empty vault on first run ──
[wiki_integration]
enabled = true
vault_root = "wiki/obsidian-vault"
# Runaway guard: the most model calls the wiki may make on its own per day
# (it counts calls, it never shortens a prompt or an answer).
# max_background_llm_calls_per_day = 200

# ── Login autostart — ON by default; toggle in Settings ──
# Launches the full desktop app (voice + Orb) at login so the wake listener
# starts right after a reboot. On a headless host (no display) this is a
# graceful no-op. Set to false to opt out, or use the "App settings" toggle.
[autostart]
enabled = true

# Phase-5/6 features (vision, computer_use, admin, missions, board, telephony)
# default to enabled = false in code. Enable them individually once configured;
# see dodaios/core/config.py for every available section and its defaults.

# -- Mission workers: tool pre-authorization (ADR-0031) --
# Tools listed here are auto-approved for a mission's OWN supervisor-granted
# calls (the gate is answered on the event bus with a full audit trail, never
# bypassed). Entries are exact tool names, or an MCP server family written
# with a trailing slash (e.g. "github/"). Prefer exact names: a family prefix
# silently authorizes every future tool that server adds. Keep messaging /
# mail / social-send tools OUT unless you explicitly accept unattended sends.
# Set worker_tool_auto_approve = false to restore "every ask-tier call waits
# for a human" behavior.
[phase6.safety]
worker_tool_auto_approve = true
auto_approve_tool_families = [
  "search_web",
  "wiki-list",
  "wiki-recall",
  "wiki-page-read",
  "wiki-ingest",
  "session-latest-turn",
]