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.
Continuous voice
GPT-Live is an experimental voice path with a separately selected Responses
thinking model. Both use the same OpenAI credential. Voice duration and backend
tokens are accounted for separately. The application does not enable cloud
recording (store: false).
Settings and migration
Select the voice, thinking model, reasoning effort and web search in API Keys. Saving creates a pre-migration configuration backup and applies to the next call. A model is never chosen or upgraded automatically. Gemini/Vertex and local realtime remain available. Provider-specific function schemas are translated at the adapter boundary.
Text chat and background work follow the active Agents selection. The OpenAI voice thinking model is also used for voice-triggered computer control without another model key. Connected services retain their own credentials.
Voice setup is a single provider panel with a shared credential, conversation settings and thinking/tool settings. Inspecting a provider does not activate it; the Live profile saves its provider and model together. The separate Tool Model tab and inline computer-control model pickers are removed. Saving only refreshes the affected voice queries and preserves a draft when a key becomes available.
Execution and media
The dodaios/live package separates the provider connection, transcript stream,
tool receipts and execution. GPT-Live uses WebRTC media plus a server-side
control connection, or a primary WebSocket for server audio integrations.
Browser/WebView capture owns echo cancellation; the voice engine does not mute
the microphone merely because it is speaking.
Tool execution remains behind the supervisor gateway and ToolExecutor. The Responses adapter retains complete function calls, sends all required outputs and explicitly continues the response. Discovery keeps tools reachable beyond the initial declaration set. Screenshots enter a vision-capable backend as images, with existing privacy settings enforced at capture.
Operation receipts prevent the same call ID from executing twice. Uncertain results are not retried automatically. Confirmations belong to the application, and already-started work can finish after speech closes. Subscription tasks use the same gateway with a scoped tool grant; unsupported isolation is reported instead of silently using another account.
An explicit task cancellation invalidates queued calls and pending approvals. The next user request gets a new cancellation token, while interrupted work retains its cancelled token. Scheduled turns keep their starting agent selection through settings changes, including nested computer-control calls.
Transcripts keep their original fragments and timestamps. The legacy archive stores one compatibility group at close, not a fabricated provider turn boundary. Voice duration updates are cumulative snapshots. Backend completion, generated speech and actual playback are separate states.
WebRTC wire limits
SDP offers and answers retain their terminal CRLF through browser and server validation. Trimming an otherwise valid offer was reproduced as HTTP 400 from Live, while the unchanged offer was accepted.
The initial tool declarations have both a count and a 24 KB JSON budget. A count
limit alone allowed large imported schemas to overflow a 64 KiB RTC data-channel
message before session.started. Discovery returns pages with next_offset;
the full catalog remains reachable through discovery and call_tool.
Qualification
Synthetic live API probes have exercised tool execution, response continuation
and spoken results on OpenAI and Gemini. Contract tests run on Windows and in a
headless Python 3.11 Linux container. A fresh base wheel installation on
python:3.11-slim boots a healthy API without PortAudio. An isolated OpenAI
synthetic probe using exactly one credential also passes there, including a
forced reconnect and exactly one tool execution. These checks do not establish
native audio parity or release readiness. Native-device audio, fresh desktop
installations, long-session recovery and comparative latency still require qualification.
Safe transport failures reconnect with jitter and a shared connection budget. Recovery restores bounded conversation history and completed tool receipts; it never replays buffered microphone audio. New actions wait for fresh user input. Pending approvals, running actions or uncertain outcomes prevent automatic reconnection. Missing final usage remains marked unconfirmed across reconnects.
Synthetic OpenAI and Gemini probes have recovered from forced connection loss, spoken the previous verified result and kept the tool execution count at one. These probes do not replace long-call or native-device qualification.
Run python scripts/verify_gpt_live.py --run-live for the opt-in OpenAI synthetic
probe. The test uses configured credentials and incurs normal API usage; it
captures neither microphone nor screen data. Gemini is selectable with
--provider gemini --model gemini-3.1-flash-live-preview.
Add --reconnect to verify recovery after a forced connection loss.
The architecture decision and remaining acceptance requirements are recorded in ADR-0036.
The settings form has been inspected in light and dark themes. Its provider
test uses the continuous session contract and waits for session.started;
browser-owned media no longer waits for the legacy desktop offer broker.
The separate Live profile owns model and voice selection without duplicate
legacy pickers on the provider card.
A synthetic check through the running application's authenticated /ws/audio
route now establishes WebRTC, receives session.started, completes the real
read-only describe-app-settings tool through the gateway and receives spoken
confirmation as non-silent RTP audio. It captures no microphone or screen data.
The transport checks do not replace physical device and acoustic-echo tests.
References: OpenAI architecture, delegation, migration, and Gemini tools.