bullshitai.dev
unofficial server open source, MIT not an OpenAI product

Codex Voice

ChatGPT transcription and voice, through your own server.

An unofficial Rust server that uses your existing Codex ChatGPT subscription login for audio transcription and voice conversations. It reproduces the reverse-engineered flow the ChatGPT app uses, through private ChatGPT services.

  • The server is the product; its browser demo is an optional way to test your server and account
  • Your audio still goes to private ChatGPT services, not OpenAI's public Realtime API; routes can change
  • Source on GitHub; no tags, binaries or images released yet
How one gateway serves both branchesfig. 1 · overview
OverviewYour client and the codex-voice gateway run on your machine. Transcriptions: the client sends an audio file to the gateway, which forwards it upstream and returns text. Realtime: the client sends a WebRTC offer to the gateway, which negotiates with upstream and returns an answer; audio and events then flow directly between the client and upstream over WebRTC, not through the gateway.on your machine · audio leaves at its edgeyour clientapp, script or browserA · audio in, text outB · offer / answercodex-voice gatewayRust, runs locally,reuses your Codex loginA · forwardedB · negotiateupstreamprivate ChatGPT serviceB · audio +events flowdirectly:client andupstream(WebRTC)
  • HTTP through the local gateway
  • WebRTC media, not via the gateway
  • your machine; arrows leaving it carry audio
  • A · Transcriptions
  • B · Realtime

Transcriptionsone audio file in, its text out

Send a recorded file as a multipart form. The gateway passes it upstream with your Codex login and returns the text in the format you asked for.

POST /v1/audio/transcriptions current route · multipart
Form fields
fieldwhat happens
filerequiredThe audio to transcribe. It can't be empty.
languageA hint, forwarded upstream. Without it the server's --language applies.
response_formatjson (default) text verbose_json. Not supported: srt, vtt.
modelAccepted so existing clients don't break, then ignored. It doesn't pick the backend.

The 10 MiB default limit covers the whole multipart body, not just the file.

requestREADME example
# local server key, not your Codex token
curl --fail-with-body http://127.0.0.1:2600/v1/audio/transcriptions \
  -H "Authorization: Bearer $CODEX_VOICE_API_KEY" \
  -F 'file=@recording.wav;type=audio/wav' \
  -F language=en -F response_format=json

127.0.0.1:2600 is the default address. Leave out the header only on a server started with --no-api-key.

response · jsonshape + example text
{"text":"Recognized speech."}

The README's example wording, not a recorded output.

Realtimea live voice session your client owns

The gateway only does the handshake. Your client sends a WebRTC offer, gets an answer back, and from then on talks to upstream directly: microphone, playback and the data channel are all yours.

Handshake, then direct mediafig. 2 · sequence
Realtime session sequenceSteps one to six, described in the numbered list below the figure.clientgatewayupstream1make offer2POST offer3negotiate4answer + session_id5audio + data channel, direct6Stop: client cleans up
  1. clientCreates a WebRTC offer with its microphone track.
  2. client → gatewayPOSTs the offer as JSON with the local key. No Codex token is ever handed to the browser.
  3. gateway ↔ upstreamNegotiates with upstream using your local Codex login. Upstream picks the voice backend.
  4. gateway → clientReturns the SDP answer and a session_id for correlation.
  5. client ↔ upstreamAudio and events flow over WebRTC. They never come back as REST JSON, and the gateway does not relay media.
  6. clientStop means the client closes the connection and releases the mic. There are no pause, resume or delete endpoints.
POST /v1/realtime/sessions private JSON · subject to change

The old /v1/voice/* routes are gone; this is the only signalling route.

request · jsonshape, not runnable
{
  "sdp": "<your peer's complete SDP>",
  "mode": "standard",
  "voice": "cove",
  "history_and_training_disabled": true
}

sdp is a placeholder; a real offer comes from your client's peer connection. Unknown fields are rejected.

response · jsonkeys from the README
sdp
The SDP answer. Apply it to the peer that made the offer.
session_id
A correlation id for logs. It isn't a handle for controlling the session.
Request fields
fieldvalues · default
sdprequiredYour complete WebRTC audio offer, up to 128 KiB.
modestandard advanced wingman. default advanced. The demo starts on Wingman with saving off, and asks you to turn saving on before a Wingman or Advanced Start. A mode picks private service behaviour, not a verified model.
voicedefault cove
languageOptional, such as en-US. Left out, the language is detected.
enable_message_streamingdefault true: user and assistant text events.
history_and_training_disableddefault true. false lets ChatGPT save the conversation. The JSON route forwards either value for every mode; only the demo requires saving for Advanced and Wingman.
timezone, timezone_offset_minutesdefault UTC, 0

Where things liveoperational boundaries

on your machine

your client

  • owns the WebRTC connection, mic, playback and data channel
  • handles Stop and cleanup
holds: the local server key

codex-voice gateway

  • transcription requests and realtime signalling
  • never relays realtime media
  • 127.0.0.1:2600, one upstream request at a time by default
holds: your Codex login (existing, local)

upstream

  • private ChatGPT service, reached with your Codex login
  • chooses the voice backend
  • its retention terms are its own; this project makes no promise about them
Two separate keys: the local server key is yours to give a client; the Codex login stays inside the gateway and never reaches the browser.

Not claimed on this page

  • official OpenAI product
  • OpenAI's Realtime API or GPT-Live
  • stable public API
  • a live Hermes or Discord voice bridge
  • a packaged release
  • free or unlimited
  • fully local or offline
  • measured speed or quality
  • zero retention

Limitsthe not-unlimited part

Codex Voice doesn't add minutes. Your ChatGPT plan has published Voice allowances, and the gateway doesn't get around your account's limits.

ChatGPT Voice in Chat, per OpenAI's Help Center (accessed 20 October 2026)
planLive usage, rolling 24 hours
Go3 hours with GPT-Live-1.1 Quip
Plus3 hours with GPT-Live-1.1
Pro ($100/month)15 hours with GPT-Live-1.1
Pro ($200/month)Unlimited GPT-Live-1.1, subject to abuse guardrails

15 hours talking. Eight hours sleeping. One hour left for bathroom breaks. Surprisingly considerate.Pro ($100) arithmetic on a nominal day, not a reset time or a promise.

  • These are OpenAI's published allowances for Voice in ChatGPT, not measured Codex Voice quotas. The private service behind the gateway may count, limit or reject sessions in its own way.
  • The window rolls over 24 hours. Nothing resets at midnight, and ChatGPT tells you when you reach a limit.
  • One Voice conversation at a time. A conversation can end at a usage limit, at the maximum session length, or when a long one reaches its context limit.
  • They don't cover file transcription (lane A) and they aren't OpenAI's public Realtime API pricing.
  • Codex Voice has no usage meter or countdown. When upstream says no, you get an error (often 429).
  • Workspace plans: Business Standard and Business Premium get 3 and 15 hours, then 1.25 credits per minute. Voice in the desktop app has separate pricing and limits.

Development status

  1. Open source on GitHub main 1af74b1

    The server, CLI, Rust library and optional browser demo, MIT and derived from codex-asr. Build it from source.

  2. File STT in production 4 Hermes gateways

    Dima's four Hermes gateways now send voice-note transcription to Codex Voice. Each passed a real public Ogg/Opus sample with LLM post-processing, with no fallback or errors.

  3. Demo accepted by its owner

    The project's owner accepted the working browser demo.

  4. Wider voice acceptance next

    Every mode in every browser, and a live Hermes or Discord voice bridge, are not done. File STT is a different, narrower path.

  5. Release not yet

    No tags, binaries or images yet. Upstream codex-asr releases are not Codex Voice builds.

Public, early.

File transcription runs in production for its author, and the demo works. Wider voice acceptance and packaged releases are still to come. This page shows how the project works; it isn't test evidence.

Source github.com/dmmeteo/codex-voice

Docs README HTTP API configuration voice guide

Codex Voice is an unofficial, independent project. It is not affiliated with or endorsed by OpenAI, and it doesn't add plan access or lift usage limits. Its code is MIT, derived from codex-asr, with the inherited copyright notice kept in its licence.

Facts on this page come from the project's README and its API, configuration and voice docs at 1af74b1. Request blocks follow them; placeholders are marked. Type on this page: Excalifont (Excalidraw) and Geist Mono (Vercel), both SIL OFL-1.1.