Use NeuronAI from AI assistants
Connect Claude, ChatGPT, Codex and other assistants to generate speech, clone voices and transcribe audio.
What it does
The NeuronAI MCP server lets an AI assistant generate speech, build multi-voice dialogues, clone voices and transcribe audio with your NeuronAI account, without leaving the chat.
It speaks Streamable HTTP. Apps sign in with NeuronAI in the browser, and scripts send an API key.
Each call is permitted, limited and billed like the same call to the API: the same key permissions, product limits and prices, paid from your organization's balance. Its usage shows against the key or the connected app that made it.
It works with Claude on the web, desktop and phone, ChatGPT, Grok, Claude Code, Codex and Gemini CLI, and with any other client that speaks Streamable HTTP.
Try asking
- “List the Uzbek voices.”
- “Read this paragraph aloud with the voice Saodat.”
- “Transcribe ~/Downloads/meeting.mp3.”
Most assistants ask you to confirm before they run a tool that spends credits or changes something. In Claude and ChatGPT, give a link to the recording instead of a file on your computer.
Connect
Choose your assistant and follow its steps. Most sign in with NeuronAI in the browser, and the CLIs can send an API key instead.
Sign in with NeuronAI
The first time, the app opens a NeuronAI page in your browser. Log in if you need to, choose the organization whose credits the app will spend, and press “Allow”. The page shows the app's name and the address it returns you to. An app picks its own name, so trust the address.
The app works with that organization only. To use it with another one, disconnect it and connect it again.
In Claude, open Customize, then Connectors, then Add custom connector.
Paste the server URL, then press Add and Connect.
URLSign in to NeuronAI, choose the organization and press Allow.
Check that it's connected
Ask Claude for the list of Uzbek voices. A list of voices means it works.
If Claude asks how to register, choose Register automatically.
Each tab ends with a way to check that it's connected. Every app you connect is listed in Developers, then MCP, where you can disconnect it.
Use an API key
For scripts, CI and clients that can't sign in through a browser, send an API key in the Authorization: Bearer header instead.
Give the key tts:write, voices:write and stt:write, which cover every tool, and a monthly spend limit.
In Developers, then MCP, go to a CLI tab and set “Sign in with” to “API key”: “Create MCP key” there makes a key with these permissions, and you give it a spend limit in the same form.
In the CLI tabs above, set “Sign in with” to “API key” for the commands that send a key; Grok Build's is under “Other”. Where the client can read the key from the environment, its settings keep only the variable's name, NEURONAI_API_KEY. Claude Code is the exception: its command saves the key itself in ~/.claude.json, so sign Claude Code in with the browser when you can. A line that holds the key, export included, stays in your shell history.
In PowerShell, set the variable with $env:NEURONAI_API_KEY = "nai_your_api_key" instead of export.
Manage the key, its limits and its spend in Developers, then API keys.
Tools
Your assistant reads each tool's name, description and parameters, and decides which to call. They are listed here as it reads them, in English, in the order the server gives them.
An assistant sees only the tools its key or connection may call. Every result carries request_id: quote it when you contact support.
list_voices
Lists the voices you can use.
- Needs
- No permission needed
- Price
- Free
What the assistant reads
Lists the voices you can speak with: NeuronAI's stock voices first, in catalog order, then the organization's own voice clones, newest first, when the API key may read voices. Filter with type (stock or custom) and language (a code such as uz, ru or en). A page holds up to limit voices (1 to 100, default 50); while next_cursor is not null, call again with it as cursor for the next page. Pass a voice's id as voice_id to text_to_speech or text_to_dialogue. Free.Parameters
| Parameter | Description |
|---|---|
typestring | Only stock voices, or only the organization's voice clones. Omit for both. Allowed values: |
languagestring | Only voices in this language, as a code such as uz, ru or en. maxLen: 16 |
limitinteger | How many voices one page holds. min: 1 · max: 100 · default: 50 |
cursorstring | The next_cursor of the previous page. maxLen: 512 |
Result
| Field | Description |
|---|---|
voicesobject[] | |
voices[]string | Pass it as voice_id. |
voices[]string | |
voices[]string | |
voices[]string[] | |
voices[]string | Allowed values: |
voices[]string | Voice clones only. |
voices[]string | null | Voice clones only: what the reference clip says. |
voices[]string | null | Voice clones only. |
next_string | null | Pass it as cursor for the next page; null on the last page. |
request_string |
text_to_speech
Turns text into speech and returns a link to the audio.
- Needs
tts:write- Price
- 0.3 credits per character, 0.5 credits with a voice clone
What the assistant reads
Turns text into speech and returns a link to the audio file, Opus unless format names mp3 or wav. Choose voice_id with list_voices: a stock voice or one of the organization's voice clones. text holds up to 10000 characters, counted as spoken: an emotion tag counts as one. Emotion tags are directions the voice acts out rather than reads: [laughter], [sigh], [dissatisfaction-hnn], [confirmation-en], [question-en], [question-ah], [question-oh], [question-ei], [question-yi], [surprise-ah], [surprise-oh], [surprise-wa], [surprise-yo]. The tool waits up to 40 seconds for the audio. If it is not ready by then, the answer has status queued or running and a job_id: call get_speech with that job_id after about 15 seconds. An identical call within 10 minutes returns the latest result and charges nothing; pass new_take: true only when the user wants a different take. If a call with new_take: true got no answer, repeat it without new_take: that returns the take it made instead of paying for another.
This spends credits from the organization's balance: 0.3 credits per character with a stock voice, 0.5 with a voice clone. Call it only when the user asks, and never call it again to retry a call that succeeded.Parameters
| Parameter | Description |
|---|---|
textstringRequired | The words to speak, up to 10000 characters counted as spoken. minLen: 1 |
voice_string | A voice id from list_voices: a stock voice or one of the organization's voice clones. maxLen: 128 |
languagestring | The language the text is written in. Allowed values: |
qualitystring | fast is quickest, high sounds best; standard is the default. Allowed values: |
speednumber | How fast the voice speaks: 1 is normal, from 0.5 to 2. min: 0.5 · max: 2 |
formatstring | The audio file's format: opus (the default, and the smallest), mp3 or wav (the largest). Allowed values: |
new_boolean | true makes a fresh take of an identical call from the last few minutes, and charges for it. Leave it out otherwise. |
Result
| Field | Description |
|---|---|
statusstring | succeeded: the audio is ready. queued or running: call get_speech with job_id after retry_after_seconds. Allowed values: |
job_string | The speech job, for get_speech. |
audioobject | The audio, as a link and as an MCP resource. Set when status is succeeded. |
audio.string | A link to the audio file. Anyone with the link can play it until expires_at, so give it only to the user. |
audio.string | When the link stops working: at most 24 hours from now. |
audio.string | The same audio as an MCP resource, for a client that cannot open the link: read it with resources/read. Free, up to 40 MB. |
audio.string | opus, mp3 or wav. Allowed values: |
character_integer | Characters billed; an emotion tag counts as one. Set when status is succeeded. |
costnumber | Credits charged. Set when status is succeeded. |
balancenumber | The organization's balance in credits after the charge. Set when status is succeeded. |
retry_integer | Set while status is queued or running: wait this long, then call get_speech. |
replayedboolean | true when this repeats an identical call from the last few minutes, which charged nothing more. |
request_string | This call's request id, for support. |
get_speech
Checks a speech job and returns its link when it's ready.
- Needs
tts:read- Price
- Free
What the assistant reads
Checks a speech job that text_to_speech answered with status queued or running, and returns the link to its audio once it is ready. If it is still not ready, call get_speech again after about 15 seconds. It is free, and it never starts a new job.Parameters
| Parameter | Description |
|---|---|
job_stringRequired | The job_id text_to_speech answered with. minLen: 26 · maxLen: 26 |
Result
| Field | Description |
|---|---|
statusstring | succeeded: the audio is ready. queued or running: call get_speech with job_id after retry_after_seconds. Allowed values: |
job_string | The speech job, for get_speech. |
audioobject | The audio, as a link and as an MCP resource. Set when status is succeeded. |
audio.string | A link to the audio file. Anyone with the link can play it until expires_at, so give it only to the user. |
audio.string | When the link stops working: at most 24 hours from now. |
audio.string | The same audio as an MCP resource, for a client that cannot open the link: read it with resources/read. Free, up to 40 MB. |
audio.string | opus, mp3 or wav. Allowed values: |
character_integer | Characters billed; an emotion tag counts as one. Set when status is succeeded. |
costnumber | Credits charged. Set when status is succeeded. |
balancenumber | The organization's balance in credits after the charge. Set when status is succeeded. |
retry_integer | Set while status is queued or running: wait this long, then call get_speech. |
replayedboolean | true when this repeats an identical call from the last few minutes, which charged nothing more. |
request_string | This call's request id, for support. |
text_to_dialogue
Speaks several lines in different voices as one recording.
- Needs
tts:write- Price
- 0.3 credits per character, 0.5 credits with a voice clone
What the assistant reads
Speaks a script of several lines, each in its own voice, as one recording, and returns a link to it with where each line starts and ends. The recording is Opus unless format names mp3 or wav. turns holds 1 to 100 lines: each has a voice_id from list_voices and its text, and may set its own language, quality and speed; the top-level language, quality and speed apply to every line that sets none. gap_ms is the silence between lines, from 0 to 5000. The whole script holds up to 10000 characters, counted as spoken: an emotion tag counts as one. Emotion tags are directions the voice acts out rather than reads: [laughter], [sigh], [dissatisfaction-hnn], [confirmation-en], [question-en], [question-ah], [question-oh], [question-ei], [question-yi], [surprise-ah], [surprise-oh], [surprise-wa], [surprise-yo]. A long script can take a minute or two. Each line is billed at its own voice's rate. An identical call within 10 minutes returns the latest result and charges nothing; pass new_take: true only when the user wants a different take. If a call with new_take: true got no answer, repeat it without new_take: that returns the take it made instead of paying for another.
This spends credits from the organization's balance: 0.3 credits per character with a stock voice, 0.5 with a voice clone. Call it only when the user asks, and never call it again to retry a call that succeeded.Parameters
| Parameter | Description |
|---|---|
turnsobject[]Required | The script, line by line: 1 to 100 lines. minItems: 1 · maxItems: 100 |
turns[]stringRequired | A voice id from list_voices. maxLen: 128 |
turns[]stringRequired | What this voice says. minLen: 1 |
turns[]string | The language the text is written in. Allowed values: |
turns[]string | fast is quickest, high sounds best; standard is the default. Allowed values: |
turns[]number | How fast the voice speaks: 1 is normal, from 0.5 to 2. min: 0.5 · max: 2 |
languagestring | The language the text is written in. Allowed values: |
qualitystring | fast is quickest, high sounds best; standard is the default. Allowed values: |
speednumber | How fast the voice speaks: 1 is normal, from 0.5 to 2. min: 0.5 · max: 2 |
gap_integer | Silence between lines, in milliseconds. min: 0 · max: 5000 |
formatstring | The audio file's format: opus (the default, and the smallest), mp3 or wav (the largest). Allowed values: |
new_boolean | true makes a fresh take of an identical call from the last few minutes, and charges for it. Leave it out otherwise. |
Result
| Field | Description |
|---|---|
audioobject | The audio, as a link and as an MCP resource. |
audio.string | A link to the audio file. Anyone with the link can play it until expires_at, so give it only to the user. |
audio.string | When the link stops working: at most 24 hours from now. |
audio.string | The same audio as an MCP resource, for a client that cannot open the link: read it with resources/read. Free, up to 40 MB. |
audio.string | opus, mp3 or wav. Allowed values: |
turnsobject[] | Where each line is in the recording, in seconds, so a player can seek to it. |
turns[]integer | |
turns[]string | |
turns[]number | |
turns[]number | |
turns[]number | |
character_integer | Characters billed; an emotion tag counts as one. |
costnumber | Credits charged. |
balancenumber | The organization's balance in credits after the charge. |
replayedboolean | true when this repeats an identical call from the last few minutes, which charged nothing more. |
request_string | This call's request id, for support. |
create_upload
Gives an assistant in your terminal a link to upload an audio file from your computer.
- Needs
stt:writeorvoices:write- Price
- Free
What the assistant reads
Creates a short-lived link to upload an audio file from the user's computer. Purpose transcription is for speech_to_text (mp3, wav, ogg, flac, m4a, webm, up to 25 MB); purpose voice_sample is for clone_voice (wav, mp3, ogg, flac, up to 10 MB). Run the curl command it returns in the user's shell, with the file's path in place of ./FILE, then pass the upload_id to that tool. On Windows, run curl.exe (PowerShell's curl is a different command). The link works for 60 minutes. Without a shell, ask the user for a link to the recording instead and pass it to speech_to_text as audio_url. A voice sample needs a shell to run the curl command, because clone_voice takes no audio_url.Parameters
| Parameter | Description |
|---|---|
purposestringRequired | transcription for speech_to_text, voice_sample for clone_voice. Allowed values: |
Result
| Field | Description |
|---|---|
upload_string | Pass it to the tool the upload is for, once the file is uploaded. |
upload_string | Where to PUT the file. Signed: it is the only credential. |
methodstring | Allowed values: |
expires_string | When the link, and the upload, stop working. |
max_integer | The largest file the link takes. |
formatsstring[] | The formats the file may be. |
curlstring | The command that uploads the file, with ./FILE to replace by its path. |
request_string | This call's request id, for support. |
speech_to_text
Transcribes a recording from a link or an upload.
- Needs
stt:write- Price
- 1 credit per second of audio
What the assistant reads
Transcribes a recording to text. Pass exactly one of upload_id or audio_url: upload_id comes from create_upload with purpose transcription, for a file on the user's computer; audio_url is a public https link to the recording, followed through at most 3 redirects. Formats: mp3, wav, ogg, flac, m4a, webm, up to 25 MB. language is one of uz, kaa, ru, en, kk, tk, tg, tr, az, ja, de, ko, or auto (the default) to detect it. Set include_segments to also get the text in timed segments. An identical call within 10 minutes returns the latest result and charges nothing; pass new_take: true only when the user wants a fresh transcription. If a call with new_take: true got no answer, repeat it without new_take: that returns the take it made instead of paying for another.
This spends credits from the organization's balance: 1 credit per second of audio. Call it only when the user asks, and never call it again to retry a call that succeeded.Parameters
| Parameter | Description |
|---|---|
upload_string | The upload_id from create_upload with purpose transcription, once the file is uploaded. Pass this or audio_url. maxLen: 64 |
audio_string | A public https link to the recording. Pass this or upload_id. maxLen: 2048 |
languagestring | The language spoken, or auto to detect it. Allowed values: default: "auto" |
include_boolean | true adds the text in timed segments. default: false |
new_boolean | true makes a fresh take of an identical call from the last few minutes, and charges for it. Leave it out otherwise. |
Result
| Field | Description |
|---|---|
textstring | What was said. |
languagestring | The language, the detected one for auto. |
duration_number | How long the recording runs. |
segmentsobject[] | The text in timed pieces, when include_segments was true. |
segments[]number | |
segments[]number | |
segments[]string | |
costnumber | What the transcription cost, in credits. |
balancenumber | The organization's balance after the charge, in credits. |
replayedboolean | true when this repeats an identical call from the last few minutes, which charged nothing more. |
request_string | This call's request id, for support. |
clone_voice
Creates a voice clone from an uploaded sample.
- Needs
voices:write- Price
- Free
What the assistant reads
Creates a voice clone from a short recording of one speaker (10 to 15 seconds) that the user has on their computer. Call create_upload with purpose voice_sample, run its curl command, then pass the upload_id; a link (audio_url) is not accepted. The speaker must have agreed to have their voice cloned, and speaker_consent: true confirms it: set it only when the user says the speaker agreed. Samples are wav, mp3, ogg, flac, up to 10 MB; ref_text is required: what the sample says, word for word, because the voice is paced from it. If the user cannot give it, transcribe the sample with speech_to_text and check the text with them. Use the new voice's id as voice_id in text_to_speech and text_to_dialogue. An identical call within 10 minutes answers with the voice it made instead of a second clone.Parameters
| Parameter | Description |
|---|---|
namestringRequired | A name for the voice. maxLen: 120 |
languagestringRequired | The language the sample is spoken in. Allowed values: |
upload_stringRequired | The upload_id from create_upload with purpose voice_sample, once the file is uploaded. maxLen: 64 |
speaker_booleanRequired | true confirms that the speaker agreed to have their voice cloned. Only true is accepted. Allowed values: |
ref_stringRequired | What the sample says, word for word. The voice is paced from it, so it must match the recording. maxLen: 1000 |
tagsstring[] | Optional labels for the voice. |
categorystring | What the voice is for. characters when left out. Allowed values: |
Result
| Field | Description |
|---|---|
voiceobject | The new voice clone. |
voice.string | Use it as voice_id. |
voice.string | |
voice.string | |
voice.string | Allowed values: |
voice.string[] | |
voice.string | |
voice.string | |
replayedboolean | true when this is the voice an identical call made, and no second clone was made. |
request_string | This call's request id, for support. |
delete_voice
Deletes one of your voice clones.
- Needs
voices:write- Price
- Free
What the assistant reads
Deletes one of the organization's voice clones for good, by the id list_voices shows for it. Stock voices cannot be deleted. Confirm with the user first: the voice cannot be brought back, and anything that uses its id stops working. If it answers not_found, the voice is already gone or was never the organization's.Parameters
| Parameter | Description |
|---|---|
voice_stringRequired | The id of the voice clone to delete. maxLen: 191 |
Result
| Field | Description |
|---|---|
deletedboolean | true once the voice clone is gone. |
voice_string | The id of the voice clone that was deleted. |
request_string | This call's request id, for support. |
Audio files
Speech comes back as a link to an Opus file, or MP3 or WAV when the assistant asks for one. Clients that show MCP Apps, such as Claude, also play it in the chat. Open the link in your browser to play or download it.
A link works for up to 24 hours, for anyone who has it, so share it as you would the file. Claude and ChatGPT don't open it for you.
An assistant that can't open links, such as one in a sandbox, can read the same audio over MCP, as the resource its result names. A read is free and carries up to 40 MB.
Files from your computer
To transcribe a recording or clone a voice from a file on your computer, the assistant calls create_upload, runs the curl command it gets back, then passes the upload_id to the tool. On Windows, it runs the command as curl.exe, because curl in PowerShell is a different command.
An upload link works for 60 minutes and takes up to 25 MB for a transcription, or 10 MB for a voice sample. Uploading needs a shell, so it works in the CLIs, not in Claude or ChatGPT. A CLI that runs commands in a sandbox, such as Codex, must let the upload reach the internet.
An organization can have up to 3 uploads in progress at once and get up to 30 upload links an hour.
Recordings on the web
speech_to_text also takes an audio_url: a public https address of the recording. That's how to transcribe in Claude and ChatGPT. The link must be the file itself: a Google Drive or OneDrive share page won't do, and a Dropbox link needs dl=1.
NeuronAI downloads it itself. It follows up to 3 redirects, refuses private and local addresses, and stops at the transcription size limit. When the download fails, the error says why, such as a blocked address or a file that is too large.
Cloning a voice
clone_voice takes an upload, never a link, and speaker_consent: true, which confirms that the speaker agreed to have their voice cloned. In Claude and ChatGPT, clone voices in the dashboard instead.
Costs and limits
Tools cost what the same API calls cost, paid from your organization's credits:
text_to_speech- 0.3 credits per character, 0.5 credits with a voice clone
text_to_dialogue- 0.3 credits per character, 0.5 credits with a voice clone
speech_to_text- 1 credit per second of audio
Every other tool is free.
An identical call within 10 minutes returns the latest result and charges nothing, so a retry never pays twice. Pass new_take: true for a different take, which is charged again.
The server takes up to 120 requests a minute per key or connected app. Each tool also keeps its product's rate and concurrency limits from the API, and a refusal says how long to wait.
text_to_speech waits up to 40 seconds for the audio. A longer text answers with a job_id, and the assistant calls get_speech with it after the suggested wait. A job that fails charges nothing.
Long calls report progress while they work, so the connection stays open. Codex stops a tool after 60 seconds unless you raise tool_timeout_sec, which its tab sets to 180. Give any other client at least 180 seconds.
Security
A connected app works for the one organization chosen when it was connected, and spends that organization's credits. Any member can open Developers, then MCP to see it and disconnect it at once.
Each connected app has its own key, marked “Connected app” on the API keys page. Give it a monthly spend limit there: a page your assistant reads can try to make it call paid tools.
Removing a member from the organization disconnects the apps they connected.
For scripts, create a key just for MCP, restricted to tts:write, voices:write and stt:write. Never commit it: keep it in an environment variable.
Clone a voice only when its speaker has agreed. clone_voice takes only an upload and asks you to confirm the agreement with speaker_consent: true.
Texts, transcripts and audio links pass through your AI provider, under its terms. Don't send anything you wouldn't share with it.
Troubleshooting
| Problem |
|---|
The NeuronAI page never opens, or it shows “This sign-in link doesn't work” Start connecting again from the app. If it keeps happening, the app may not support NeuronAI yet: connect it with an API key instead, and tell us which app it is. |
Claude asks how to register Choose “Register automatically”. |
The app spends the wrong organization's credits Open Developers, then MCP, disconnect the app, then connect it again and choose the right organization. |
ChatGPT has no “Add custom MCP server”, or can only read Every tool works in ChatGPT Business, Enterprise and Edu, on the web, and on Business an admin adds the server. ChatGPT Pro can only list voices and check speech jobs. Plus, Go and Free don't offer custom MCP servers. |
Gemini CLI doesn't open the browser Run |
A tool is missing, or answers The key or the connection lacks the tool's permission, and an assistant sees only the tools it may call. Give the key the permission the message names. |
Long calls time out Give the client a tool timeout of at least 180 seconds, as the Codex tab does. An identical call within 10 minutes returns the latest result without charging again. |
The client sends An app opened from the Dock or the Start menu doesn't see variables you set in a terminal. Open it from a terminal where |
Claude Code saves a long transcript to a file Claude Code keeps a tool result over 50,000 characters in a file and shows the model only part of it. Ask it to read the file, and leave |
When the key or the sign-in is refused
The server answers these before any tool runs, and the app usually shows the code.
| Code | What to do |
|---|---|
invalid_401 | The key is unknown, or the app was disconnected or its sign-in has ended. Check the key, or connect the app again. |
api_401 | The key was revoked. Create a new key, or connect the app again. |
api_401 | The key is switched off. Switch it back on in Developers, then API keys. |
api_401 | The key is past its expiry. Open Developers, then API keys and give it a later expiry date, or create a new key. |
ip_403 | The call came from an address outside the key's IP allowlist. Add the address, or call from a listed one. Claude and ChatGPT call from their own servers, so a connected app's key needs an empty allowlist. |