Use your own API key to draft music prompts with an OpenAI-compatible model
MusicPrompt’s optional AI assistant drafts text and musical choices. Audio generation remains a separate step. You choose the provider, model and API credentials, and the local compiler still works without them.
Configure a connection you control
In Studio, open AI writing assistant below the editor. Select OpenAI, DeepSeek or an OpenAI-compatible service. Enter a Chat Completions model ID that your account can access. Model names are not guessed from a pricing tier or a marketing announcement.
Use the API base URL supplied by your provider, for example a URL ending in /v1. The relay appends /chat/completions; it also accepts an address that already ends with that path. The hosted relay accepts public HTTPS endpoints on port 443. It does not connect to localhost, private networks or URLs containing embedded credentials.
Enter your key in the password field. Changing the provider clears the previous key, and Clear key removes it from the current form. Reloading the page clears the session value. Do not use a key for an unrelated account or paste it into the creative brief.
Pick the smallest useful writing task
Use Suggest sound choices only when the words already work. It preserves the title and lyrics. Use Rewrite lyrics and structure only when you want to keep the sound. Draft from the brief and Revise the full draft can change the title, language, sound, lyrics and exclusions.
Write a concrete direction such as “keep the chorus words, shorten the verses and remove decorative fills.” Open the data preview before generating if you want to inspect the draft. The assistant sends the current title, brief, musical choices, lyrics and your instruction, not the project’s listening history.
The request passes through this site to the endpoint you configured. Your key stays in page memory and is not written to projects, backups, share links or analytics. The application does not store request bodies. The provider’s own billing and data policies still apply. Only use an endpoint you trust with the key and draft.
Review before adopting a suggestion
The preview shows changed fields alongside the previous text. The result must match MusicPrompt’s schema and pass blocking checks before it can become a new version. A sound-only suggestion can be saved before you write vocal lyrics; you will still need lyrics before exporting a vocal prompt. Applying a suggestion preserves the original version. If you edit the draft or switch projects while a request is pending, the old suggestion is discarded to avoid overwriting new work.
Cancellation stops waiting and aborts the request here; it cannot guarantee that the provider has not already processed or billed it. A timeout is not a reason for an automatic retry.
Resolve compatibility problems deliberately
The default format is JSON mode. For a service that does not support it, choose prompt-only JSON in the compatibility options. Both paths validate the returned object. The output budget is 4,096 tokens; a model that spends many tokens reasoning may need a shorter task or a different configuration.
The output-token parameter can be max_tokens or max_completion_tokens. Choose the one documented by your provider. OpenAI describes JSON response handling and credential handling; DeepSeek documents its JSON output mode. A rejected key, rate limit or truncated output leaves the project unchanged.
For local use, run npm run dev from the application directory. Vercel deployments include api/assist.js; copying only static HTML to another host will not provide that endpoint. Developers can also use the agent workflow with credentials supplied through the process environment.