Dictate the prompt, not the code

Capture the audio, send it to a dictation API, and paste the returned text into the agent. The part that matters is what sits in between: a verbatim transcript carries your false starts and filler into the prompt, so the Dictation API applies an LLM cleanup pass in the same call and returns the instruction you meant. Blurt is a free, MIT-licensed macOS app that does exactly this — hold a key, talk, and finished text lands in whatever has focus, including a terminal running an agent.

You dictate the prompt. This is not a speech-to-code tool that converts spoken words into literal syntax or editor commands — those map utterances to keystrokes and need you to learn their grammar. Here the agent still writes the code; you are just describing what you want in prose, which is the input coding agents already take. Nothing about the request is agent-specific.

Because transcription models are built to be faithful, and speech is full of things you did not mean to send. Disfluencies, repetitions and self-corrections — "in the the auth module, no wait, the middleware" — are a correct transcript and a misleading instruction. Spoken numbers also arrive as words rather than digits unless something normalizes them. The cleanup pass resolves the correction to what you landed on, drops the filler, and applies the output shape you asked for.

Pass them as keyterms, which biases recognition toward names the model would otherwise spell phonetically — the same idea other vendors call phrase hints or contextual biasing. The Dictation API takes up to 100 terms per request in keyterms_prompt, included in the rate. You can also add an llm_instruction telling the rewrite to keep file, module and library names exactly as spoken, so cleanup does not "correct" them.

Those return a transcript, so you still own the second step: your own LLM call to turn speech into a usable prompt, plus the prompt engineering, latency and failure handling that comes with it. The Dictation API does transcription and the rewrite in one request, with the rewrite instruction under your control. On accuracy, Universal-3.5 Pro ranks first across independent benchmarks and posts a mean of 3.87% normalized word error rate on short-form audio — the length dictated prompts actually are.

One rate covers the whole request — transcription and the LLM rewrite together, billed on audio length rather than tokens, so there is no second model to forecast. Keyterms, prompting and output instructions are all included rather than priced as add-ons. Current rates are on the pricing page and the Dictation API product page.

The rewrite is best-effort on a five-second deadline, and past it the API returns HTTP 200 with the verbatim transcript and llm_error: "timeout" — not an error your client has to handle. In the Python SDK, result.final_text does the fallback for you. Design for plainer text in the worst case, not for an error path.

Any of them, because the API returns finished text and knows nothing about what happens next. It sits in front of Claude Code, Cursor or GitHub Copilot equally well, and in front of orchestration layers like Anthropic’s Claude Agent SDK and OpenRouter. The integration point is wherever your tool accepts a text prompt.