Loquela Remote

Turn speech into text, on a computer we own.

An OpenAI-compatible API and an MCP server for English audio. Recordings are held in memory while they are transcribed, then erased. We don't keep them and we don't train on them.

Get an API key Read the docs
memo.wav, 5.8 seconds answered in 0.34 s

Remind me to send a contract to Dana on Thursday, and book the small room for 10.

A real response with word timings. The voice is synthetic, and the model heard "a contract" where it said "the contract". Point at a word to see where it falls.

From your code

The API speaks OpenAI's transcription format, so an OpenAI SDK works once you change the base URL. You get json, text, srt, vtt, or verbose_json with word timings like the ones above. Every option is in the API reference.

from openai import OpenAI

client = OpenAI(base_url="https://api.getloquela.com/v1",
                api_key="lqr_...")
with open("memo.wav", "rb") as f:
    text = client.audio.transcriptions.create(
        model="loquela-remote-beta", file=f).text

From an AI assistant

Add the MCP server to Claude, ChatGPT, Cursor or your own agent. It can transcribe a link, a short clip, or a file you send to a one-time upload link, from the same minutes as the API. The docs show the setup for each assistant.

claude mcp add --transport http loquela \
  https://api.getloquela.com/mcp

What happens to a recording

  1. It travels encrypted to our computer in North Carolina.
  2. It is held in memory and never written to disk.
  3. Our own hardware turns it into text.
  4. You get the text back and the audio is erased.

Each step is explained in more detail on How your audio is handled.

Pricing

Starter is $9.99 a month for 4,000 minutes. Pro and Business are founding plans with a limited number of places. Or buy minutes as you need them at $0.003 each, from $10; they last two years. When your minutes run out, requests stop, and nothing is charged unless you turn on automatic top-up. The pricing page compares the plans.

In our own run on LibriSpeech test-clean, the model got 1.8% of words wrong.