Tooling
The transcript is the interface between audio and models
No LLM reads audio directly, not really - it reads text about audio, or a transcript underneath it. Once you accept that, the transcript stops being an output you glance at and becomes the API surface between spoken audio and whatever model or agent wants to work with it.
Audio is not an interface, text is
A model can summarize a meeting, extract action items, or answer questions about an interview - but only after someone turns the audio into text. That step is not optional and it is not free; it is the actual interface. Get it wrong - drop timestamps, merge speakers, lose the raw text - and everything built on top inherits the loss.
Treat transcription as infrastructure, not a one-off convenience. If audio is a regular input to your pipeline, the thing that turns it into text is a dependency like any other, and it should be scriptable.
From a script: vocateca transcript <url|file>
The CLI takes a URL or a local file and returns a transcript: vocateca transcript <url|file>. A YouTube link, a podcast episode, an audio or video file on disk - same command, same output shape, run from a shell script, a cron job, or a CI pipeline.
No dashboard, no upload step, no API key to provision. The command runs on-device and exits with a transcript you can pipe straight into the next step - grep it, feed it to a model, write it to disk.
From an agent: the MCP server
An agent working through the Model Context Protocol does not shell out to a CLI - it calls a tool. Vocateca's MCP server exposes transcription as one of those tools directly, so an agent can hand it a URL or a file mid-task and get text back without a human in the loop.
That is the difference between a transcript you fetch by hand and one an agent fetches for itself while chasing some other goal - research a topic, answer a question, summarize a backlog of recordings.
Open-core, so you can see what it does
The CLI and the MCP server are open-core, Apache-2.0. Nothing about how audio becomes text is a black box you have to trust from a landing page - read the code, run it yourself, fork it if you need to.
The transcript is the API. Once audio is text, any LLM can read it - the model, not the audio format, is the choice you get to keep making.