AssemblyAI: 2 reads and 1 write, one client

Speech-to-text with speaker labels and summarisation. Every call below is available over HTTP, through the SDK and as a tool for a coding agent, with the same arguments and the same meter.

What AssemblyAI can do here

The integration exposes 2 reads and 1 write, each with a declared input and output schema rather than a page of documentation you would have to translate. Speech-to-text with speaker labels and summarisation.

How it authenticates

AssemblyAI authenticates with a apiKey, supplied per request or stored once against your account and sealed at rest. On our side the same value is read from apiKey when we broker the call for you.

Called by name, not by capability

AssemblyAI does not currently map onto a vendor-neutral capability, so you call its operations by name. That is the right shape when you genuinely mean this provider — which, for a service with its own model of the world, is most of the time.

• createTranscript — write: Submit audio for transcription. ASYNCHRONOUS — poll getTranscript for the text.
• getTranscript — read: One transcript, with its status and text once ready.
• listTranscripts — read: Recent transcripts. The connection test.

Questions

Do I have to bring my own API keys?

You can, and that lane is free for ever — pass your key on the request or store it against your account, and we are one proxy hop. The paid lane is the other direction: we call 226 providers with our credentials so you never register anywhere. Both go through the same client and the same operation names.

What happens to a key I store here?

It is sealed with AES-256-GCM before it reaches the database and is only ever opened to make the call you asked for. It is scoped to your account, it is never logged, and deleting the connection deletes it. For an OAuth provider we hold a refresh token instead, and revoking the link at the provider stops the calls here immediately.

What happens when a provider changes its API?

The operation contract is declared, so a response that no longer matches is a refusal rather than a silently wrong answer. The contract does not coerce types — a field declared a number that arrives as a string is rejected at the edge, on the way in and on the way out.