Transcription
Use the connected PyAI tool schemas as the source of truth; clients may prefix tool IDs with the server name. If PyAI is disconnected, direct the user to the host's browser sign-in. Keep credentials out of chat. Respect the selected project and the user's authorized task.
Tools: whoami, transcribe_audio, create_transcription_job, get_transcription_job, list_transcription_jobs
For a short supplied recording, use transcribe_audio with base64 audio (up to 8 MiB decoded). For a larger recording already available at an authorized HTTPS URL, use create_transcription_job. Inspect current schemas for diarization, language and other options; never promise an option absent from the contract.
Preserve the returned job_id. Poll get_transcription_job at a reasonable interval until completion, an error, or the task's waiting limit. Pending is not failure. If waiting ends, return the actual ID and status so work can resume; do not resubmit the recording.
Return the transcript and any speaker/timestamp data actually supplied. For microphone streaming, use the Hear streaming API or SDK described in https://docs.pyai.com; an MCP transcription job is not a live audio stream.