Build it in your project
Paste this line into your agent (Claude Code, Codex, Cursor, OpenCode, Cline). It sets up your key, then builds the pipeline in your project:The code
One file per language. Both readRUNCRATE_API_KEY from the environment; nothing secret is in the code.
What the proof run cost
The Python file on a 34-second three-person launch sync dated 2026-10-19 (a Monday). Measured on 2026-10-23 againstapi.runcrate.ai; costs are what the account was billed, read from each response’s X-LLM-Cost header or usage.cost.
Transcription was billed at its 2026-10-23 price; it costs less now, at the rate on the live model pricing page.
Both runs found the three action items with the right due dates (“by Thursday” became 2026-10-22) and the Friday 10:00 follow-up. The TypeScript run, also $0.0120, attributed Priya’s item to “Unidentified speaker”.
Scale it
- Every meeting is independent: run it from your recorder’s upload webhook.
- Hours of audio: transcription is billed per minute of audio at the model’s price on the live model pricing page; whisper-large-v3-turbo costs less than Voxtral Small, which also gives no timestamps.
- Long meetings fit: DeepSeek V4.1 Flash takes a million tokens of context.
Gotchas found in the proof run
- Transcripts carry no speaker labels, so attribution depends on people saying names. One of two runs missed an owner who only introduced herself at the start (“Priya here”). If it matters, diarise first or ask people to name owners.
- DeepSeek V4.1 Flash took 34 s for the extraction: it reasons before answering. gpt-oss-120b is faster if you need minutes in seconds.
- Strict JSON schemas with nullable fields (
"type": ["string", "null"]) worked on both runs.