agentic-transcribe

Package · Support

agentic-transcribe is an open-source audio transcription scaffold with OpenAI, Deepgram, and AssemblyAI request paths, a JavaScript API, CLI, and Model Context Protocol (MCP) server.

Install

npm install agentic-transcribe

Use Node.js 18 or later. TypeScript declarations are included. The package has two direct runtime dependencies.

Preview your first request

npx agentic-transcribe transcribe --file recording.mp3 --vendor openai

Expect exit code 0 and this JSON:

{"vendor":"openai","audioFile":"recording.mp3","text":"","segments":[]}

This preview requires no audio file or credentials. It performs no file read or network request. An empty transcript confirms preview behavior; it does not confirm transcription.

import { transcribeAudio } from 'agentic-transcribe';
console.log(await transcribeAudio('recording.mp3', { vendor: 'deepgram' }));

Features and coverage

• Preview: Return the transcript shape without vendor calls; dry run is the default across all interfaces.
• Response mapping: Map vendor response text and timestamp fields to text and segments.
• Errors: Expose missing API keys, network failures, HTTP failures, and invalid JSON.

Three vendor paths exist. All are scaffolds; none provides verified local-file transcription.

| Integration | Status | Live-path environment key | Current limit | | --- | --- | --- | --- | | OpenAI | scaffold | OPENAI_API_KEY | Sends an empty file Blob; never reads local audio | | Deepgram | scaffold | DEEPGRAM_API_KEY | Sends an empty byte array; never reads local audio | | AssemblyAI | scaffold | ASSEMBLYAI_API_KEY | Sends the input as audio_url; has no upload or job polling |

Tests use mocked responses. This campaign performs zero live credential checks. The OpenAI path hardcodes gpt-4o-transcribe and verbose_json; request compatibility is unverified.

Runtime surfaces and limits

CLI operations are transcribe, mcp, --help/-h, and --version/-V. --file names the input; --vendor selects a path. --live enables the incomplete vendor request. CLI usage errors exit 2; other errors exit 1.

Start stdio MCP with npx agentic-transcribe-mcp or npx agentic-transcribe mcp. The single tool, transcribe_audio, accepts audioFile, optional vendor, and optional live. The SDK enables vendor calls only with dryRun: false. It accepts apiKey and fetchFn overrides.

Export the matching key before a live request. The package does not load .env files. It has no file upload implementation, automatic retries, timeout setting, streaming, or background job polling. The repository Worker handler exposes preview-only POST /transcribe; it is not a managed service.

Status and links

Version 0.1.1 is an early scaffold. APIs can change before 1.0. Use GitHub issues for support. No public documentation site is configured. Public case studies are not yet published.

License

MIT.