voiceforge-mcp

MCP server for voice AI development. Recommend, benchmark, and scaffold STT+LLM+TTS stacks — powered by production data from 4 years of voice AI at enterprise scale.

Install

npx voiceforge-mcp

Or add to Claude Code:

claude mcp add voiceforge-mcp node /path/to/voiceforge-mcp/dist/index.js

Tools

voiceforge_recommend

Get the optimal STT+LLM+TTS combination for your language and use case. You can request structured output with output_format: "json" for automation.

"What's the best voice stack for Thai debt collections?"
→ Deepgram nova-3 → ElevenLabs eleven-turbo → ElevenLabs eleven_v3
  142ms latency | 4.1/5 quality | $0.004/min

voiceforge_benchmark

Browse all tested combinations with latency, quality, and cost data. Filter by language or provider. Supports output_format: "markdown" | "json" for automation.

voiceforge_compare

Compare two stacks side-by-side: latency, quality, cost, and language coverage. combo_a/combo_b support provider-only or provider+model syntax. When provider-only is used and multiple benchmark rows match, the highest-scoring match is selected.

Examples:

• combo_a: "Deepgram + OpenAI + Cartesia"
• combo_b: "Deepgram nova-3 + OpenAI gpt-4.1-mini + Cartesia sonic-3"

voiceforge_scaffold

Generate a complete voice AI project with the recommended stack.

• LiveKit Agents (Python) — agent.py, requirements.txt, voiceforge.yaml, .env
• Next.js + ElevenLabs (TypeScript) — voice widget component, env config (currently ElevenLabs-based for MVP)

voiceforge_validate

Validate a custom stack (provider + model tuple) against benchmark coverage and scaffoldability. Model names are validated against known provider catalogs; unsupported values return clear suggestions.

Use this before pitching/live demos so YC reviewers can verify a candidate stack is real and runnable.

voiceforge_providers

List all supported providers: Deepgram, ElevenLabs, Cartesia, OpenAI, Anthropic, Google, Groq, AssemblyAI, Speechmatics, PlayHT, Rime.

voiceforge_config

Generate a production-ready YAML config with pipeline settings, quality targets, and monitoring thresholds. Supports output_format: "markdown" | "json" and provider/model sanitization.

voiceforge_health

Use before interviews to show readiness at a glance:

• total benchmark rows available
• language support and framework coverage
• top performer summary (fastest, best quality, cheapest)
• health output includes data_snapshot for reproducible demo evidence

YC Demo Checklist

1. Run voiceforge_health and save the output.
2. Keep voiceforge_health {"output_format":"json"} output in your final submission folder for YC review screenshots.
3. Run voiceforge_recommend for your top use case with output_format: "json".
4. Run voiceforge_validate on your final stack before recording.
5. Run voiceforge_scaffold and share generated starter files/screenshots.
6. Run voiceforge_providers and voiceforge_config for one-page evidence before submission.

YC Demo Script

voiceforge_health {"output_format":"json"}
voiceforge_recommend {"language":"English","use_case":"customer-support","output_format":"json"}
voiceforge_validate {"stt_provider":"Deepgram","stt_model":"nova-3","llm_provider":"OpenAI","llm_model":"gpt-4.1-mini","tts_provider":"Cartesia","tts_model":"sonic-3","framework":"livekit","output_format":"json"}
voiceforge_scaffold {"language":"English","use_case":"customer-support","framework":"livekit","output_format":"json"}

Stability Improvements

• Canonical normalization now handles provider aliases and avoids ambiguous partial matches.
• Use-case aliases are normalized (customer support, debt collections, etc.).
• Scaffold output is now safe for filesystem-friendly agent names.
• Unknown provider/language filters return actionable supported-value hints.
• voiceforge_compare now parses provider+model tuples and returns ambiguity metadata when needed.

Benchmark Data

12 tested STT+LLM+TTS combinations across 10 languages, from production deployments at Morgan Stanley, Prudential, and HSBC.

| Metric | Range | |--------|-------| | Latency | 142ms — 287ms | | Quality | 4.0 — 4.8 UTMOS | | Cost | $0.004 — $0.022/min | | Languages | 10 (English, Thai, Vietnamese, Japanese, Korean, ...) |

Why VoiceForge?

Every team building voice agents does trial and error with provider combinations. ElevenLabs benchmarks itself. Deepgram benchmarks itself. Nobody benchmarks combinations. We do.

Built by Beknazar Abdikamalov — 4 years building production voice AI at Hupo, deployed to Morgan Stanley, Prudential, HSBC across 8 languages.

License

MIT