Connect your AI

Nightjar does not hard-wire a single AI vendor. Transcription and summaries each target a provider you choose in the server config. This guide shows the exact settings for each.

All settings live in the server’s .env file. Change them, restart the server, and the new provider is live. Nothing else needs to change on the device or in the app.

Summaries and chat (the language model)

Set LLM_PROVIDER to one of ollama, openai_compatible, or anthropic.

Fully local (Ollama)

The default. Everything stays on your machine, no API key, no cloud. Install Ollama, pull a model, and point Nightjar at it.

NIGHTJAR_LLM_PROVIDER=ollama
NIGHTJAR_OLLAMA_URL=http://localhost:11434
NIGHTJAR_OLLAMA_MODEL=llama3.1

OpenAI, or anything OpenAI-compatible

One setting covers OpenAI, Groq, OpenRouter, Together, LocalAI, vLLM, LM Studio, and Gemini’s OpenAI-compatible endpoint. Just change the base URL and model.

NIGHTJAR_LLM_PROVIDER=openai_compatible
NIGHTJAR_LLM_BASE_URL=https://api.openai.com/v1
NIGHTJAR_LLM_API_KEY=sk-...
NIGHTJAR_LLM_MODEL=gpt-4o-mini

For Groq (fast and cheap), use https://api.groq.com/openai/v1. For a local model server on your LAN such as LM Studio, use http://localhost:1234/v1.

Anthropic Claude

Claude uses its own integration.

NIGHTJAR_LLM_PROVIDER=anthropic
NIGHTJAR_LLM_API_KEY=sk-ant-...
NIGHTJAR_LLM_MODEL=claude-opus-4-8

Transcription (speech to text)

Set TRANSCRIPTION_PROVIDER to local_whisper (self-hosted) or openai_compatible.

Local, self-hosted (faster-whisper)

NIGHTJAR_TRANSCRIPTION_PROVIDER=local_whisper
NIGHTJAR_WHISPER_MODEL=base
NIGHTJAR_WHISPER_DEVICE=cpu
NIGHTJAR_WHISPER_COMPUTE_TYPE=int8

Larger models (small, medium, large-v3) are more accurate but slower. Set DEVICE=cuda if you have a GPU.

A hosted transcription API

NIGHTJAR_TRANSCRIPTION_PROVIDER=openai_compatible
NIGHTJAR_TRANSCRIPTION_BASE_URL=https://api.openai.com/v1
NIGHTJAR_TRANSCRIPTION_API_KEY=sk-...
NIGHTJAR_TRANSCRIPTION_MODEL=whisper-1

Which should I choose?

• Maximum privacy, zero cost per minute: local Whisper + a local Ollama model. Nothing leaves your machine.
• Best quality, least setup: a commercial API for both. You pay per use, and accuracy is top-tier.
• A balance: local transcription for privacy, a commercial model for the summary polish.

Not sure, or would rather not manage any of this? Our hosted plans run it all for you.