LLM Providers

The LLM provider handles all text generation in the pipeline: web research, script writing (Creative Director), image/video prompt optimization, stock query reformulation, and quality critique. OpenReels uses the Vercel AI SDK under the hood, so all providers share the same structured output and tool-calling interface.

Supported providers

Usage

Set the LLM provider with the --provider flag:

pnpm start "the apollo 13 disaster" --provider anthropic
pnpm start "the apollo 13 disaster" --provider openai
pnpm start "the apollo 13 disaster" --provider gemini

OpenRouter

Access 300+ models through a single API key:

pnpm start "the apollo 13 disaster" --provider openrouter
pnpm start "the apollo 13 disaster" --provider openrouter --llm-model google/gemini-2.5-flash

OpenAI-compatible endpoints

Use any endpoint that speaks the OpenAI API (Ollama, Together, Groq, vLLM):

pnpm start "the apollo 13 disaster" --provider openai-compatible \
  --llm-base-url http://localhost:11434/v1 \
  --llm-model llama3:8b \
  --search-provider none

Model override

Override the default model for any provider with --llm-model:

pnpm start "topic" --provider anthropic --llm-model claude-opus-4-6

Convenience shortcuts

The google shortcut sets the LLM to Gemini along with all other providers to their Google equivalents:

# Sets LLM=Gemini, Image=Gemini, TTS=Gemini, Video=Veo, Music=Lyria
pnpm start "the apollo 13 disaster" --provider google

Environment variables

Each provider reads its API key from a specific environment variable:

# .env
ANTHROPIC_API_KEY=sk-ant-...       # Anthropic Console
OPENAI_API_KEY=sk-...              # OpenAI Platform
GOOGLE_API_KEY=AI...               # Google AI Studio
OPENROUTER_API_KEY=sk-or-...       # OpenRouter
OPENREELS_LLM_API_KEY=...          # For custom OpenAI-compatible endpoints

Only the key for your chosen provider is required. If you use --provider gemini, you only need GOOGLE_API_KEY.

How the LLM is used

The pipeline makes several LLM calls per run:

Search providers

Control how the research agent searches the web with --search-provider:

Native providers use their built-in search tools by default:

• Anthropic — webSearch_20250305 tool
• OpenAI — webSearch tool
• Gemini — googleSearch tool

OpenRouter and OpenAI-compatible providers auto-detect Tavily if TAVILY_API_KEY is set, otherwise fall back to parametric knowledge.

# Override search on any provider
pnpm start "topic" --provider anthropic --search-provider tavily
# Disable search entirely
pnpm start "topic" --provider openrouter --search-provider none

Tavily requires TAVILY_API_KEY (free tier: 1000 searches/month at tavily.com).

Cost comparison

LLM costs are typically the smallest part of the total bill. A typical run uses 3-7 LLM calls.

Gemini is roughly 30x cheaper than Anthropic per token. For budget-conscious runs, combine --provider gemini with --tts-provider kokoro and --music-provider bundled.

VLM verification model

Stock footage verification uses a separate VLM (Vision Language Model) call. By default, this uses the same provider as your main LLM. The verification model examines downloaded stock footage and scores how well it matches the requested visual. The models used for verification are:

• Anthropic — claude-sonnet-4-6
• OpenAI — gpt-4o
• Gemini — gemini-2.5-flash
• OpenRouter — anthropic/claude-sonnet-4 (or your --llm-model)
• OpenAI-compatible — falls back to Anthropic if no OPENREELS_LLM_BASE_URL is set