Providers

The Omni-Provider Engine

Initially, Parallax was exclusively hardcoded to pure Gemini API schemas. We chose that route because generic "provider-agnostic" abstractions often strip away the low-level metadata required for strict deterministic tool-calling.

However, as Parallax has evolved into a massively parallelized swarm orchestrator, we've broken down that limitation by rolling our own native translation layers. Parallax now supports OpenAI, Anthropic, Ollama, LMStudio, and vLLM securely, translating their streaming payloads into raw Gemini ReAct schemas instantly under the hood.

Supported Providers

You can switch to any of these models instantly via the /model dropdown, or by typing /model <prefix:model-name>:

Google Gemini (Default)

Parallax uses the @google/gemini-cli-core authentication context. You don't even need an API key; just sign in globally via the Gemini CLI context and Parallax natively detects it.

gemini:gemini-3.1-pro-preview (Used for massive reasoning tasks)
gemini:gemini-3-flash-preview (Used natively by the swarm subagents)

OpenAI & Anthropic

To invoke commercial tier-1 logic engines natively, simply expose your keys to the terminal context when booting Parallax:

export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
parallax

openai:gpt-4o
openai:o1
anthropic:claude-3-7-sonnet-latest

Local / Offline execution (Ollama, LMStudio, vLLM)

For strict privacy, Parallax effortlessly routes the exact same autonomous parallel loop into offline inferencing clusters using standard OpenAI schema shims.

Ensure your local server is running, and Parallax connects automatically:

Ollama: ollama:llama3.3 (Defaults to localhost:11434/v1)
LMStudio: lmstudio:llama3 (Defaults to localhost:1234/v1)
vLLM: vllm:mistral (Defaults to localhost:8000/v1)

Note: Parallax's ToolLoopAgent relies on highly precise strict JSON schema structures. We recommend only running 14B+ parameter models locally (like qwen2.5:14b or llama3.3:70b) to prevent the agent from hallucinating function calls inside the parallel subagent worker loops.