Skip to content

Models and providers

TabulaFlow supports hundreds of models across more than 25 cloud and local providers. It uses one model for the main conversation and another for parallel subagent work. Most users only need to export one provider API key and choose models in /config.

Quick setup

Choose one provider and configure it before launching TabulaFlow:

export OPENAI_API_KEY="your-api-key"
tabulaflow
export ANTHROPIC_API_KEY="your-api-key"
tabulaflow
export VLLM_BASE_URL="http://127.0.0.1:8000/v1"
# For authenticated endpoints:
# export VLLM_API_KEY="your-api-key"
tabulaflow
export FIREWORKS_API_KEY="your-api-key"
tabulaflow

TabulaFlow automatically selects models for OpenAI, Anthropic, and single-model vLLM endpoints. Use /config to change them or select among multiple vLLM models. Credentials are read from the environment and never saved.

vLLM tool calling

TabulaFlow relies on automatic tool calling. Follow the current vLLM tool-calling guide to configure it for your model.

Choose models

Run /config to choose the main model, subagent model, reasoning effort, and maximum model request rate. The main model handles the conversation and plans tool use; the subagent model handles parallel extraction, classification, and other row-wise work.

Model identifiers use the Pydantic AI provider:model format:

openai:gpt-5.6-sol
anthropic:claude-sonnet-5
vllm:Qwen/Qwen3-8B

The picker includes recommended models, the current selection, and compatible models from the installed provider stack. You can also type a complete custom identifier. TabulaFlow saves model selections in ~/.tabulaflow/app_config.json, but continues to read credentials from the environment.

Supported providers

The model picker includes models from all the provider routes below. TabulaFlow uses Pydantic AI's environment variables and authentication unchanged; it does not store these credentials. Set the listed variables before launching TabulaFlow, then select a provider:model identifier in /config when automatic setup does not apply.

API key providers

Provider Prefix Environment variable
OpenAI openai: OPENAI_API_KEY
Anthropic anthropic: ANTHROPIC_API_KEY
Google Gemini google: GOOGLE_API_KEY or GEMINI_API_KEY
OpenRouter openrouter: OPENROUTER_API_KEY
DeepSeek deepseek: DEEPSEEK_API_KEY
xAI xai: XAI_API_KEY
Groq groq: GROQ_API_KEY
Mistral mistral: MISTRAL_API_KEY
Together AI together: TOGETHER_API_KEY
Hugging Face huggingface: HF_TOKEN
Fireworks AI fireworks: FIREWORKS_API_KEY
Cohere cohere: CO_API_KEY
Cerebras cerebras: CEREBRAS_API_KEY
Alibaba Cloud Model Studio alibaba: ALIBABA_API_KEY or DASHSCOPE_API_KEY
Moonshot AI moonshotai: MOONSHOTAI_API_KEY
Nebius AI Studio nebius: NEBIUS_API_KEY
Z.AI zai: ZAI_API_KEY
Crusoe crusoe: CRUSOE_API_KEY
OVHcloud AI Endpoints ovhcloud: OVHCLOUD_API_KEY

Cloud and hosted authentication

Provider Prefix Authentication
Amazon Bedrock bedrock: AWS credential chain and AWS_DEFAULT_REGION, or AWS_BEARER_TOKEN_BEDROCK
Azure OpenAI azure: AZURE_OPENAI_API_KEY and AZURE_OPENAI_ENDPOINT; OPENAI_API_VERSION for versioned endpoints
Google Cloud google-cloud: Application Default Credentials with GOOGLE_CLOUD_PROJECT and optional GOOGLE_CLOUD_LOCATION; or GOOGLE_API_KEY for Vertex AI Express Mode
GitHub Copilot github-copilot: GITHUB_COPILOT_API_KEY, GITHUB_COPILOT_API_TOKEN, or COPILOT_GITHUB_TOKEN
Vercel AI Gateway vercel: VERCEL_AI_GATEWAY_API_KEY or VERCEL_OIDC_TOKEN
Snowflake Cortex snowflake: SNOWFLAKE_ACCOUNT and SNOWFLAKE_TOKEN

Local and custom endpoints

Provider Prefix Connection settings
Ollama ollama: OLLAMA_BASE_URL; optional OLLAMA_API_KEY
vLLM vllm: VLLM_BASE_URL; optional VLLM_API_KEY

Provider requirements can change independently of TabulaFlow. Follow the linked Pydantic AI provider page for account setup and provider-specific details. If a required setting is missing, TabulaFlow names it when model activation fails.

Model requirements

The Data Agent sends instructions and tool schemas and expects models to produce dependable tool calls. Choose models with:

  • Text generation and tool-calling support
  • A large context window
  • Reliable structured output for extraction and subagent work

Disabling the LLM in /config turns off conversational analysis while keeping data connections and the data explorer available.