@databricks/ai-sdk-provider weekly downloads statistics

@databricks/ai-sdk-provider

Databricks provider for the Vercel AI SDK.

Features

• 🚀 Support for all Databricks endpoint types: • Responses (agent/v1/responses) - Foundation model and agent responses API (docs) • Chat Completions (llm/v1/chat) - Foundation model chat completions API (docs) • Chat Agent (agent/v2/chat) - Legacy Databricks chat agent API (docs) • 🔄 Stream and non-stream (generate) support for all endpoint types • 🛠️ Tool calling and agent support • 🔐 Flexible authentication (bring your own tokens/headers) • 🎯 Full TypeScript support

Installation

Peer Dependencies

This package requires the following peer dependencies:

To use the provider with AI SDK functions like generateText or streamText, also install:

Quick Start

Authentication

The provider requires you to pass authentication headers:

API Reference

Main Export

createDatabricksProvider(settings)

Creates a Databricks provider instance.

Parameters:

• settings.baseURL (string, required): Base URL for the Databricks API calls • settings.headers (object, optional): Custom headers to include in requests • settings.provider (string, optional): Provider name (defaults to "databricks") • settings.fetch (function, optional): Custom fetch implementation • settings.formatUrl (function, optional): Optional function to format the URL • settings.useRemoteToolCalling (boolean, optional): Enable remote tool calling mode (defaults to false). See Remote Tool Calling below.

Returns: DatabricksProvider with three model creation methods:

• responses(modelId: string): Create a Responses model • chatCompletions(modelId: string): Create a Chat Completions model • chatAgent(modelId: string): Create a Chat Agent model

Remote Tool Calling

The useRemoteToolCalling option controls how tool calls from Databricks agents are handled. When enabled, tool calls are marked as dynamic: true and providerExecuted: true, which tells the AI SDK that:

1. Dynamic: The tools are not pre-registered - the agent decides which tools to call at runtime 2. Provider-executed: The tools are executed remotely by Databricks, not by your application

When to use useRemoteToolCalling: true

Enable this option when your Databricks agent handles tool execution internally:

• Databricks Agents with built-in tools: Agents that use tools like Python execution, SQL queries, or other Databricks-managed tools • Agents on Apps: When deploying agents that manage their own tool execution • MCP (Model Context Protocol) integrations: When tools are executed via MCP servers managed by Databricks

When NOT to use useRemoteToolCalling

Keep this option disabled (the default) when:

• You define and execute tools locally: Your application registers tools with the AI SDK and handles their execution • Standard chat completions: You're using the Chat Completions endpoint without agent features • Hybrid scenarios: You want to intercept tool calls and handle some locally

Example: Remote tool calling with Databricks agents

Examples

Responses Endpoint

Links

• Vercel AI SDK Documentation • Databricks Documentation • GitHub Repository

Contributing

This package is part of the databricks-ai-bridge monorepo.