LLM Provider Configuration

Hive uses a unified provider interface to manage interactions with Large Language Models. While the framework is designed to be provider-agnostic, it primarily utilizes the LiteLLMProvider to offer seamless integration with OpenAI, Anthropic, Google Gemini, and 100+ other providers.

The LiteLLM Provider

The LiteLLMProvider is the standard interface for agent reasoning within Hive. It handles model routing, API authentication, and token tracking.

Programmatic Configuration

When initializing an AgentRuntime, you provide an instance of the provider configured with your chosen model and credentials.

from framework.llm import LiteLLMProvider

llm = LiteLLMProvider(
    model="claude-3-5-sonnet-20240620", # Model identifier
    api_key="your-api-key-here",       # Optional if set in environment
    api_base="https://api.example.com"  # Optional: for proxy/alternative endpoints
)

Supported Provider Formats

Hive follows standard model string formats to identify the backend provider:

| Provider | Model String Format | Required Environment Variable | | :--- | :--- | :--- | | Anthropic | anthropic/claude-3-5-sonnet-20240620 | ANTHROPIC_API_KEY | | OpenAI | openai/gpt-4o | OPENAI_API_KEY | | Google | gemini/gemini-1.5-pro | GEMINI_API_KEY | | Local (Ollama) | ollama/llama3 | N/A |

Credential Management

Hive includes a robust credential store located in core/framework/credentials/models.py to manage secrets securely. Credentials are stored as CredentialObject instances, preventing accidental logging of sensitive keys through the use of Pydantic's SecretStr.

Credential Types

The framework categorizes credentials to handle different authentication flows:

• API_KEY: Standard keys for services like OpenAI or Brave Search.
• OAUTH2: Supports refresh tokens for long-lived agent sessions.
• BEARER_TOKEN: For JWT-based authentication.

Usage in Runtime

When building an agent, you typically pass the LLM provider into the create_agent_runtime factory function:

from framework.runtime.agent_runtime import create_agent_runtime

runtime = create_agent_runtime(
    graph=my_graph,
    llm=llm,          # The LiteLLMProvider instance
    tools=my_tools,
    storage_path="./agent_data"
)

Environment Variables

For production deployments, it is recommended to configure providers via environment variables. Hive and the underlying LiteLLM library will automatically detect these:

# OpenAI Configuration
export OPENAI_API_KEY='sk-...'

# Anthropic Configuration
export ANTHROPIC_API_KEY='sk-ant-...'

# Custom API Base (e.g., for Azure or LiteLLM Proxy)
export AZURE_API_BASE='https://my-resource.openai.azure.com/'
export AZURE_API_KEY='...'

Monitoring and Logs

LLM interactions are captured by Hive's three-level runtime logging system:

1. Level 1 (Summary): Total token counts and success status for the entire run.
2. Level 2 (Node Detail): Token usage and latency per node execution.
3. Level 3 (Tool Logs): Granular tool calls and the raw LLM text responses (see NodeStepLog).

This data is viewable in real-time via the Aden TUI Dashboard or by inspecting the RunSummaryLog generated after an agent execution.