[BUG] MCP Server: Anthropic provider fails with 400 error when temperature not explicitly set

[barweiss45]

MCP Server: Anthropic provider fails with 400 error when temperature not explicitly set in config

Description

When using the Anthropic LLM provider without explicitly setting temperature in config.yaml, the API returns a 400 error because None is passed instead of a valid float.

The MCP server's LLMConfig model in mcp_server/src/config/schema.py defaults temperature to None:

temperature: float | None = Field(
    default=None, description='Temperature (optional, defaults to None for reasoning models)'
)

This None value is then passed directly to the Anthropic API, which expects a numeric value.

Error

Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'temperature: Input should be a valid number'}}

Environment

• graphiti-mcp-server v0.24.3
• graphiti-core v0.24.3
• Provider: Anthropic (claude-sonnet-4-5-20250929)

Root Cause

1. graphiti_core/llm_client/config.py defines a sensible default:

DEFAULT_TEMPERATURE = 1

class LLMConfig:
    def __init__(
        self,
        temperature: float = DEFAULT_TEMPERATURE,  # Works correctly
        ...
    ):

2. mcp_server/src/config/schema.py (L151-153) overrides this with None:

temperature: float | None = Field(
    default=None, description='Temperature (optional, defaults to None for reasoning models)'
)

3. mcp_server/src/services/factories.py (L206-212) passes None explicitly, overriding the core default:

llm_config = GraphitiLLMConfig(
    api_key=api_key,
    model=config.model,
    temperature=config.temperature,  # Passes None, overriding core default
    max_tokens=config.max_tokens,
)
return AnthropicClient(config=llm_config)

4. Anthropic API receives None and returns 400.

Why This Happens

The None default was added to support OpenAI reasoning models (o1, o3, gpt-5) which reject the temperature parameter entirely. However, the implementation is backwards - the edge case (OpenAI reasoning models) drives the default, breaking the normal case (all other providers).

When using graphiti-core directly (without MCP server), the default works correctly because the core's LLMConfig defaults temperature to 1.0.

Workaround

Explicitly set temperature in config.yaml:

llm:
  provider: "anthropic"
  model: "claude-sonnet-4-5-20250929"
  temperature: 1.0

Suggested Fix

The None default exists to support OpenAI reasoning models, but this is the edge case, not the norm. The fix should be:

1. Change the schema default to 1.0 in mcp_server/src/config/schema.py:

temperature: float | None = Field(
    default=1.0, description='Temperature (optional, set to None for OpenAI reasoning models)'
)

2. Handle OpenAI reasoning models explicitly in mcp_server/src/services/factories.py OpenAI case - detect reasoning models (o1, o3, gpt-5) and omit temperature for those.

This way:

• Anthropic/Gemini/Groq work out of the box with sensible default
• OpenAI reasoning models are handled as the exception they are
• Users can still override temperature in config if needed

Related

• [BUG] Can't use GPT-5 model family - Unsupported parameter: temperature is not supported with this model. #874 (GPT-5 temperature parameter handling)
• [BUG] OpenAIClient passes unsupported 'temperature' parameter to GPT-5 models causing 400 Bad Request errors #878 (OpenAI reasoning model temperature fix)