BerriAI/litellm · critical · ImportError

MCP SDK is not installed. Please install it with: pip instal

Error message

MCP SDK is not installed. Please install it with: pip install 'litellm[proxy]'

What it means

ImportError from MCPToolRegistry.convert_tools_to_mcp_sdk_tool_type (tool_registry.py:77): the module-level `from mcp.types import Tool` failed at import time (MCPToolSDKTool is None), meaning the `mcp` SDK package is absent from the environment. LiteLLM's proxy MCP server needs that SDK to serialize local registry tools into MCP SDK Tool objects for tools/list responses.

Source

Thrown at litellm/proxy/_experimental/mcp_server/tool_registry.py:77

    def unregister_tools_with_prefix(self, prefix: str) -> int:
        """Remove tools whose registered name starts with ``prefix``.

        Used when an OpenAPI-backed MCP server leaves the runtime registry so
        stale tool handlers cannot be invoked after eviction.
        """
        if not prefix:
            return 0
        removed = 0
        for name in list(self.tools.keys()):
            if name.startswith(prefix):
                del self.tools[name]
                removed += 1
                verbose_logger.debug("Unregistered MCP tool %s", name)
        return removed

    def convert_tools_to_mcp_sdk_tool_type(self, tools: list[MCPTool]) -> list["MCPToolSDKTool"]:
        if MCPToolSDKTool is None:
            raise ImportError("MCP SDK is not installed. Please install it with: pip install 'litellm[proxy]'")
        return [
            MCPToolSDKTool(
                name=tool.name,
                description=tool.description,
                inputSchema=tool.input_schema,
            )
            for tool in tools
        ]

    def load_tools_from_config(
        self,
        mcp_tools_config: dict[str, Any] | None = None,
        config_file_path: str | None = None,
    ) -> None:
        """
        Load and register tools from the proxy config

        Args:

View on GitHub (pinned to 77b7c6c40c)

Solutions

1. Install the extra: pip install 'litellm[proxy]' (or pip install mcp), then restart the proxy/process
2. Pin the extra in requirements/pyproject: litellm[proxy]>=<version> so CI and images always include mcp
3. Rebuild Docker images from a clean layer after adding the extra to flush stale caches
4. Verify with python -c "import mcp.types" before starting the proxy

Example fix

# before
pip install litellm  # mcp missing -> ImportError on tools/list

# after
pip install 'litellm[proxy]'

import importlib.util

def mcp_sdk_available() -> bool:
    return importlib.util.find_spec("mcp") is not None

assert mcp_sdk_available(), "install the proxy extra: pip install 'litellm[proxy]'"

def safe_convert(registry, tools):
    if getattr(registry, "_MCP_SDK_missing", False):
        return None  # caller decides to skip or hard-fail
    return registry.convert_tools_to_mcp_sdk_tool_type(tools)

try:
    sdk_tools = registry.convert_tools_to_mcp_sdk_tool_type(tools)
except ImportError as e:
    raise RuntimeError("MCP tool listing requires the mcp SDK: pip install 'litellm[proxy]'") from e

When it happens

Trigger: Listing local mcp_tools via an MCP session when litellm was installed bare (pip install litellm) without the proxy extra; slim Docker images that strip optional deps; a venv built from requirements that pin only litellm; CI environments reusing a cached env created before mcp became required.

Common situations: Embedding litellm in a custom app without litellm[proxy]; upgrading litellm in an env whose mcp pin was dropped; running the proxy from a source checkout without installing extras; dependency resolver removing mcp after a conflict.

Related errors

• OpenTelemetry OTLP HTTP exporter is not available. Install `
• OpenTelemetry OTLP gRPC exporter is not available. Install `
• OpenTelemetry OTLP gRPC log exporter is not available. Insta
• OpenTelemetry OTLP gRPC metric exporter is not available. In
• OpenTelemetry OTLP gRPC metric exporter is not available. In

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/ecb0b65f502769c7. Report an issue: GitHub.