# MCP Server Tool Evaluation Support
## Overview
Add support for evaluating tools from remote MCP servers without
requiring Python callables. Enables direct evaluation of any
MCP-compatible tool server.
## What's New
### Core Features
- **`MCPToolRegistry`**: Evaluate tools from a single MCP server
- **`CompositeMCPRegistry`**: Evaluate tools from multiple MCP servers
simultaneously
- **Automatic loaders**: `load_from_stdio()` and `load_from_http()` to
fetch tools from running servers
- **Automatic namespacing**: Tools prefixed with server name (e.g.,
`server_tool_name`)
- **Smart name resolution**: Use short names if unique, full names if
ambiguous
- **OpenAI strict mode**: Automatic schema conversion prevents parameter
hallucinations
### Usage
**Automatic Loading:**
```python
from arcade_evals import load_from_stdio, MCPToolRegistry
# Load tools automatically from MCP server
tools = load_from_stdio(["npx", "-y", "@modelcontextprotocol/server-github"])
registry = MCPToolRegistry(tools)
```
**Single MCP Server:**
```python
from arcade_evals import MCPToolRegistry, ExpectedToolCall
registry = MCPToolRegistry(mcp_tools)
suite = EvalSuite(catalog=registry)
suite.add_case(
expected_tool_calls=[
ExpectedToolCall(tool_name="tool_name", args={...})
]
)
```
**Multiple MCP Servers:**
```python
from arcade_evals import CompositeMCPRegistry, load_from_stdio
# Load from multiple servers
github_tools = load_from_stdio(["npx", "-y", "@modelcontextprotocol/server-github"])
slack_tools = load_from_stdio(["npx", "-y", "@modelcontextprotocol/server-slack"])
composite = CompositeMCPRegistry(
tool_lists={
"github": github_tools,
"slack": slack_tools,
}
)
suite = EvalSuite(catalog=composite)
suite.add_case(
expected_tool_calls=[
ExpectedToolCall(tool_name="github_list_issues", args={...})
]
)
```
## Implementation
### Files Changed
- **`libs/arcade-evals/arcade_evals/registry.py`** (NEW): Registry
abstractions and implementations
- **`libs/arcade-evals/arcade_evals/loaders.py`** (NEW): Automatic tool
loading from MCP servers
- **`libs/arcade-evals/arcade_evals/eval.py`** (MODIFIED): Enhanced
`ExpectedToolCall` and evaluation logic
- **`libs/arcade-evals/arcade_evals/__init__.py`** (MODIFIED): Exported
new registries and loaders
### Key Technical Details
- Added `BaseToolRegistry` interface for abstraction
- `MCPToolRegistry` handles single server tools
- `CompositeMCPRegistry` manages multiple servers with collision
detection
- `load_from_stdio()` and `load_from_http()` for automatic tool
discovery
- Fixed name normalization bug: MCP tools use underscores (not dots)
- Optimized tool copying: 2.5x faster via shallow copy
## Testing
- ✅ 41 tests passing (25 new tests added)
- ✅ `test_eval_mcp_registry.py`: MCPToolRegistry functionality
- ✅ `test_eval_composite_mcp.py`: CompositeMCPRegistry with multiple
servers
- ✅ Verified backward compatibility with Python tools
## Backward Compatibility
✅ **100% backward compatible** - No breaking changes
## Breaking Changes
**None**
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> Adds end-to-end eval UX: examples, a robust CLI runner, and rich
outputs.
>
> - **New examples**: `eval_arcade_gateway.py`,
`eval_stdio_mcp_server.py`, `eval_http_mcp_server.py`,
`eval_comprehensive_comparison.py` with timeouts, error handling, and
track-based comparisons; detailed `README.md`
> - **CLI runner**: `arcade_cli/evals_runner.py` to execute
evals/capture in parallel with progress, error isolation, failed-only
filtering, context inclusion, and multi-provider/model support
> - **Output formatters**: `arcade_cli/formatters/` (txt, md, html,
json) for evals and capture; comparative and multi-model HTML with tabs
and context rendering
> - **Display refactor**: `display.py` now supports writing multiple
formats, failed-only disclaimers, include-context, and improved console
summaries
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
ff8acf9c34a6b61462a019a1ee9df081006517d0. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
---------
Co-authored-by: Francisco Liberal <francisco@arcade.dev>
Co-authored-by: Mateo Torres <torresmateo@gmail.com>
283 lines
11 KiB
Python
283 lines
11 KiB
Python
"""EvalSuite internal unified tool registry (not part of the public API)."""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any, Literal, TypedDict
|
|
|
|
from arcade_core.converters.anthropic import to_anthropic
|
|
from arcade_core.converters.utils import normalize_tool_name
|
|
|
|
from arcade_evals._evalsuite._anthropic_schema import convert_mcp_to_anthropic_tool
|
|
from arcade_evals._evalsuite._openai_schema import convert_to_strict_mode_schema
|
|
|
|
ToolFormat = Literal["openai", "anthropic"]
|
|
|
|
|
|
class _MCPToolDefinitionRequired(TypedDict):
|
|
"""Required fields for MCP-style tool definition."""
|
|
|
|
name: str
|
|
|
|
|
|
class MCPToolDefinition(_MCPToolDefinitionRequired, total=False):
|
|
"""MCP-style tool definition structure.
|
|
|
|
This is the format expected by `add_tool_definitions()` and used internally
|
|
by the EvalSuiteToolRegistry.
|
|
|
|
Required:
|
|
name: The unique tool name.
|
|
|
|
Optional:
|
|
description: Human-readable description (defaults to "").
|
|
inputSchema: JSON Schema for input parameters
|
|
(defaults to {"type": "object", "properties": {}}).
|
|
"""
|
|
|
|
description: str
|
|
inputSchema: dict[str, Any]
|
|
|
|
|
|
class EvalSuiteToolRegistry:
|
|
"""
|
|
A minimal internal registry that stores tools in MCP-style descriptors:
|
|
|
|
{
|
|
"name": "...",
|
|
"description": "...",
|
|
"inputSchema": { ... JSON Schema ... }
|
|
}
|
|
|
|
EvalSuite converts Python tools into this format too, so there is only one
|
|
runtime path for OpenAI tool formatting.
|
|
|
|
Note: Tools are stored with their original names (e.g., "Google.Search"),
|
|
but Anthropic requires underscores (e.g., "Google_Search"). The registry
|
|
maintains a mapping to look up tools by either format.
|
|
"""
|
|
|
|
def __init__(self, *, strict_mode: bool = True) -> None:
|
|
self._tools: dict[str, dict[str, Any]] = {}
|
|
self._strict_mode = strict_mode
|
|
# Mapping from normalized names (underscores) to original names (dots)
|
|
# e.g., {"Google_Search": "Google.Search"}
|
|
self._normalized_to_original: dict[str, str] = {}
|
|
# Store original MaterializedTool objects for direct Anthropic conversion (Python tools only)
|
|
self._materialized_tools: dict[str, Any] = {}
|
|
|
|
@property
|
|
def strict_mode(self) -> bool:
|
|
return self._strict_mode
|
|
|
|
@strict_mode.setter
|
|
def strict_mode(self, value: bool) -> None:
|
|
self._strict_mode = value
|
|
|
|
def add_tool(
|
|
self,
|
|
tool_descriptor: MCPToolDefinition | dict[str, Any],
|
|
materialized_tool: Any = None,
|
|
) -> None:
|
|
"""Add a tool to the registry.
|
|
|
|
Args:
|
|
tool_descriptor: MCP-style tool definition.
|
|
materialized_tool: Optional MaterializedTool for direct Anthropic conversion (Python tools only).
|
|
"""
|
|
if "name" not in tool_descriptor:
|
|
raise ValueError("Tool descriptor must have a 'name' field")
|
|
name = tool_descriptor["name"]
|
|
if name in self._tools:
|
|
raise ValueError(
|
|
f"Tool '{name}' already registered. "
|
|
"Each tool name must be unique across all sources (MCP servers, gateways, catalogs)."
|
|
)
|
|
self._tools[name] = dict(tool_descriptor)
|
|
|
|
# Store MaterializedTool if provided (for direct Anthropic conversion)
|
|
if materialized_tool is not None:
|
|
self._materialized_tools[name] = materialized_tool
|
|
|
|
# Build normalized name mapping for Anthropic/OpenAI lookups
|
|
# e.g., "Google.Search" -> normalized key "Google_Search"
|
|
normalized_name = normalize_tool_name(name)
|
|
if normalized_name != name:
|
|
# Check for collision: if the normalized name already exists as a direct tool
|
|
# (e.g., registering "Google.Search" when "Google_Search" already exists),
|
|
# the normalized lookup would be ambiguous
|
|
if normalized_name in self._tools:
|
|
# The underscore version is registered directly, so normalized lookups
|
|
# should prefer that. Don't add to mapping to avoid ambiguity.
|
|
pass
|
|
elif normalized_name in self._normalized_to_original:
|
|
# Another dotted tool already maps to this normalized name
|
|
# e.g., "A.B" and "A_B" (as "A.B") would both normalize to "A_B"
|
|
# Keep the first mapping to avoid silent overwrites
|
|
pass
|
|
else:
|
|
self._normalized_to_original[normalized_name] = name
|
|
|
|
def add_tools(self, tools: list[MCPToolDefinition] | list[dict[str, Any]]) -> None:
|
|
for tool in tools:
|
|
self.add_tool(tool)
|
|
|
|
def list_tools_for_model(self, tool_format: ToolFormat = "openai") -> list[dict[str, Any]]:
|
|
if tool_format == "openai":
|
|
return self._to_openai_format()
|
|
elif tool_format == "anthropic":
|
|
return self._to_anthropic_format()
|
|
else:
|
|
raise ValueError(f"Tool format '{tool_format}' is not supported")
|
|
|
|
def _to_openai_format(self) -> list[dict[str, Any]]:
|
|
"""Convert stored MCP tools to OpenAI function calling format.
|
|
|
|
Note: Tool names are normalized (dots replaced with underscores) because
|
|
OpenAI function names don't allow dots.
|
|
"""
|
|
openai_tools: list[dict[str, Any]] = []
|
|
for tool in self._tools.values():
|
|
parameters = tool.get("inputSchema", {"type": "object", "properties": {}})
|
|
if self._strict_mode and isinstance(parameters, dict):
|
|
parameters = convert_to_strict_mode_schema(parameters)
|
|
|
|
# Normalize tool name for OpenAI (e.g., "Google.Search" -> "Google_Search")
|
|
# OpenAI function names don't allow dots
|
|
tool_name = normalize_tool_name(tool["name"])
|
|
|
|
openai_tool: dict[str, Any] = {
|
|
"type": "function",
|
|
"function": {
|
|
"name": tool_name,
|
|
"description": tool.get("description", ""),
|
|
"parameters": parameters,
|
|
},
|
|
}
|
|
if self._strict_mode:
|
|
openai_tool["function"]["strict"] = True
|
|
openai_tools.append(openai_tool)
|
|
|
|
return openai_tools
|
|
|
|
def _to_anthropic_format(self) -> list[dict[str, Any]]:
|
|
"""Convert stored tools to Anthropic format.
|
|
|
|
Uses direct to_anthropic() from arcade-core for Python tools (when MaterializedTool available),
|
|
falls back to convert_mcp_to_anthropic_tool() for MCP/remote tools (JSON descriptors only).
|
|
"""
|
|
anthropic_tools: list[dict[str, Any]] = []
|
|
for tool_name, tool_descriptor in self._tools.items():
|
|
# Python tools: use direct converter (we have MaterializedTool)
|
|
if tool_name in self._materialized_tools:
|
|
anthropic_tool = to_anthropic(self._materialized_tools[tool_name])
|
|
anthropic_tools.append(dict(anthropic_tool))
|
|
else:
|
|
# MCP/remote tools: convert from JSON descriptor (no MaterializedTool available)
|
|
# Used for tools from: load_mcp_remote_async(), load_from_stdio_async(),
|
|
# load_arcade_mcp_gateway_async(), or add_tool_definitions()
|
|
anthropic_tools.append(convert_mcp_to_anthropic_tool(tool_descriptor))
|
|
|
|
return anthropic_tools
|
|
|
|
def _resolve_tool_name(self, tool_name: str) -> str | None:
|
|
"""Resolve a tool name to its original registry key.
|
|
|
|
Handles both original names (e.g., "Google.Search") and
|
|
normalized names (e.g., "Google_Search" from Anthropic).
|
|
|
|
Args:
|
|
tool_name: The tool name to resolve.
|
|
|
|
Returns:
|
|
The original tool name if found, None otherwise.
|
|
"""
|
|
# First, try direct lookup
|
|
if tool_name in self._tools:
|
|
return tool_name
|
|
# Then, check if it's a normalized name (from Anthropic)
|
|
original_name = self._normalized_to_original.get(tool_name)
|
|
if original_name and original_name in self._tools:
|
|
return original_name
|
|
return None
|
|
|
|
def normalize_args(self, tool_name: str, args: dict[str, Any]) -> dict[str, Any]:
|
|
"""Apply schema defaults to arguments.
|
|
|
|
Fills in default values from the tool schema for:
|
|
- Missing parameters (key not in args)
|
|
- Null parameters (value is None), which OpenAI strict mode sends for optional params
|
|
|
|
This ensures that optional parameters with defaults are properly filled
|
|
even when the model sends null values.
|
|
"""
|
|
resolved_name = self._resolve_tool_name(tool_name)
|
|
tool = self._tools.get(resolved_name) if resolved_name else None
|
|
if not tool:
|
|
return args
|
|
|
|
schema = tool.get("inputSchema", {})
|
|
if not isinstance(schema, dict):
|
|
return args
|
|
|
|
properties = schema.get("properties", {})
|
|
if not isinstance(properties, dict):
|
|
return args
|
|
|
|
normalized = dict(args)
|
|
for prop_name, prop_schema in properties.items():
|
|
# Apply default if parameter is missing OR if it's null (None)
|
|
# OpenAI strict mode sends null for optional parameters that weren't provided
|
|
should_apply_default = (
|
|
isinstance(prop_schema, dict)
|
|
and "default" in prop_schema
|
|
and (prop_name not in normalized or normalized[prop_name] is None)
|
|
)
|
|
if should_apply_default:
|
|
normalized[prop_name] = prop_schema["default"]
|
|
return normalized
|
|
|
|
def get_tool_schema(self, tool_name: str) -> dict[str, Any] | None:
|
|
resolved_name = self._resolve_tool_name(tool_name)
|
|
return self._tools.get(resolved_name) if resolved_name else None
|
|
|
|
def has_tool(self, tool_name: str) -> bool:
|
|
return self._resolve_tool_name(tool_name) is not None
|
|
|
|
def resolve_tool_name(self, tool_name: str) -> str | None:
|
|
"""Public method to resolve a tool name to its original registry key.
|
|
|
|
This is useful for callers that need to look up tools by names
|
|
returned from providers (e.g., Anthropic returns underscore names).
|
|
|
|
Args:
|
|
tool_name: The tool name to resolve.
|
|
|
|
Returns:
|
|
The original tool name if found, None otherwise.
|
|
"""
|
|
return self._resolve_tool_name(tool_name)
|
|
|
|
def process_tool_call(self, tool_name: str, args: dict[str, Any]) -> tuple[str, dict[str, Any]]:
|
|
"""Resolve tool name and apply schema defaults in one step.
|
|
|
|
This combines name resolution (for Anthropic underscore -> dot conversion)
|
|
with schema default application.
|
|
|
|
Args:
|
|
tool_name: The tool name (may be in provider format like "Google_Search").
|
|
args: The arguments from the tool call.
|
|
|
|
Returns:
|
|
Tuple of (resolved_name, args_with_defaults).
|
|
resolved_name will be the original registered name (e.g., "Google.Search")
|
|
or the input name if not found in registry.
|
|
"""
|
|
resolved_name = self._resolve_tool_name(tool_name) or tool_name
|
|
args_with_defaults = self.normalize_args(tool_name, args)
|
|
return resolved_name, args_with_defaults
|
|
|
|
def tool_names(self) -> list[str]:
|
|
return list(self._tools.keys())
|
|
|
|
def tool_count(self) -> int:
|
|
return len(self._tools)
|