Skip to content

πŸ™Œ OpenHands: AI-Driven Development

[eval]: disable MCP for SWE-Bench evaluation (#8574)

Xingyao Wang2ecc39ffccc0
diff --git a/evaluation/benchmarks/swe_bench/run_infer.py b/evaluation/benchmarks/swe_bench/run_infer.py
index 887cf90f97..8e23125945 100644
--- a/evaluation/benchmarks/swe_bench/run_infer.py
+++ b/evaluation/benchmarks/swe_bench/run_infer.py
@@ -261,6 +261,7 @@ def get_config(
enable_jupyter=False,
enable_browsing=RUN_WITH_BROWSING,
enable_llm_editor=False,
+ enable_mcp=False,
condenser=metadata.condenser_config,
enable_prompt_extensions=False,
)
diff --git a/openhands/cli/main.py b/openhands/cli/main.py
index bed9a731f2..8ba7ef532f 100644
--- a/openhands/cli/main.py
+++ b/openhands/cli/main.py
@@ -251,7 +251,8 @@ async def run_session(
)
# Add MCP tools to the agent
- await add_mcp_tools_to_agent(agent, runtime, memory, config.mcp)
+ if agent.config.enable_mcp:
+ await add_mcp_tools_to_agent(agent, runtime, memory, config.mcp)
# Clear loading animation
is_loaded.set()
diff --git a/openhands/core/config/agent_config.py b/openhands/core/config/agent_config.py
index 6800e41690..c6222fb51e 100644
--- a/openhands/core/config/agent_config.py
+++ b/openhands/core/config/agent_config.py
@@ -28,6 +28,8 @@ class AgentConfig(BaseModel):
"""Whether to enable finish tool"""
enable_prompt_extensions: bool = Field(default=True)
"""Whether to enable prompt extensions"""
+ enable_mcp: bool = Field(default=True)
+ """Whether to enable MCP tools"""
disabled_microagents: list[str] = Field(default_factory=list)
"""A list of microagents to disable (by name, without .py extension, e.g. ["github", "lint"]). Default is None."""
enable_history_truncation: bool = Field(default=True)
diff --git a/openhands/core/main.py b/openhands/core/main.py
index 80a7a0144b..e9cbc45b26 100644
--- a/openhands/core/main.py
+++ b/openhands/core/main.py
@@ -129,7 +129,8 @@ async def run_controller(
)
# Add MCP tools to the agent
- await add_mcp_tools_to_agent(agent, runtime, memory, config.mcp)
+ if agent.config.enable_mcp:
+ await add_mcp_tools_to_agent(agent, runtime, memory, config.mcp)
replay_events: list[Event] | None = None
if config.replay_trajectory_path:
diff --git a/openhands/server/session/agent_session.py b/openhands/server/session/agent_session.py
index e1a11dc3b7..e388b9c37a 100644
--- a/openhands/server/session/agent_session.py
+++ b/openhands/server/session/agent_session.py
@@ -149,7 +149,7 @@ class AgentSession:
# NOTE: this needs to happen before controller is created
# so MCP tools can be included into the SystemMessageAction
- if self.runtime and runtime_connected:
+ if self.runtime and runtime_connected and agent.config.enable_mcp:
await add_mcp_tools_to_agent(agent, self.runtime, self.memory, config.mcp)
if replay_json:
diff --git a/tests/unit/test_agent_controller.py b/tests/unit/test_agent_controller.py
index 79998cd0b8..c5cca37be2 100644
--- a/tests/unit/test_agent_controller.py
+++ b/tests/unit/test_agent_controller.py
@@ -57,6 +57,10 @@ def mock_agent():
agent.llm.metrics = Metrics()
agent.llm.config = AppConfig().get_llm_config()
+ # Add config with enable_mcp attribute
+ agent.config = MagicMock(spec=AgentConfig)
+ agent.config.enable_mcp = True
+
# Add a proper system message mock
system_message = SystemMessageAction(
content='Test system message', tools=['test_tool']
diff --git a/tests/unit/test_agent_session.py b/tests/unit/test_agent_session.py
index b3e1220f6f..66fa9b40c1 100644
--- a/tests/unit/test_agent_session.py
+++ b/tests/unit/test_agent_session.py
@@ -35,6 +35,7 @@ def mock_agent():
# Configure the agent config
agent_config.disabled_microagents = []
+ agent_config.enable_mcp = True
# Set up the chain of mocks
llm.metrics = metrics