feat: AgentCore tool search plugin for Strands Agents

src/bedrock_agentcore/gateway/integrations/strands/plugins/__init__.py

@@ -0,0 +1,5 @@
+"""Gateway Strands plugins."""
+
+from .agentcore_tool_search import AgentCoreToolSearchPlugin
+
+__all__ = ["AgentCoreToolSearchPlugin"]

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/README.md

@@ -0,0 +1,138 @@
+# Strands AgentCore Tool Search Plugin
+
+A semantic tool discovery plugin for [Strands Agents](https://github.com/strands-agents/sdk-python) that uses the [Amazon Bedrock AgentCore Gateway](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-using-mcp-semantic-search.html) `x_amz_bedrock_agentcore_search` tool. This enables agents to dynamically load only the relevant tools for each invocation by deriving user intent from conversation history, even when hundreds of tools are registered on the gateway.
+
+## Features
+
+- **Semantic tool discovery** — uses AgentCore Gateway's built-in search to find relevant tools
+- **Intent-based loading** — derives user intent via LLM before searching
+- **No list_tools call** — tools are built directly from search results
+- **Pluggable intent provider** — swap the default intent provider with your own
+- **Agent model reuse** — by default, the intent classifier uses the same model as the parent agent
+
+## Installation
+
+```bash
+pip install agentcore-tool-search-plugin
+```
+
+## Usage
+
+```python
+from mcp_proxy_for_aws.client import aws_iam_streamablehttp_client
+from strands import Agent
+from strands.tools.mcp import MCPClient
+from agentcore_tool_search_plugin import AgentCoreToolSearchPlugin
+
+mcp_client = MCPClient(lambda: aws_iam_streamablehttp_client(
+    endpoint="https://<gateway-id>.gateway.bedrock-agentcore.<region>.amazonaws.com/mcp",
+    aws_region="us-east-1",
+    aws_service="bedrock-agentcore",
+))
+
+mcp_client.start()
+
+agent = Agent(plugins=[AgentCoreToolSearchPlugin(mcp_client=mcp_client)])
+
+agent("Find me afternoon flights to New York")
+```
+
+Or using a context manager:
+
+```python
+with mcp_client:
+    agent = Agent(plugins=[AgentCoreToolSearchPlugin(mcp_client=mcp_client)])
+    agent("Find me afternoon flights to New York")
+```
+
+## How It Works
+
+![Tool Search Flow](images/agentcore_tool_search_plugin.png)
+
+On each agent invocation:
+
+1. **User query** — The user sends a query to Strands agent.
+2. **Hook** — The agent triggers the `AgentCoreToolSearchPlugin` before model invocation
+3. **Derive intent** — The `IntentProvider` sends the last N messages from conversation history to the configured LLM to produce a concise intent string
+4. **Search gateway** — The intent is passed to AgentCore Gateway's `x_amz_bedrock_agentcore_search` tool to obtain most relevant tools.
+5. **Invoke LLM** — The agent invokes the LLM with the user query along with the matched tools from registered MCP targets (Lambda, API Gateway, MCP Server)
+
+Previously loaded tools are cleared before each search, so the agent always has the most relevant tools available.
+
+## Intent Provider
+
+An `IntentProvider` is responsible for analyzing conversation messages and producing a concise intent string that drives tool search. The plugin calls `derive_intent(messages, model)` before each invocation to determine what tools to load.
+
+### Default Intent Provider
+
+`StrandsIntentProvider` uses an LLM to classify the last few conversation messages into a concise intent string. By default it uses the agent's model.
+
+**Basic usage (uses the agent's model automatically):**
+
+```python
+from bedrock_agentcore.gateway.integrations.strands.plugins import AgentCoreToolSearchPlugin
+
+agent = Agent(plugins=[
+    AgentCoreToolSearchPlugin(mcp_client=mcp_client)
+])
+```
+
+**With a custom model for intent classification:**
+
+```python
+from strands.models.bedrock import BedrockModel
+from bedrock_agentcore.gateway.integrations.strands.plugins import AgentCoreToolSearchPlugin
+from bedrock_agentcore.gateway.integrations.strands.plugins.agentcore_tool_search.intent_providers import StrandsIntentProvider
+
+intent_model = BedrockModel(model_id="us.anthropic.claude-haiku-4-5-20251001-v1:0")
+agent = Agent(plugins=[
+    AgentCoreToolSearchPlugin(
+        mcp_client=mcp_client,
+        intent_provider=StrandsIntentProvider(model=intent_model),
+    )
+])
+```
+
+**With a custom system prompt:**
+
+```python
+from bedrock_agentcore.gateway.integrations.strands.plugins import AgentCoreToolSearchPlugin
+from bedrock_agentcore.gateway.integrations.strands.plugins.agentcore_tool_search.intent_providers import StrandsIntentProvider
+
+agent = Agent(plugins=[
+    AgentCoreToolSearchPlugin(
+        mcp_client=mcp_client,
+        intent_provider=StrandsIntentProvider(
+            system_prompt="Classify the user's intent in one sentence. Focus on the action, not details."
+        ),
+    )
+])
+```
+
+### Custom Intent Provider
+
+You can provide your own intent derivation strategy by subclassing `IntentProvider`:
+
+```python
+from bedrock_agentcore.gateway.integrations.strands.plugins.agentcore_tool_search.intent_providers import IntentProvider
+
+class MyIntentProvider(IntentProvider):
+    def derive_intent(self, messages: list[dict], model=None) -> str:
+        # custom logic to derive intent
+        return "intent string"
+
+agent = Agent(plugins=[
+    AgentCoreToolSearchPlugin(
+        mcp_client=mcp_client,
+        intent_provider=MyIntentProvider(),
+    )
+])
+```
+
+## Prerequisites
+
+- An AgentCore Gateway with **[semantic search](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-using-mcp-semantic-search.html) enabled**
+- Tools registered on the gateway with descriptions
+- AWS credentials with access to the gateway
+
+For more details, see the [AgentCore Gateway Documentation](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-building.html).

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/__init__.py

@@ -0,0 +1,6 @@
+"""AgentCore Tool Search plugin for Strands Agents."""
+
+from .intent_providers import StrandsIntentProvider, IntentProvider
+from .plugin import AgentCoreToolSearchPlugin
+
+__all__ = ["AgentCoreToolSearchPlugin", "IntentProvider", "DefaultIntentProvider"]

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/intent_providers/__init__.py

@@ -0,0 +1,6 @@
+"""Intent provider interfaces and implementations."""
+
+from .intent_provider import IntentProvider
+from .strands_intent_provider import StrandsIntentProvider
+
+__all__ = ["StrandsIntentProvider", "IntentProvider"]

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/intent_providers/intent_provider.py

@@ -0,0 +1,26 @@
+"""Intent provider abstract interface."""
+
+from abc import ABC, abstractmethod
+
+
+class IntentProvider(ABC):
+    """Abstract interface for deriving user intent from conversation messages.
+
+    Subclasses must implement the `derive_intent` method to analyze conversation
+    messages and return a concise intent string.
+    """
+
+    @abstractmethod
+    def derive_intent(self, messages: list[dict], model=None) -> str:
+        """Analyze conversation messages and return a concise intent string.
+
+        Args:
+            messages: List of conversation message dicts in Strands format.
+            model: Optional model instance from the parent agent. Implementations
+                can use this for LLM-based intent derivation.
+
+        Returns:
+            A plain text string describing the user's intent.
+            Returns empty string if intent cannot be determined.
+        """
+        ...

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/intent_providers/strands_intent_provider.py

@@ -0,0 +1,73 @@
+"""Default LLM-based intent provider implementation."""
+
+import logging
+
+from strands import Agent
+
+from .intent_provider import IntentProvider
+
+logger = logging.getLogger(__name__)
+
+INTENT_SYSTEM_PROMPT = (
+    "You are an intent classifier. Given the recent conversation messages, "
+    "produce a concise one-sentence description of what the user is trying to accomplish. "
+    "Focus on the type of task, not the specific details. "
+    "Reply with ONLY the intent description, nothing else."
+)
+
+
+class StrandsIntentProvider(IntentProvider):
+    """LLM-based intent provider that classifies the last N messages."""
+
+    def __init__(self, message_window: int = 5, model=None, system_prompt: str = INTENT_SYSTEM_PROMPT):
+        """Initialize StrandsIntentProvider.
+
+        Args:
+            message_window: Number of recent messages to consider.
+            model: Optional explicit model for intent classification.
+            system_prompt: System prompt for the intent classifier. Defaults to INTENT_SYSTEM_PROMPT.
+        """
+        self._message_window = message_window
+        self._explicit_model = model
+        self._system_prompt = system_prompt
+
+    def derive_intent(self, messages: list[dict], model=None) -> str:
+        """Derive intent using an LLM. Falls back to agent's model if no explicit model set."""
+        try:
+            recent_messages = messages[-self._message_window :] if messages else []
+            if not recent_messages:
+                return ""
+
+            kwargs = {"system_prompt": self._system_prompt, "tools": []}
+            # Priority: explicit model > agent's model > Strands default
+            resolved_model = self._explicit_model or model
+            if resolved_model:
+                kwargs["model"] = resolved_model
+
+            intent_agent = Agent(**kwargs)
+            response = intent_agent(self._format_messages_for_prompt(recent_messages))
+            return str(response).strip()
+        except Exception as e:
+            logger.error("Failed to derive intent: %s", e)
+            return ""
+
+    def _format_messages_for_prompt(self, messages: list[dict]) -> str:
+        """Format user messages into a text prompt for the intent LLM.
+
+        Only includes user-role messages to avoid leaking PII or sensitive data
+        from tool results or assistant responses.
+        """
+        parts = []
+        for msg in messages:
+            role = msg.get("role", "")
+            if role != "user":
+                continue
+            content = msg.get("content", [])
+            text = ""
+            if isinstance(content, list):
+                text = " ".join(
+                    block.get("text", "") for block in content if isinstance(block, dict) and "text" in block
+                )
+            if text.strip():
+                parts.append(text.strip())
+        return "\n".join(parts)

src/bedrock_agentcore/gateway/integrations/strands/plugins/agentcore_tool_search/plugin.py

@@ -0,0 +1,123 @@
+"""AgentCore tool search plugin for Strands Agents."""
+
+import json
+import logging
+
+from mcp.types import Tool as MCPTool
+from strands.hooks import BeforeInvocationEvent
+from strands.plugins import Plugin, hook
+from strands.tools.mcp import MCPClient
+from strands.tools.mcp.mcp_agent_tool import MCPAgentTool
+
+from .intent_providers import StrandsIntentProvider, IntentProvider
+
+logger = logging.getLogger(__name__)
+
+
+class AgentCoreToolSearchPlugin(Plugin):
+    """Plugin that dynamically loads tools from AgentCore Gateway based on semantic intent.
+
+    Args:
+        mcp_client: MCPClient connected to an AgentCore Gateway.
+        intent_provider: Strategy for deriving intent. Defaults to DefaultIntentProvider.
+    """
+
+    name = "agentcore-tool-search-plugin"
+
+    def __init__(
+        self,
+        mcp_client: MCPClient,
+        intent_provider: IntentProvider | None = None,
+    ):
+        """Initialize the plugin.
+
+        Args:
+            mcp_client: MCPClient connected to an AgentCore Gateway.
+            intent_provider: Strategy for deriving intent. Defaults to DefaultIntentProvider.
+        """
+        super().__init__()
+        self._intent_provider = intent_provider or StrandsIntentProvider()
+        self._mcp_client = mcp_client
+        self._loaded_tool_names: set[str] = set()
+
+    @property
+    def tools(self):
+        """Return empty list; tools are loaded dynamically via the hook."""
+        return []
+
+    @hook
+    def on_before_invocation(self, event: BeforeInvocationEvent) -> None:
+        """Derive intent, search gateway, and load matching tools."""
+        messages = event.messages or []
+
+        # Pass the agent's model to the intent provider
+        intent = self._intent_provider.derive_intent(messages, model=event.agent.model)
+        logger.info("Derived intent: %s", intent)
+
+        # Clear all previously loaded dynamic tools
+        for name in list(self._loaded_tool_names):
+            event.agent.tool_registry.registry.pop(name, None)
+        self._loaded_tool_names.clear()
+
+        if not intent:
+            return
+
+        try:
+            result = self._mcp_client.call_tool_sync(
+                tool_use_id="intent-search",
+                name="x_amz_bedrock_agentcore_search",
+                arguments={"query": intent},
+            )
+            agent_tools = self._build_tools_from_search_result(result)
+        except Exception as e:
+            logger.error("AgentCore Gateway search failed: %s", e)
+            return
+
+        for agent_tool in agent_tools:
+            try:
+                # Skip if a non-dynamic tool with this name already exists
+                if (
+                    agent_tool.tool_name in event.agent.tool_registry.registry
+                    and agent_tool.tool_name not in self._loaded_tool_names
+                ):
+                    logger.debug("Skipping tool %s: already registered as a static tool", agent_tool.tool_name)
+                    continue
+                event.agent.tool_registry.register_tool(agent_tool)
+                self._loaded_tool_names.add(agent_tool.tool_name)
+            except Exception as e:
+                logger.error("Failed to register tool %s: %s", agent_tool.tool_name, e)
+
+        logger.info("Loaded tools: %s", self._loaded_tool_names)
+
+    def _build_tools_from_search_result(self, result) -> list[MCPAgentTool]:
+        """Build MCPAgentTool objects from the gateway search response."""
+        tools = []
+        if not result or not isinstance(result, dict):
+            return tools
+
+        tool_defs = []
+        structured = result.get("structuredContent")
+        if isinstance(structured, dict) and "tools" in structured:
+            tool_defs = structured["tools"]
+        else:
+            for block in result.get("content", []):
+                if isinstance(block, dict) and "text" in block:
+                    try:
+                        data = json.loads(block["text"])
+                        if isinstance(data, dict) and "tools" in data:
+                            tool_defs = data["tools"]
+                            break
+                    except (json.JSONDecodeError, TypeError):
+                        continue
+
+        for tool_def in tool_defs:
+            if not isinstance(tool_def, dict) or "name" not in tool_def:
+                continue
+            mcp_tool = MCPTool(
+                name=tool_def["name"],
+                description=tool_def.get("description", ""),
+                inputSchema=tool_def.get("inputSchema", {"type": "object", "properties": {}}),
+            )
+            tools.append(MCPAgentTool(mcp_tool=mcp_tool, mcp_client=self._mcp_client))
+
+        return tools