Problem Statement
MCPClient._start_session() calls session.initialize() but discards the InitializeResult. This means Host applications have no
way to access the server's instructions field — a standard MCP feature for providing LLMs with contextual knowledge about tool interdependencies and operational constraints.
# mcp_client.py line ~504
await session.initialize() # InitializeResult (including instructions) is discarded
MCP clients like Claude Code, VSCode, and Goose inject server instructions into the LLM's system prompt. Strands-based Host applications cannot do this because the data is inaccessible.
Proposed Solution
Store InitializeResult.instructions on the MCPClient instance:
# In MCPClient._start_session():
result = await session.initialize()
self.server_instructions = result.instructions or ""
Host applications can then access it:
with mcp_client:
instructions = mcp_client.server_instructions
agent = Agent(
system_prompt=f"...\n{instructions}",
tools=mcp_client.list_tools_sync(),
)Use Case
I'm building an Agent (Host) that connects to an MCP Server with detailed workflow instructions (multi-phase slide generation). The server returns these via the standard instructions field in InitializeResult, but my Agent cannot access them through MCPClient.
Without this, I have to either:
- Hardcode the MCP Server's instructions in the Agent's system prompt (breaks DRY, drifts)
- Make a separate HTTP call to the MCP Server just to get instructions (wasteful)
Per the MCP spec, the Host is responsible for deciding how to use instructions — but the Client (MCPClient) needs to make them accessible first.
References:
Alternatives Solutions
- Host could make a separate
initialize HTTP request to get instructions, but this creates a redundant connection
- Instructions could be duplicated in every tool description, but this wastes context window and violates MCP best practices
Additional Context
The fix is ~3 lines: store the result of session.initialize() and expose instructions as a public attribute. The SDK's responsibility is only to make the data accessible — how/whether to inject into the system prompt remains the Host's decision.
Problem Statement
MCPClient._start_session()callssession.initialize()but discards theInitializeResult. This means Host applications have noway to access the server's
instructionsfield — a standard MCP feature for providing LLMs with contextual knowledge about tool interdependencies and operational constraints.MCP clients like Claude Code, VSCode, and Goose inject server instructions into the LLM's system prompt. Strands-based Host applications cannot do this because the data is inaccessible.
Proposed Solution
Store
InitializeResult.instructionson theMCPClientinstance:Host applications can then access it:
Use Case
I'm building an Agent (Host) that connects to an MCP Server with detailed workflow instructions (multi-phase slide generation). The server returns these via the standard
instructionsfield inInitializeResult, but my Agent cannot access them throughMCPClient.Without this, I have to either:
Per the MCP spec, the Host is responsible for deciding how to use instructions — but the Client (MCPClient) needs to make them accessible first.
References:
Alternatives Solutions
initializeHTTP request to get instructions, but this creates a redundant connectionAdditional Context
The fix is ~3 lines: store the result of
session.initialize()and exposeinstructionsas a public attribute. The SDK's responsibility is only to make the data accessible — how/whether to inject into the system prompt remains the Host's decision.