Why does MCP have no async callback support?

MCP was designed as synchronous request-response from the start. The protocol specification has no defined mechanism for a tool to notify the host that an async operation completed. The 2026 stateless RC focused on transport not async semantics.

What happens when a long-running MCP tool is called?

The agent blocks and waits with the HTTP connection open. If the operation takes longer than the connection timeout the call fails with no result. There is no way to poll, subscribe, or receive a callback when the work completes.

What workarounds exist for async MCP tool calls?

asyncmcp wraps MCP over AWS SNS+SQS queues. Some builders return a job ID from the tool and poll with a second tool. Neither is a protocol-level solution and both require custom infrastructure.

What kinds of tools need async support?

Code execution environments, large file processing, external API calls with delayed responses (payment processing, batch ML inference), and any operation that takes more than a few seconds to complete.

← Back to dashboard

clawsmith.com/signal/mcp-no-async-callbacks-long-running-tools-block

⚠ IssueWide Openai_agent_mcpLive

MCP has no async callback mechanism so long-running tool calls block the agent indefinitely

MCP was designed for synchronous request-response tool calls. When a tool needs minutes (code execution, file processing, external API waits) the agent hangs or times out with no way to receive a callback when the tool completes. Community workarounds use AWS SNS+SQS or webhooks but there is no protocol-level solution. The MCP stateless RC does not add async support.