Amazon Bedrock AgentCore Now Supports Stateful MCP with User Input and LLM Sampling

Amazon has introduced stateful MCP (Model Context Protocol) client capabilities on Bedrock AgentCore Runtime, enabling interactive, multi-turn agent workflows that require mid-execution pauses for user input and real-time feedback.

The update addresses a core limitation of previous stateless MCP implementations: agents couldn't pause execution to request clarification from users, generate content via LLM, or provide progress updates during long-running operations. Stateful mode removes this constraint by maintaining session continuity and enabling three new client capabilities.

The Three New Client Capabilities

Elicitation allows servers to pause execution and request structured user input mid-workflow. Servers send an elicitation/create request with a message and optional JSON schema describing expected response format. The client renders an input interface, and users can accept (providing data), decline, or cancel. Two modes are supported: form mode for direct data collection and URL mode for sensitive interactions like OAuth flows or payment processing.

Sampling enables servers to request LLM-generated completions from the client via sampling/createMessage. This allows tool logic on the server to use language model capabilities without maintaining its own model credentials. Practical uses include generating personalized summaries, creating natural-language explanations of structured data, or producing recommendations based on conversation context.

Progress notifications allow servers to report incremental progress during long-running operations using ctx.report_progress(progress, total). Clients can display updates as progress bars or status indicators, keeping users informed instead of watching blank screens during multi-step operations like cross-source data searches.

Technical Implementation

Enabling stateful mode requires a single configuration flag: setting stateless_http=False in server startup. When activated, AgentCore Runtime provisions a dedicated microVM for each user session with isolated CPU, memory, and filesystem. Sessions persist for up to 8 hours or until 15 minutes of inactivity (configurable via idleRuntimeSessionTimeout).

The protocol maintains continuity through an Mcp-Session-Id header. The server returns this identifier during initialization handshake; clients include it in every subsequent request to route back to the same session. All three capabilities are opt-in at the client level—clients declare supported capabilities during initialization, and servers must only use advertised capabilities.

Stateful mode contrasts with the previous stateless implementation, where each HTTP request was independent with no shared context between calls. Stateless mode remains suitable for simple tool serving, while stateful mode supports interactive, multi-turn workflows.

What This Means

This release completes MCP's bidirectional protocol implementation on AWS's managed agent platform. Teams can now build more sophisticated agents that interact naturally with users and leverage LLM capabilities without creating separate model connections. The session-based architecture with configurable timeouts provides flexibility for both short interactions and extended operations. For organizations already standardized on MCP, this reduces friction in moving agents from development to Bedrock production.