AgentCore MCP Server
A stateless MCP server sample for the 2026-07-28 protocol, deployable to Amazon Bedrock AgentCore Runtime and probed via SigV4-signed requests.
README
AgentCore MCP 2026-07-28 sample
A minimal, reproducible sample for exercising the MCP 2026-07-28 (stateless)
protocol on Amazon Bedrock AgentCore — both a Runtime-hosted MCP server and a
SigV4 probe you can point at a Gateway or a Runtime.
MCP 2026-07-28 makes the protocol stateless: no initialize handshake, no
protocol-level Mcp-Session-Id, every request self-contained. The two artifacts here
let you see that end to end.
Contents
| File | What it is |
|---|---|
server.py |
A stateless MCP server built on the mcp Python SDK v2 (MCPServer), ready to deploy to AgentCore Runtime. |
requirements.txt |
Runtime dependencies (mcp>=2.0.0). |
mcp_probe.py |
A tiny SigV4-signed MCP client for probing an AgentCore Gateway /mcp endpoint or a Runtime /invocations endpoint with IAM auth. |
Tested with
mcp==2.0.0(LATEST_PROTOCOL_VERSION == "2026-07-28"),bedrock-agentcore==1.19.0,bedrock-agentcore-starter-toolkit==0.3.11, Python 3.13, inus-east-1. Pin yourmcpversion — a different major version changes the server API (see note below).
The server
server.py uses the v2 high-level API. Note that mcp v2 removed FastMCP —
the class is now MCPServer (mcp.server.mcpserver). It serves a stateless
streamable-HTTP app at 0.0.0.0:8000/mcp, which is what AgentCore Runtime expects.
Key choices:
stateless_http=True— the 2026-07-28 model and the recommended AgentCore default.enable_dns_rebinding_protection=False— the AgentCore front door forwards a non-localhostHostheader, so the v2 default host validation must be relaxed for the container.
Run locally
python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python server.py # serves http://0.0.0.0:8000/mcp
Compare the two protocol shapes locally:
# old protocol -> SSE, no envelope
curl -s http://localhost:8000/mcp \
-H 'Content-Type: application/json' -H 'Accept: application/json,text/event-stream' \
-H 'MCP-Protocol-Version: 2025-03-26' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'
# 2026-07-28 stateless -> plain JSON with resultType/cacheScope envelope
curl -s http://localhost:8000/mcp \
-H 'Content-Type: application/json' -H 'Accept: application/json,text/event-stream' \
-H 'MCP-Protocol-Version: 2026-07-28' -H 'Mcp-Method: tools/list' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{"_meta":{
"io.modelcontextprotocol/protocolVersion":"2026-07-28",
"io.modelcontextprotocol/clientInfo":{"name":"local","version":"1.0"},
"io.modelcontextprotocol/clientCapabilities":{}}}}'
Deploy to AgentCore Runtime
Using the starter toolkit's direct_code_deploy (no local container build; IAM inbound
auth is the default when you omit a JWT authorizer):
pip install bedrock-agentcore-starter-toolkit
agentcore configure -e server.py -n <agent-name> --protocol MCP \
-r us-east-1 -rf requirements.txt --non-interactive
agentcore deploy
There is no MCP-version flag on a Runtime — the runtime declares only
serverProtocol: MCP. Which MCP versions the server speaks is decided entirely by the
mcp SDK version in your requirements.txt.
The probe
mcp_probe.py SigV4-signs a POST (default service bedrock-agentcore) so you can hit
IAM-auth AgentCore endpoints with your own credentials — no Cognito/OAuth setup.
pip install boto3
# Point at a Gateway MCP endpoint or a Runtime /invocations URL:
URL="https://<gateway-id>.gateway.bedrock-agentcore.us-east-1.amazonaws.com/mcp"
# 2026-07-28 stateless tools/list
python mcp_probe.py --url "$URL" --proto 2026-07-28 --mcp-method tools/list --method tools/list \
--meta '{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientInfo":{"name":"probe","version":"1.0"},"io.modelcontextprotocol/clientCapabilities":{}}'
# tools/call
python mcp_probe.py --url "$URL" --proto 2026-07-28 --mcp-method tools/call --mcp-name add_numbers \
--method tools/call --name add_numbers --arguments '{"a":2,"b":40}' \
--meta '{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientInfo":{"name":"probe","version":"1.0"},"io.modelcontextprotocol/clientCapabilities":{}}'
For a Runtime, use the data-plane URL with the URL-escaped runtime ARN:
https://bedrock-agentcore.us-east-1.amazonaws.com/runtimes/<URL-ESCAPED-ARN>/invocations?qualifier=DEFAULT
Gotchas this sample demonstrates
- A stateless
2026-07-28request must carryprotocolVersion,clientInfo, andclientCapabilitiesinside_metaon every call — omit any and you get-32602. A real client SDK does this for you; hand-rolling from an abbreviated example fails. mcpv2 droppedFastMCP→ useMCPServer. (Pinmcp>=1.28,<2if you must keepFastMCP.)- v2 DNS-rebinding protection blocks the non-localhost
Hostbehind AgentCore → disable it for the container. - On the wire: old protocol responds via SSE with no envelope;
2026-07-28responds with plain JSON carryingresultType/cacheScope(andstructuredContent/ttlMswhere applicable).
Prerequisites
- Python 3.10+ (3.13 tested)
- AWS credentials in your environment with permission to deploy/invoke AgentCore
- For probing: an AgentCore Gateway or Runtime configured with IAM (SigV4) inbound auth
References
- How AgentCore Gateway supports the MCP 2026-07-28 spec
- Deploy MCP servers in AgentCore Runtime
- MCP 2026-07-28 specification & changelog
License
MIT-0. See LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。