anthilll-inside
MCP server that reproduces SSE response-size hangs with Anthropic MCP clients by providing a configurable 'generate_lipsum' tool.
README
anthilll-inside
PoC MCP server for reproducing Anthropic hosted MCP client SSE response-size hangs.
The server uses @modelcontextprotocol/sdk@1.29.0 with stateless StreamableHTTPServerTransport in default SSE mode (enableJsonResponse: false), matching the conditions described in the bug report. A single tool — generate_lipsum — returns a configurable number of pseudo-Latin sentences so you can bracket the suspected ~10–333 KiB threshold.
Quick start (local)
cp .env.example .env
npm install
npm run dev
The MCP endpoint is http://127.0.0.1:3000/mcp.
Tool: generate_lipsum
sentence_count |
Approx. SSE response |
|---|---|
| 50 | ~8 KiB |
| 500 | ~55 KiB |
| 2000 | ~222 KiB |
| 3000 | ~333 KiB |
Test with curl (SSE)
# Initialize + list tools (small SSE response)
curl -sS -X POST http://127.0.0.1:3000/mcp \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}'
# Call generate_lipsum with 3000 sentences (~333 KiB SSE — matches bug report)
curl -sS -X POST http://127.0.0.1:3000/mcp \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
-d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"generate_lipsum","arguments":{"sentence_count":3000}}}'
Workaround mode (JSON)
Set ENABLE_JSON_RESPONSE=true to serve Content-Type: application/json instead of SSE — the known workaround from the bug report.
Logging
This PoC logs everything useful for layer-by-layer debugging:
Server-side (Pino structured logs to stdout):
request_received/request_completed— per-request access log withreqId,responseBytes,durationMs,writableEndedjsonrpc_request_body— inbound JSON-RPC body (Bearer tokens redacted)mcp_request_dispatch— method,Acceptheader, transport modetool_invoked/lipsum_generated/tool_response_ready— tool lifecycle with byte counts
Client-visible (MCP notifications/message via logging capability):
tool_start,lipsum_generated,tool_response_readyevents withpayloadBytesandtoolReqId
Note: client-visible log notifications are only delivered in SSE mode; they are dropped when ENABLE_JSON_RESPONSE=true (SDK behavior).
Docker
docker build -t mcp-sse-repro .
docker run --rm -p 3000:3000 mcp-sse-repro
Kubernetes (Helm)
docker build -t your-registry/mcp-sse-repro:latest .
docker push your-registry/mcp-sse-repro:latest
helm upgrade --install mcp-sse-repro ./helm/mcp-sse-repro \
--set image.repository=your-registry/mcp-sse-repro \
--set image.tag=latest \
--set ingress.enabled=true \
--set ingress.hosts[0].host=mcp.example.com \
--set env.ALLOWED_HOSTS=mcp.example.com
When deploying behind an ingress, set ALLOWED_HOSTS to your public hostname for DNS rebinding protection.
Configuration
| Variable | Default | Description |
|---|---|---|
PORT |
3000 |
HTTP listen port |
HOST |
127.0.0.1 |
Bind address (0.0.0.0 in container/K8s) |
ALLOWED_HOSTS |
— | Comma-separated allowed Host headers |
ENABLE_JSON_RESPONSE |
false |
true = JSON workaround; false = SSE repro |
LOG_LEVEL |
debug |
Pino log level |
LOG_REQUEST_BODY |
true |
Log inbound JSON-RPC bodies |
LOG_RESPONSE_BODY_PREVIEW_BYTES |
2048 |
Max bytes of request body to log |
MAX_SENTENCE_COUNT |
10000 |
Upper bound for sentence_count |
Bug report
See anthropic-mcp-sse-bug-report.md for the full write-up with layer-by-layer evidence. File issues at anthropics/claude-ai-mcp.
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。