mcp-uplift
Wraps a legacy MCP stdio server and presents it as a modern stateless 2026-07-28 protocol server, translating the legacy handshake and server-initiated requests.
README
mcp-uplift
MCP 2026-07-28 removed the initialization handshake, sessions, ping,
logging configuration, resource subscriptions, and server-initiated requests.
Older MCP servers still use those protocol features. mcp-uplift wraps one
legacy stdio server and presents it as a modern stateless server.
Security: wrapping a server executes that server with your operating-system permissions. This bridge is not a sandbox. Only run packages and commands you trust.
Demo

The unmodified official @modelcontextprotocol/server-filesystem running behind
the 2026-07-28 protocol: server/discover is synthesized from the legacy
handshake, the server's own roots/list request becomes a keyed input_required
result, and answering it resumes the call and returns all 14 tools.
Usage
Run without installing:
npx mcp-uplift <legacy-command> [args...]
For example:
npx mcp-uplift npx -y @modelcontextprotocol/server-filesystem /tmp
The wrapped command receives a minimal environment by default. Forward a needed
credential explicitly, before --:
npx mcp-uplift --env BRAVE_API_KEY -- npx -y @modelcontextprotocol/server-brave-search
--inherit-env is available for compatibility but exposes every environment
variable to the wrapped process. Run npx mcp-uplift --help for resource and
timeout controls.
Client configuration
Before, a client launches the legacy server directly:
{
"mcpServers": {
"files": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
}
}
}
After, launch the same command through mcp-uplift:
{
"mcpServers": {
"files": {
"command": "npx",
"args": ["mcp-uplift", "npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
}
}
}
Compatibility
Because 2026-07-28 is stateless, every request must carry its own envelope in
params._meta: the protocol version and the client's capabilities, plus
optional client identity. A request missing them is rejected rather than
guessed at.
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/list",
"params": {
"_meta": {
"io.modelcontextprotocol/protocolVersion": "2026-07-28",
"io.modelcontextprotocol/clientCapabilities": { "roots": {} },
"io.modelcontextprotocol/clientInfo": { "name": "my-client", "version": "1.0.0" }
}
}
}
If a wrapped server needs a capability the client did not declare, the bridge
returns -32021 naming what was required, rather than failing deep inside the
call.
| Feature | Behavior |
|---|---|
Legacy initialize |
Performed once and exposed as server/discover |
| Per-request version metadata | Validated as 2026-07-28, else -32022 |
| Tools, prompts, and resources | Forwarded with modern result and cache metadata |
| Legacy resource-not-found errors | Mapped from -32002 to -32602 |
| Sampling, elicitation, and roots requests | Translated into multi-round-trip input_required results |
| Removed methods | Rejected with -32601 method-not-found |
| Upstream failures | Returned as -32603 internal errors |
Answering a server's question
A legacy server can interrupt its own call to ask the client something.
Because a modern client cannot receive that push, the bridge returns an
input_required result instead, keyed by request id:
{
"resultType": "input_required",
"inputRequests": {
"ir_1000": {
"method": "elicitation/create",
"params": { "message": "Which environment?" }
}
},
"requestState": "c0a2db7b-62bc-4420-b2e9-31da87f8f999"
}
To resume, send the same request again with the requestState you were
given and an inputResponses map using those same keys. Each value is the
response body itself, not wrapped in result:
{
"name": "deploy",
"arguments": {},
"requestState": "c0a2db7b-62bc-4420-b2e9-31da87f8f999",
"inputResponses": {
"ir_1000": { "action": "accept", "content": { "env": "prod" } }
}
}
The parked call then finishes and returns its ordinary "complete" result.
Shapes per method: elicitation/create takes { action, content } with
action one of accept, decline, or cancel; roots/list takes
{ roots: [...] }; sampling/createMessage takes { model, role, content }.
The resumed request must match the original, every key must be answered, and
requestState is single-use and expires, so an invalid resume is rejected
rather than half-applied.
Known limitations
Real-world validation confirmed discovery and tool listing against 39 distinct
legacy MCP packages. The official filesystem server also completed a real
roots/list MRTR round trip and returned all 14 tools. These are tested
examples, not a guarantee that every server or session-dependent behavior can
be translated.
- All legacy calls are serialized. The
legacy protocol never links a
sampling/createMessage,elicitation/create, orroots/listrequest back to the call that caused it, so the bridge keeps one call in flight through all of its MRTR rounds. This favors correct attribution over throughput. - Legacy notifications have no home and are dropped.
2026-07-28moved streamed notifications onto a dedicatedsubscriptions/listenstream; this bridge is a plain one-request-one-response stdio proxy and does not implement that stream, sonotifications/progressandnotifications/messagefrom the wrapped server are discarded rather than delivered. - Real-server checks require downloads and remain outside the offline suite.
- The kill switch terminates the launched process tree on POSIX and Windows, but trusted code can deliberately daemonize into a new OS process session. Use an OS sandbox or container when stronger confinement is required.
Security reports
Do not include credentials or exploit details in a public issue. Report a suspected vulnerability privately through the repository's GitHub security advisory page. Ordinary bugs can use the public issue tracker.
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。