GitHub user reeb-dev added a comment to the discussion: [General][Java SDK (Component)] Dubbo-MCP: Model Context Protocol (MCP) server for Apache Dubbo (Connecting Cursor, Claude Code & AI Coding Agents)
This is a useful developer-experience direction, especially because generic invocation removes the friction of maintaining a local client dependency graph. The staging/production boundary is the part I would make even more explicit: an MCP server turns an AI client into a powerful RPC caller, so its security model should be as deliberate as the Dubbo Admin model. A few guardrails would make the production story easier to trust: 1. Do not expose one generic `invoke` tool in production. Generate narrow MCP tools from an allowlist of service/method pairs, with typed input schemas and an explicit read-only classification. That gives clients less room to discover or call an unintended method. 2. Bind authorization to the authenticated caller, not to the MCP client alone. The policy should decide which service, method, parameter shape, environment, and data classification a user may access. 3. Make every production call auditable: correlation id, principal, tool name, normalized arguments with sensitive fields redacted, result status, and latency. Keep logs separate from prompt/context capture so business data is not unintentionally retained. 4. Use short-lived credentials and require an explicit environment selector. A default that can reach production is risky; development should be the safe default, with a visible warning and stronger approval for staging/production. 5. Treat tool discovery as sensitive metadata too. Even a method signature can reveal internal capabilities, so discovery should follow the same authorization policy as invocation. The two-stage discovery approach is valuable for token efficiency, but a published threat model plus a small end-to-end authorization test matrix would likely be the strongest next addition to the project. GitHub link: https://github.com/apache/dubbo/discussions/16565#discussioncomment-19411016 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
