A shim that connects to external MCP servers via STDIO and proxies their tools through SSE transport. Reads settings[0].local_mcpServers from shared config to discover servers and their tools.
This MCP server provides a single dynamically-proxied tool ('local_<server>_<tool>') that bridges external MCP servers. Critical issues: (1) The tool definition itself is not explicit in the source code, it is dynamically generated at runtime based on discovered external servers. The code mentions '_discover_tools' and 'tool_registry' but the actual tool registration logic is incomplete/cut off in the provided source. (2) The tool description provided ('Dynamic proxy tools bridged from external MCP servers...') is present but vague and does not follow LLM-optimized conventions (it is 165 chars, within baseline range, but reads more like implementation notes than user-facing guidance). (3) No input schema is visible in the source, the code comment states 'Operation arguments are forwarded to the external server verbatim (pass-through proxy); no schema-level validation happens on our side', which means this tool accepts arbitrary JSON and provides no schema constraints to the LLM. (4) No output schema is documented.
Dynamic proxy tools bridged from external MCP servers. Tool names are dynamically generated based on server name and original tool name. Description can be overridden per-server via ai_description config parameter.
Tool definition is inferred/dynamic, not explicitly registered in code. Tool names and schemas are generated at runtime; no static registration visible.
No input schema visible or documented. Tool accepts arbitrary arguments as pass-through with no LLM-visible constraints. Comment states 'no schema-level validation happens on our side'.
No output schema documented. LLM cannot reason about what fields to expect from bridged tool calls.
Tool description is generic and implementation-focused ('Dynamic proxy tools bridged from external MCP servers...') rather than user/LLM-optimized. Does not explain WHEN to use, WHAT it does from user perspective, or WHAT to expect back.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 29 | <=2025-11-25 | v2 |
Tool name 'local_<server>_<tool>' is templated and does not start with a clear action verb. LLM cannot infer intent from name alone.
No error handling guidance documented. If a bridged tool fails, what should the LLM do next?
Tool accepts arbitrary parameters with no validation. Comment (#B3) states 'Operation arguments are forwarded to the external server verbatim... no schema-level validation happens on our side'. LLM has no visibility into valid parameters for any bridged tool.
Tool definition code is incomplete in provided source (cut off mid-function). Cannot verify actual tool registration or schema construction.