A Neovim plugin that provides a hub for managing and interacting with Model Context Protocol (MCP) servers
mcphub.nvim is a Neovim plugin that acts as a bridge to MCP servers, exposing two meta-tools (use_mcp_tool, access_mcp_resource) that allow calling arbitrary tools on remote MCP servers. The core issue: these are not true tool definitions with concrete schemas and error handling, they are pass-through proxies. Tool definitions are inferred/dynamic, not statically registered. Naming is acceptable (verb_noun pattern), but descriptions are minimal (10-11 chars baseline for both tools, well below the 34-392 char range of production tools). Schemas are present but sparse: no output schema documented, no error handling guidance, no per-tool discovery or composition support. Transport is unknown (likely STDIO via Neovim plugin architecture). This limits usability for production agentic workloads.
get resources on MCP servers.
calls tools on MCP servers.
Descriptions are critically short (10-11 chars). 'calls tools on MCP servers' and 'get resources on MCP servers' provide no actionable guidance on WHEN to use each tool, what prerequisites exist, or what the LLM should expect. Production baseline is 34-392 chars; these fall far below the 20-char minimum threshold.
No output schema documented. Tools return results from remote MCP servers, but the response structure is never specified. LLMs cannot plan downstream calls or extract fields without knowing what fields exist. Production rule: 100% of A+ tools document return types.
No error handling or recovery guidance. Tool descriptions say nothing about failure modes, retryability, or what to do if a remote server is unreachable or a tool does not exist. Per pattern:recovery-guide, errors must tell the LLM what to do next.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 47 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
tool_input parameter has additionalProperties=false but no minProperties/maxProperties constraints, and the description 'Input object for the tool call' is generic. LLMs cannot infer what fields are required vs optional for arbitrary remote tools. This is inherent to the proxy design and cannot be fixed without runtime schema discovery.
Tool definitions are inferred dynamically at runtime (tools are created in create_static_tools() based on hub state), not statically registered. Per scoring rules, inferred tool definitions are capped at 50 overall. This prevents static verification and upstream MCP client discovery.
No parameter descriptions for server_name and tool_name are actionable. 'Name of the server to call the tool on. Must be from one of the available servers' assumes the LLM knows which servers are available, but this is only revealed at runtime via a system prompt. No enum constraint or discovery mechanism is offered.
Risk classification is present (IRREVERSIBLE for use_mcp_tool, READ_ONLY for access_mcp_resource) but tool descriptions do NOT mention this. Per pattern:command-tool, if a tool modifies state, the description must say so explicitly. Relying on side-band risk metadata instead of explicit docstring guidance is weak.