MCP server that lets AI agents ask users questions mid-task via inline chat UI, Electron dialogs, or MCP elicitation
The server defines 5 tools with complete input schemas and descriptions. However, the quality varies significantly by tool. Three tools (user_input_dialog, user_input_elicitation, user_input_inline__present_ui) have well-structured schemas with typed parameters and meaningful descriptions. Two tools (__internal_do_not_use__submit_inline and user_input_inline__await_presented_ui) are incomplete or internal-only. Output schemas are not documented, critical for LLM planning. Error handling is minimal; no recovery guidance. The naming convention is inconsistent (some tools use 'user_input_' prefix, others use '__internal_do_not_use__'). The server lacks idempotency guidance, timeout documentation, and per-tool permission hints. Overall, definition quality is fair: core tools are usable, but completeness and LLM-optimization are lacking.
Internal tool for MCP Apps UI to store user responses. Do not call directly.
Opens an Electron GUI dialog for user input. The Electron binary is downloaded automatically on first use. Accepts the same options as the inline tool: prompt, title, options (string array of choices shown as buttons, or null for free-text only), and showAdditionalFreeInputButton (adds an "Other..." button that reveals a free-text input).
Ask the user a direct question via the MCP elicitation API and return their answer as plain text. Prefer this tool over user_input_dialog when the client supports elicitation.
Retrieves the user's response from a user_input_inline__present_ui call. Params: requestId (required). Blocks until the user submits.
Display an inline interactive form to collect user input via rich inline UI (requires MCP Apps support). All params required: prompt (string), title (string or null), options (string array of choices or null for free-text only), showAdditionalFreeInputButton (bool — show an extra "Other..." option for pure free-text entry). After calling, you MUST call user_input_inline__await_presented_ui with the returned requestId.
No output schemas documented. LLMs cannot infer what these tools return, forcing them to guess field names and types for downstream tool calls. For example, user_input_dialog returns a user response, but the schema (type, format, structure) is invisible to the LLM.
Tool naming is inconsistent and confusing. __internal_do_not_use__submit_inline uses a non-standard prefix that violates verb_noun convention. user_input_inline__present_ui and user_input_inline__await_presented_ui use double underscores and are difficult to parse. Consider rename to: present_inline_form, await_form_response, submit_form_response.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | 2025-06-18+ | v2 |
| 2026-03-09 | F | 48 | - | v1 |
No error handling guidance. Descriptions do not explain what errors are possible, how the LLM should recover, or whether operations are retryable. For example, if user cancels the dialog, does user_input_dialog return a special value or throw an error? How should the LLM respond?
No timeout or latency documentation. user_input_dialog and user_input_inline__await_presented_ui are blocking operations that wait for user input. The descriptions do not specify maximum wait time, how long the server will block, or what happens on timeout. This leaves the LLM unaware of potential hangs.
Interdependent tool pair (present_ui / await_presented_ui) has no explicit sequencing guidance. The description of user_input_inline__present_ui says 'you MUST call user_input_inline__await_presented_ui with the returned requestId', but LLMs often miss imperative phrasing. Consider using Multi-Round-Trip Requests (result: 'input_required') instead, which the MCP spec supports and which signal to the client that the request is incomplete.
Permissions and scope not declared. None of the tools declare what permissions they require (e.g. 'read:user_input', 'write:user_input'). Agents should know what authority is needed to call each tool.
No idempotency guidance. If an LLM retries user_input_dialog due to a timeout, will it prompt the user twice? Or is the second call a no-op? Idempotency contract is not documented.