MCP server for automating the adaptive multi-step form at sixmo.ru, including form completion, file upload, and result extraction
Single tool 'run_sixmo_form' has a clear action verb and comprehensive schema with all parameters typed and described. However, the tool description (127 chars) is borderline short and lacks critical context about prerequisites, side effects, and error recovery. The schema is well-structured with nested objects and proper constraints (minLength, format), but output schema is not documented, LLMs cannot predict the response structure. Tool annotations are present (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) which is good, but the destructiveHint=false is questionable for a tool that uploads files and modifies external state. Error handling returns a structured result but lacks actionable recovery guidance.
Automatically complete the adaptive multi-step form at sixmo.ru, upload the required file on step 2, and return the final identifier plus diagnostics.
Output schema not documented. Tool returns a complex nested object (success, artifacts, diagnostics, error) but LLMs have no schema to predict structure. This forces LLMs to guess field names and types for downstream reasoning.
Tool description lacks critical context. At 127 chars, it states WHAT but not WHEN to use it, prerequisites (file must exist, form must be accessible), or what happens on failure. LLMs cannot determine if this is the right tool without more guidance.
Destructive operation marked as non-destructive. Tool uploads files and modifies external form state (destructiveHint=false), but this is a write operation with side effects. Annotation should be true to signal to agents that this is irreversible.
Error responses lack recovery guidance. buildFatalResult() returns a generic 'MCP_TOOL_FAILED' code with only a message. No guidance on whether to retry, what caused the failure, or what the LLM should do next.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2025-06-18+ | v2 |
No pagination or result limits documented. If the tool returns large artifact lists (screenshots, DOM snapshots), there is no limit or pagination mechanism visible. Response could bloat context window.