Relay MCP has well-structured tool definitions with complete input schemas and clear descriptions. All 8 tools are explicitly registered with name, description, and inputSchema. Descriptions average ~120 chars and explain WHAT the tool does and WHEN to use it. However, several tools lack parameter-level descriptions (figma_set_fills, figma_set_strokes have minimal 'id' and 'fills'/'strokes' descriptions), and output schemas are not documented. Error handling is present but generic. The figma_execute tool is a high-risk escape hatch with minimal guardrails. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification (READ_ONLY vs WRITE vs IRREVERSIBLE).

• Add output schema documentation to each tool. Example: figma_search_components should document it returns {components: [{id, name, type, childCount}, ...], total: number}. This enables LLMs to chain calls and extract IDs for downstream operations.
• Enhance figma_set_fills and figma_set_strokes parameter descriptions. Provide a concrete Paint object example in the description: 'Figma Paint[] array, e.g. [{type: "SOLID", color: {r: 1, g: 0, b: 0}, opacity: 1}] for solid red, or [{type: "GRADIENT", ...}] for gradients.'
• Add tool annotations to the tool definitions. Mark figma_search_components, figma_find_node, figma_export_node as readOnlyHint: true. Mark figma_set_* and figma_execute as destructiveHint: true. This signals to MCP clients which tools mutate state.
• Restrict figma_execute with a confirmation pattern. Require agents to call a separate 'confirm_execute' tool with the code snippet before execution, or add a dry_run parameter that returns the result without applying mutations.
• Improve error messages with recovery guidance. When Relay times out, suggest 'Check that the Figma plugin is open and worker mode is enabled. Retry the call.' When the plugin reports failure, include the plugin's error message and suggest relevant tools (e.g., 'Node not found. Try figma_search_components to locate it.').