Agent-eyes provides 13 tools with basic schemas and descriptions, but falls short of production quality. Most tools have descriptions (10-150 chars, below the 194-char baseline), and input schemas are present but lack depth. Parameter descriptions are sparse or missing entirely. Error handling is minimal, tools return JSON with isError flags but lack recovery guidance. No tool annotations (readOnlyHint/destructiveHint). The ai_orchestrate_workflow and ai_plan_goal tools appear to be inferred from test fixtures rather than explicitly registered, capping their scores at 50. Overall, the server reads like a working prototype rather than a production-grade agent toolkit.

• Add tool annotations: mark browser_interact and browser_evaluate_js with destructiveHint=true; mark all read-only tools with readOnlyHint=true. This enables LLMs to reason about side effects without parsing descriptions.
• Expand parameter descriptions to 50-100 chars each. E.g., browser_interact's 'target' should read: 'CSS selector (e.g., button.submit) or data-agent-id attribute value identifying the element to interact with.'
• Document output schemas for all tools. E.g., browser_get_content returns {content: string, interactive_map?: object, spilled: boolean, chars: number}. Publish these in tool definitions so LLMs can chain calls.
• Add recovery guidance to error messages. E.g., 'Navigation failed: URL returned 404. Try browser_search() to find a valid URL, or ask the user for clarification.'
• Consolidate ai_orchestrate_workflow and ai_plan_goal into the main server registration (src/index.ts) with explicit tool definitions, not test fixtures. Verify they are discoverable via the MCP tools list.