Zeno Agent exposes 4 CLI tools via MCP with minimal documentation. All tools have descriptions (10-50 chars), but lack depth and actionable guidance. Input schemas are present but sparse: parameters have types and brief descriptions, but lack enums, constraints, or format specifications. No output schemas are documented. Error handling is absent, no recovery guidance, no categorization of failures. Tool names follow verb_noun convention (connector_refresh_tools, connector_tool_list, connector_tool_set, connector_tool_bulk), which is good, but descriptions are generic and do not explain WHEN to use each tool or what distinguishes them. Parameters like 'permission' and 'category' accept constrained values but are not declared as enums, forcing LLMs to guess valid options. No security annotations, no idempotency hints, no audit trail guidance.

• Add output schemas to all tools. Document the structure of responses, including field names, types, and what data is returned on success. Example: connector_tool_list should document that it returns an array of {tool_name, permission, category} objects.
• Convert constrained parameters to enums. For 'permission', declare enum: ['always_allow', 'ask', 'never']. For 'category', declare enum: ['read', 'write', 'interactive']. This makes valid options machine-readable and prevents hallucinated values.
• Expand descriptions to 50-200 characters with actionable context. Example: 'connector_tool_set: Set permission for a single tool in a connector. Use this to grant or deny access to a specific tool. Permissions: always_allow (auto-execute), ask (confirm before use), never (block).'
• Add tool annotations. Mark connector_tool_bulk and connector_tool_set with destructiveHint=true (they modify state). Mark connector_tool_list with readOnlyHint=true. This guides LLM retry logic and safety reasoning.
• Document error scenarios and recovery paths. Example: 'If connector not found, call connector_list() to see available connectors. If permission value is invalid, use one of: always_allow, ask, never.'