The network-mcp server provides 42 tools across multiple network infrastructure domains (Arista, Batfish, Loki, PostgreSQL, Prometheus, TACACS). While tool NAMES are generally well-formed with action verbs (run_commands, get_bgp_summary, list_snapshots), the implementation exhibits critical gaps in schema depth, parameter documentation, and output structure specification. Examination of arista_tools.py, batfish_tools.py, and related modules reveals that most tools have descriptions present but lack comprehensive input/output schema documentation and parameter-level descriptions. The average tool description is adequate (60-120 chars observed), but parameter descriptions are inconsistent: some tools like run_commands document their parameters clearly (device_name, commands, encoding), while others (e.g., list_snapshots, get_latest_snapshot) have minimal or no parameter descriptions beyond the name. Output schemas are inferred from the source code (typically JSON-serialized results) but are NOT explicitly documented in tool registration. Error handling is present in code (try-catch blocks with JSON error returns) but lacks actionable recovery guidance for LLMs. No tool exhibits tool annotations (readOnlyHint, destructiveHint, idempotentHint), and there is no evidence of output schema declarations. This places the server in the C/D range: functional definitions with significant documentation gaps.

• CRITICAL: Add explicit output schemas to all tool definitions. For example, run_commands should declare: { "type": "object", "properties": { "result": { "type": "array", "description": "Array of command results" }, "error": { "type": "object" } } }. This enables LLMs to understand response structure without reading code.
• CRITICAL: Add tool annotations. Mark load_snapshot with destructiveHint=true. Mark all READ_ONLY tools with readOnlyHint=true. This helps agents reason about side effects and plan safely.
• HIGH: Rename duplicate 'query' tools: query_logql (for Loki) and query_tacacs_audit (for TACACS). Verb-noun naming eliminates ambiguity.
• HIGH: For empty-parameter tools (get_parse_status, get_node_properties, get_latest_snapshot, health), add descriptions explaining: (a) what data structure is returned, (b) when this tool should be called relative to others (e.g., 'Call after load_snapshot()'), (c) any prerequisites.
• MEDIUM: Document pagination and result limits. For each tool accepting 'limit', state: 'Maximum: 500. Default: 20.' For tools without explicit limits, add truncation warnings: 'get_running_config may return very large results; consider using run_commands with specific filtering.'
• MEDIUM: Add recovery guidance to error descriptions. Examples: 'If 'Unknown Arista device' error, call list_devices() to see available device names.' 'If reachability check fails, verify src_ip and dst_ip are within the simulated network topology using traceroute.'
• MEDIUM: Document mutual exclusivity. For tools like get_bgp_sessions (accepts both node and snapshot_id optionally), clarify: 'Both node and snapshot_id are optional; if omitted, defaults to latest snapshot with all nodes.'