The n8n-claw server exposes 13 tools across multiple bridges (email, file, browser, Discord). While all tools have names and most have descriptions, significant quality gaps emerge across schema definition, parameter documentation, and error handling. Tool naming generally follows verb conventions (read-emails, send-email, upload, run_task), but descriptions are inconsistent in depth and actionability. Parameter schemas are present but incomplete, many lack proper type constraints, enums for constrained values, and descriptions that would guide LLM selection. Output schemas are largely undocumented. Error handling is minimal: most tools return generic error messages without recovery guidance or actionability hints. The server mixes naming conventions (kebab-case: read-emails, send-email; snake_case: run_task, list_sessions; path-based: files/:id). Security risks are acute: email bridge accepts plaintext credentials as parameters (host, user, password, port) rather than using server-side injection. Browser bridge accepts arbitrary user_id and domain without validation. File bridge stores user-uploaded content without clear authorization checks. These patterns violate pattern:secret-injection and pattern:permission-gate fundamentally. Tool composition is reasonable, tools are narrowly scoped, but several parameters are underdocumented. For example, run_task's 'user_id' parameter lacks format guidance ("Qualified user ID (e.g. 'telegram:1810565648')") but no validation schema or error message specifies what formats are valid. The 'domain' parameter in run_task and close_session is unexplained: is it a domain regex, an exact match, or something else? File bridge lacks pagination despite storing files with TTL and presumably returning lists. Email read-emails tool returns up to a limit but no total count or next cursor, making repeated queries error-prone. Discord reply tool's automatic chunking at 2000 chars is undocumented in the schema, forcing LLMs to guess behavior.

• Document complete output schema for all 13 tools, showing field names, types, and required vs. optional fields. Use JSON Schema format matching input schemas.
• Migrate email bridge credentials to server-side injection: (1) remove host, user, password parameters; (2) add optional 'credential_id' parameter; (3) load credentials from environment or vault by ID; (4) implement credential rotation and expiration.
• Standardize tool naming to snake_case verb_noun: read_emails, send_email, upload_file, upload_file_base64, download_file, get_file_metadata, delete_file, cleanup_files, forward_file, run_task, list_sessions, close_session, reply_discord.
• Add formal constraint to user_id parameter in browser bridge (pattern: '^[a-z]+:[a-z0-9]+$' for 'telegram:123456'); validate against authenticated agent identity before executing.
• Implement pagination for read_emails and list_sessions: add 'offset' and 'limit' parameters (default limit=20, max=100); return 'total' count and optional 'next_cursor' in response.
• Add error guidance to all tool descriptions: specify which errors are retryable (e.g., 'timeout: try with smaller max_steps'), which are user-fixable (e.g., 'invalid email address'), and which are fatal (e.g., 'credential expired').
• Add domain whitelist to files/:id/forward; reject private IPs (127.0.0.1, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16); add DNS rebinding check; require https:// scheme.
• Implement permission checks for destructive operations: verify that agent calling delete_file or close_session is authorized (belongs to same user/team); throw 403 Forbidden if not.