Memcode has 36 tools with significant structural gaps. Most tools lack proper input schemas (many show only parameter names without type definitions in visible code). Descriptions are present but often generic (10-50 chars). No evidence of output schemas documented. Tools are not verb-prefixed consistently (gw_* prefix dominates, obscuring action intent). Parameter descriptions exist but lack constraints, ranges, or format specifications. Error handling is minimal, no recovery guidance visible. Security concerns: admin tools accept free-form strings for sensitive operations (channel names, agent names, policy documents) without validation hints. Composition issues: gw_agent combines 9+ responsibilities (model, reasoning, browser, objective, toolsets, autonomous, paused) in one tool. No pagination support visible despite tools returning lists. Overall, this reads as an internal admin interface rather than a production-grade agent toolkit.

• Split gw_agent into 9 focused tools: create_agent, set_agent_model, set_agent_reasoning, enable_agent_browser, set_agent_objective, configure_agent_toolsets, set_agent_autonomous, pause_agent, resume_agent. Each tool does one thing.
• Document output schemas for all 36 tools. Example: gw_overview should return {channels: [{name, enabled, agent, projects}], agents: [{name, model, reasoning}], ...}. LLMs need this to chain calls.
• Add enums to all free-form string parameters. gw_channel.field → enum [allow_add, allow_remove, agent, pairing, respond_to_all, voice_replies, poll, projects]. gw_pairing.action → enum [approve, deny, revoke]. Prevents hallucinated values.
• Rename tools to verb_noun convention: gw_overview → get_gateway_overview, gw_channel → configure_channel, gw_agent → create_agent (after split), gw_schedule → create_schedule, gw_service → control_gateway_service. Verb-first names let LLMs infer intent.
• Add error handling guidance to descriptions. Example: 'If agent not found, call list_agents() to see available agents.' Classify errors as retryable (network timeout) vs user-fixable (invalid channel name) vs fatal (permission denied).