ALL MEMOS Download .docx

Key Insight

Claude Code v2.1.149 (May 23 2026) upgraded /usage to show per-category token cost breakdown: skills, subagents, plugins, and per-MCP-server. This makes it possible to see exactly which component is burning quota.

Key cost ratios to know:

• Skills: ~100 tokens of context until invoked; 50 skills installed costs almost nothing unused.
• MCP servers: a 5-server setup with 58 tools eats ~55,000 tokens before the first prompt. Anthropic's Tool Search feature (load tool definitions on demand) cuts this by ~85%.
• Subagents: pairing a Haiku subagent with a scoped MCP server + preloaded skill is the lowest cost pattern for parallelizable work.

The /usage command also shows the /insights crash fix (20+ bug fixes in v2.1.149) and Markdown GFM task-list checkbox rendering.

Action Items

• Run /usage after the next TITAN /feed or /forge session to establish a baseline per-category cost profile.
• Audit TITAN's MCP server configuration: if any server has > 10 tools loaded eagerly, enable Tool Search to switch to on-demand loading and cut context cost by ~85%.
• When spawning ORACLE or DARWIN sub-tasks, prefer Haiku + scoped MCP + skill over a full Sonnet session with all tools.
• Use per-category /usage data to tell DARWIN which sub-agent patterns are cost-inefficient.

Relevance

• TITAN capabilities: /usage breakdown is now the primary diagnostic for TITAN session cost attribution. Previously invisible, now queryable in-session.
• Business context: TITAN runs 250+ sessions/day; per-MCP-server cost visibility enables targeted optimization without guesswork.