CLAUDE.md routing audit + MCP token-cost audit, 2026-05-01

Triggered by Nate Herk's "32 Tricks to Level Up Claude Code" hacks #15 (CLAUDE.md routes to other files) and #24 (API endpoints over MCP servers). Justin asked to apply both.

Hack 15: CLAUDE.md routing audit

forge/CLAUDE.md is 165 lines. The video's recommended cap is 150 to 200, so we are at the high end of healthy. The structure is already routing-shaped: Doctrine essentials inlined, everything else routes to FORGE-DOCTRINE.md, system-map/, MEMORY.md, memory/handoffs/.

Verdict: no immediate trim. Active structure is good. The risk is future-bloat from inlining new content here.

Going-forward rule, added to memory

When tempted to add a new section to forge/CLAUDE.md, add a one-line index entry that routes to a new system-map/ or memory/general/ file instead. Treat CLAUDE.md the way MEMORY.md is treated: it is an index, not a knowledge base, with the bare doctrine essentials inlined because nothing else auto-loads.

Brand-scoped CLAUDE.md files (currently only brands/nova-design/CLAUDE.md at 36 lines) follow the same pattern: thin briefing, route to the brand README and recent handoffs.

Sections that could move (if future trim is needed)

Section	Lines	Trim option	Trade-off
Doctrine Hard Rules	13 to 45	Move to FORGE-DOCTRINE-ESSENTIALS.md, link from CLAUDE.md	Loses inlined-on-every-session guarantee. The whole point of inlining was Section 12 ("only file auto-loaded"). Don't trim.
Quick Commands table	61 to 84	Move to system-map/commands.md	Loses at-a-glance lookup for the user and for Claude. Keep.
Where to Look for More	148 to 159	Could go anywhere	Already routing, no win.

Conclusion: leave as is, apply the going-forward rule.

Hack 24: MCP server audit, candidates for direct-API replacement

Six MCP servers are active in this session. Tools are deferred (lazy-loaded via ToolSearch), which softens the cost, but each server still incurs session-startup overhead and pollutes the ToolSearch index.

MCP	Tool count	Has direct-API alternative on forge?	Recommendation
claude-in-chrome	~30	No, this is the canonical browser surface	Keep
plugin:context7	2	No, current-docs lookup is novel	Keep
plugin:github	~40	Yes, gh CLI is already installed and on PATH	Drop, lean on gh
claude_ai_Gmail	OAuth provider	Mostly: reference_email_workflows.md covers reads + drafts via direct Gmail API; email-triage skill currently uses the MCP though	Investigate before dropping (email-triage flow needs to be ported to direct API first)
claude_ai_Google_Calendar	OAuth provider	Yes, forge_google_calendar.py is canonical (per reference_google_calendar_client.md, n8n calendar already retired 2026-04-30)	Drop
claude_ai_Google_Drive	OAuth provider	Yes, forge_gdrive_*.py (Phase 4.6 gdrive subsystem)	Drop

Concrete next actions, not yet executed

1. Drop GitHub MCP. Lowest risk, biggest token win (~40 tools). Action: remove the github plugin from the active plugin set. Replace any current usage with gh CLI calls. Update reference_claude_code_plugins.md.
2. Drop Google Calendar MCP. Action: same pattern, remove the OAuth connection. The forge_google_calendar.py direct client is fully wired and is what every cron / bot already calls.
3. Drop Google Drive MCP. Same pattern; the gdrive subsystem replaces it.
4. Defer Gmail MCP decision. Open question: does the email-triage skill rely on the MCP, or on direct API? If MCP, schedule a port to the same direct-API pattern as Calendar / Drive, then drop.

These are not destructive operations (just unwiring plugins / OAuth). Reversible. Each save is on the order of a few thousand tokens at session start plus a cleaner ToolSearch index.

What changed today

• forge/CLAUDE.md: no edits needed; going-forward rule recorded in this handoff and in the new feedback memory feedback_claude_md_routing.md.
• No MCPs removed yet. Justin to decide which to pull. Recommendation: start with GitHub (zero loss, gh already on the box).

Validated against doctrine

• Per feedback_subscription_over_api_for_automation: external-service APIs (GitHub, Google) bill on their own free tiers, not Anthropic API. Cutover saves Claude tokens with zero new external cost.
• Per feedback_robust_over_quick: replacing MCPs with direct clients is the more stable long-term path. Notion and Calendar already proved this.