Files
frmorettoandClaude Opus 4.6 a7197e5534 Edition 1.4: 37 tools, 4 platforms, 13 corrections, 2 contributors
9 new tool cards (29–37): iOS Reminders CRUD suite (5 tools),
visualize:show_widget, gmail_create_draft, anthropic_api_in_artifacts,
persistent_storage. 13 v1.3 corrections including user_time/location
reclassified as always-loaded, memory_user_edits 500-char limit,
chart_display_v0 intermittent (not deterministic), window.storage
session-scoped. New sections: MCP connector state, artifact execution
layer, Project context behavior. iOS split from Android as distinct
platform. Contributors: anonymous security researcher, DMontgomery40.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-17 19:29:02 +01:00

3.5 KiB

AGENTS.md — Context for AI Coding Agents

What This Repo Is

A technical book ("Claude's Hidden Toolkit") documenting 37 undocumented internal tools in Claude.ai's consumer interfaces. Fully reverse-engineered through systematic testing across browser, desktop app, Android, and iOS.

File Roles

File Role Editable?
Claude_Hidden_Toolkit.md Source of truth Yes — all content changes here
README.md GitHub landing page Yes
LICENSE CC BY 4.0 No

Domain Knowledge

  • 37 tools documented with confirmed JSON schemas
  • 4 platforms with different tool inventories:
    • Browser: 21 always-loaded, tool_search conditionally available with MCP connectors
    • Desktop: 22 always-loaded + 32 MCP tools via tool_search
    • Android: 22 always-loaded + 11 deferred consumer tools
    • iOS: 22 always-loaded + 16 deferred consumer tools (Android set + 5 Reminders)
  • MCP connector state affects tool inventory — connecting Gmail/Calendar adds tools to always-loaded pool
  • Tool cards in the Technical Annex follow a strict template (see Appendix C in the book)
  • Figures are numbered sequentially — all currently marked [Upcoming] pending screenshot capture
  • Response formats are all empirically confirmed — no inferred markers remain
  • Two external contributors credited: an anonymous security researcher (11 findings) and DMontgomery40 (iOS Reminders suite)

Conventions

  • Tool names in backticks: web_search, event_create_v0
  • Platform tags on every claim: "Mobile App only", "Browser + Desktop only", "iOS only", "All"
  • Version suffixes: _v0 (initial/experimental), _v1 (breaking upgrade)
  • MCP connector tools use Provider:tool_name prefix on browser (e.g., "Google Calendar:gcal_list_calendars")
  • Confirmed vs unverified clearly marked — never speculate
  • Anonymous contributor referenced only as "An independent security researcher who prefers to remain anonymous" — never use real name

Architecture Context

The book documents a multi-tier tool system:

  • Always-loaded tools: injected into every conversation context (~21-22 depending on platform)
  • Deferred tools: discovered at runtime via tool_search (mobile only, 11-16 consumer tools depending on OS)
  • MCP tools: discovered via tool_search on desktop (32 integration tools) or conditionally on browser
  • MCP connector tools: appear in always-loaded pool when connectors (Gmail, Calendar) are active
  • Artifact execution layer: anthropic_api_in_artifacts (Claude-in-Claude), persistent_storage (session-scoped), MCP endpoint access from React artifacts — desktop/browser only, not mobile

The calendar system has a dual-layer architecture where Projects use MCP Connector tools with email-based IDs, while standard mobile chats use deferred consumer tools with numeric string IDs. The two layers are incompatible.

visualize:show_widget and chart_display_v0 are mutually exclusive by surface: chart_display_v0 is mobile only (deferred, intermittent instability), show_widget is desktop/browser only.

Quality Gates

Before any content change:

  1. Has the claim been tested on the relevant platform(s)?
  2. Is the platform availability clearly stated?
  3. Does the tool card follow the template structure?
  4. Are figure numbers sequential with no gaps?
  5. Is the response format marked as confirmed or unverified?
  6. For contributor content: has it been independently verified before documenting as confirmed?