Uncovering Claude’s 37 Hidden Tools: A Complete Reverse‑Engineered Guide

Introduction

The author introduces a community‑driven open‑source effort that reverse‑engineers Claude’s hidden internal tools, many of which have never been documented publicly.

Methodology

Over the course of one week the researcher systematically examined every platform where Claude runs – the web UI (claude.ai), the desktop client, Android, and iOS. For each discovered tool the author recorded the tool name, full JSON schema, performed functional tests, explored edge cases, and verified behavior on all four platforms.

Key Findings

Claude is not a single monolithic product; it comprises three separate toolsets that are distributed unevenly across platforms:

Browser (claude.ai) : 21 always‑loaded tools, no tool_search, no delayed‑load tools.

Desktop client : Same baseline as the web UI plus a tool_search that can discover 32 MCP‑integrated tools (Chrome + filesystem).

Android app : Same baseline plus 11 consumer‑grade delayed‑load tools (alarm, timer, calendar, chart, location, time) accessed via tool_search.

iOS app : Same baseline plus 16 delayed‑load tools (the 11 Android tools plus 5 reminder‑related tools) accessed via tool_search.

The web version, often perceived as the most feature‑complete, is actually the most limited in tool diversity; the mobile apps contain the richest built‑in architecture.

Tool Catalog (Selected Examples)

The 37+ tools are grouped into functional categories. Below are representative entries (the full list contains JSON schemas for each tool).

Context & Device Tools : user_time_v0 (gets current time, all platforms), user_location_v0 (gets location, all platforms), ask_user_input_v0 (requests user input, all platforms), message_compose_v1 (drafts email/message, all platforms).

Visualization : chart_display_v0 (renders interactive charts, browser + desktop), visualize:show_widget (shows UI widget, browser + desktop).

Calendar & Time Management : a suite of 6 calendar tools, alarm (sets alarm, mobile only), timer (count‑down timer, mobile only).

Search & Data : web_search, image_search, places_search, map_display, sports_data, google_drive (requires MCP connector).

Memory & History : memory_user_edits (user memory edit, 500‑char limit, all platforms), conversation_search (searches dialogue history), recent_chats (lists recent chats).

Automation & System : bash (executes bash commands, desktop only), filesystem (file system ops, desktop delayed‑load), artifacts (creates/manages artifacts, browser + desktop), skill_system (custom skill management, desktop).

Artifact Execution Layer : anthropic_api_in_artifacts allows artifacts to call Claude’s own API without an external API key, effectively turning artifacts into a lightweight application platform.

Meta Tool : tool_search (enumerates available tools, all platforms; browser requires MCP).

Additional Findings

The study also documents the MCP (Model Context Protocol) connector state‑change tools, which add or remove functionality when Gmail or Google Calendar is linked or unlinked, and outlines the persistent storage mechanism ( persistent_storage) that provides session‑wide caching.

Content Overview

The accompanying document is organized into 12 chapters covering architecture, dual‑layer loading, platform inventories, MCP connector details, version control, specific tool categories, a full reference index, and a forward‑looking discussion of “ghost” tools and future directions.

Target Audience

Advanced users – learn the exact prompts that trigger hidden capabilities.

Developers – understand the gap between the consumer product and the public API.

AI researchers – obtain a case study of production‑grade AI tool architecture and the impact of platform fragmentation on user experience.

Disclaimer & Risks

These hidden tools are experimental, region‑restricted, or beta‑only. Using them may violate platform terms of service and could lead to account or data risks; the author assumes no responsibility for any such outcomes.