Your Claude Code.
In full view.

Megatron runs alongside your existing Claude Code workflow. It gives you visibility into skills, plugins, and local session usage. Claude Code remains where you do the work.

Skill inspection does not edit definitions or execute skill scripts. Resident tax runs Claude Code’s local /context command, which can start configured SessionStart hooks and MCP servers. Plugin management uses Claude Code’s own CLI commands, which may access the network. Inventory and usage analysis run locally.

Megatron keeps a local usage archive across scans and cache rebuilds, without storing prompt text in that archive. It retains activity already observed when Claude Code removes older source files. History deleted before Megatron saw it cannot be recovered. Activity and History use the same retained prompt records for matching time intervals. The History panel shows annual summaries, a yearly calendar, and a monthly project timeline. Coverage notes separate missing history from days with no activity; streaks count consecutive observed days with prompts. A day can still show token details when its prompt history is unavailable. Incomplete source scans preserve the previously archived history. Recorded paths and existing location links stay intact; similar names do not automatically merge projects.

Total tokens processed includes input, output, cache reads, and cache writes. Thinking is included in output. Figures for a time period are estimated allocations of recorded session totals; missing and undated token data are identified separately. Compact values show their exact counts on hover or keyboard focus; those counts retain the same estimation limits.

Resident tax runs Claude Code’s /context from an empty folder to measure your global baseline, excluding project instructions, skills, and MCP servers. It measures on first open each launch and again when you choose Refresh. Deferred tools and the autocompact buffer are shown separately. On context windows larger than 200K tokens, the bar is drawn against a 200K reference so small categories stay visible; totals and percentages still use the full window. Category tables come from /context; other details read the first request of your newest interactive Claude Code 2.1.300+ session on demand, without storing it. Those session details can include project configuration. The website preview uses synthetic examples throughout.

Megatron uses cost information recorded by Claude Code. Time and skill breakdowns are approximate allocations. They represent estimated API-equivalent usage, not subscription charges or invoices. Resumed, rewound, and forked conversations are counted once; when Claude Code has cleaned up part of a branched conversation, its cost is marked as an estimate. Missing data is shown as a coverage gap.

The desktop app runs on Apple silicon Macs and 64-bit Windows. Install it with Homebrew or the .dmg on a Mac, or with the installer on Windows. It is specifically built for Claude Code. Skills from other tools and Linux distribution are outside the current scope.

Megatron checks GitHub for a newer release 10 seconds after launch and every 6 hours. Each check is one request to github.com, which sees your IP address and platform; no skills, usage data, or paths are sent. Turn checks off in Settings → Updates; Check now works either way. Windows downloads updates in the background and installs them when you choose Restart to update or quit. macOS can’t install its own updates without a Developer ID signature, so upgrade with brew upgrade --cask megatron or a new DMG. The website preview never checks for updates.

Megatron is currently free and open source under the MIT license. You can inspect the code, run it locally, and contribute on GitHub.

Yes. Megatron watches global skills, plugin stores, user configuration, and approved project folders. It updates after file changes settle and periodically checks for missed changes. Live inventory updates leave Usage views and historical activity alone. If a discovered project is deleted, its inventory entries are removed and a selected project returns to the nearest available parent. Missing folders granted directly and unreadable folders retain cached data. Warnings stay with the affected folder, and unreadable sources retry automatically. Settings keeps the result of your manual rescan through later background updates. The website demonstrates these behaviors with sample files only.

Project folders require an explicit grant. The first-run welcome suggests folders Claude Code has run in; later, open Settings → Project folders → Manage…, add your repository or parent folder containing multiple projects, and rescan if needed. Grants never overlap: adding a parent folds in repositories granted inside it, a repository already inside a granted parent is skipped, and you can narrow a parent to just the folders you choose. A newly granted folder stays quiet until discovery confirms its availability; a known failure remains visible during retry. Usage keeps historical project identities independently of current folder access. Linking Usage locations does not grant access to folders. Automatic discovery skips temporary and build directories; explicitly granting a repository inside one still includes it.