Task Manager, for your AI.

Yes. It is MIT licensed and the whole source is on GitHub, the team half included. There is no account, no upsell and no paid tier. Use it as it is, or fork it and change whatever you like.

Download the installer for your platform from the GitHub release: a universal .dmg for macOS, Intel and Apple Silicon alike, or a setup .exe for Windows 10 and 11. Neither is code-signed yet. On macOS, open the app once and let macOS block it, then open System Settings, choose Privacy & Security, scroll to Security and click Open Anyway; after that it opens normally. On macOS 14 or earlier, right-click the app and choose Open instead. On Windows, SmartScreen will warn: choose More info, then Run anyway. Prefer to read what you run? Clone the repository and build it in about ten minutes; the README has the steps for both platforms.

No. Your limits, spend, folder names and client names stay on the machine. Each tool's token goes to that tool's own vendor API and nowhere else. There are two exceptions, both off by default: the MCP Trust Index lookup fetches a public list once a day, and update checks ask GitHub for the latest version number. Neither sends anything about you. The privacy page lists every network call and the commands that prove it.

Almost certainly you are looking at the Extra usage row, which is Anthropic's monthly pay-as-you-go cap, not the weekly limit. It does not move with the week; it moves with the billing cycle. The row says “monthly cap” and the real percentage for exactly this reason. If a refresh you asked for could not reach a provider, the footer says so and names the time the shown numbers are from. And the app never asks a vendor twice inside a minute, so clicking Refresh right after its own poll no longer trips a rate limit and freezes the card.

They are a floor, not an invoice. Spend is priced from your tools' local logs at API rates, and on Claude that reads about 23% low because 1-hour cache writes are billed at twice the rate the catalogue assumes. The app checks its own catalogue against the vendor's recorded cost and says when it is off, rather than quietly reporting a number it cannot stand behind.

Yes, that is the Inventory tab: every MCP server across Claude Code, Claude Desktop, VS Code, Cursor, Windsurf, Gemini CLI and Codex, pinned or not, which copies are live right now and how much memory they hold, and a way to end one. Names and counts only, never a key.

Yes. Running now lists every agent process with its folder, client and token pace; Inventory shows each agent definition with its runs and cost over 30 days; the audit checks agent guardrails; and a team policy can require them. That folder now resolves on Windows as well as macOS. Ending an agent is deliberately not a button, because killing one mid-task destroys work.

No, and it will not. Group Spend by Session and every session of the last 30 days is listed by cost with its age and the size of its log; Reveal shows you the file so you can archive it yourself. The app never deletes a session, because those logs are also where its spend figures come from. It changes your machine in exactly two places, both behind a confirmed click: ending a running server, and pinning a package version.

23 providers, including Claude, Codex, GitHub Copilot, Cursor, OpenRouter, Grok, DeepSeek, Kimi, Qwen, Z.ai and Ollama. Honestly: only Claude and Copilot are verified end to end on macOS so far; the rest came across from the Windows original and are being checked. Inventory › Sign-ins shows exactly where each one looked on your machine.

macOS and Windows 10 and 11, from one codebase. It builds on either kind of Mac; the machine I test on is an Intel Mac, so that is the one with the most hours on it.

English, 中文, Русский, Español, Français, Deutsch, 日本語, Português (Brasil) and 한국어: nine in all. It follows the system language by default, and can be set explicitly in Settings. Every view translates, findings, audit checks and notifications included.

Switch on Check for updates under Settings, Network; it is off by default. Turned on, the app asks the GitHub release feed for the latest version when it launches, when you open it and every four hours, sending nothing about you, and a button in the footer installs a new one only when you click it. Or download the next release yourself, or pull and rebuild.

Yes, and that half is free too. A headless agent reports each seat's inventory to a collector you host yourself: one process, one folder of JSON files, a token on every request. A policy file on the collector turns its dashboard into a conformance view: allowed and blocked MCP packages, pinned versions required, remote servers refused, a minimum number of deny rules, which AI tools may be present, agents must list their tools, the shell must be denied, and named hooks must exist. The verdict is computed from what the seat reported, never claimed by the seat. The report carries counts and package names, never paths, prompts, clients or credentials, and a test plants each of those to prove they cannot appear.

The interface is the real one, compiled for the browser. The numbers come from the real engine run against a made-up computer: an invented freelancer, invented clients. Nothing you click in it goes anywhere.