Task Manager, for your AI.
Remember opening Task Manager to see what was really running? This sits in your menu bar and does that for your AI tools: limits, spend, subscriptions and every MCP server on the machine.
Private beta. Builds are not code-signed yet, so there is no public download. The demo below is the real interface running on made-up data.
- Price
- Free
- Runs on
- Mac + Winmenu bar · system tray
- License
- MITbuilt on open source
- Data collected
- Noneno telemetry, no account
- Size
- 26 MBinstalled, macOS
- Languages
- 3English · 中文 · Русский
One small window. A lot behind it.
It opens at 380 pixels wide and stays out of the way. Click anything and it goes a level deeper, in the same window.
Every limit, one glanceSession, weekly and model limits across your tools, with a warning on the one that runs out before it resets. 
When does it run out?A forecast from your recent pace, in a sentence. It says which pace it used, and goes flat when you have been idle. 
Your week, by the hourWhich hours actually eat the limit, with the next reset outlined. Hours nobody measured are left blank, not guessed. 
Spend by clientMap folders to clients once. Every dollar lands in exactly one row, with a budget per client and a CSV for the invoice. 
What is installed, and what to fixMCP servers, agents and skills across seven apps. Opportunities are computed from your machine, never generic advice. 
One-click pin, with a previewSee the exact line that changes before it does. One string is replaced, a backup is saved, nothing else moves. 
Is the plan worth it?Your own prices against what the same work would cost pay-as-you-go. It never guesses what you pay. 
A scored audit of your setupChecks passed over checks that apply. Only real gaps cost points, and it exports as Markdown. 
Somebody lives in AboutPoke him. Then poke him nine more times.
Every screenshot is the real app showing a fictional machine. Acme Co and Northwind are not clients of anyone.
Don’t take the screenshots’ word for it. Use it.
This is the actual interface, compiled for the browser, running on numbers the real engine produced from a made-up computer. Nothing you click here goes anywhere.
- Click Claude’s name to open its detail page: the chart, the forecast, your week.
- In Spend, change Group by to Client, then click a bar to see the sessions behind it.
- Hit Wide in the top corner. Same window, twice the room, no pop-out.
- Open Inventory and press Pin… on a server to preview the change.
- Open Subscriptions and find the plan that is not earning its keep.
- Run the Audit from Inventory. Then find the robot.
Why I built this.
I run a lot of AI at once: several subscriptions, a pile of MCP servers, agents working across client folders. No single place could tell me what all of it was doing, what it cost, or who it was for.
The old answer to “what is this computer doing?” was Ctrl+Shift+Esc. There was no equivalent for AI tools, so this is it. It reads what is already on your machine, the logs and config files your tools leave behind, and turns it into answers: what is running, what it costs, who it was for, and what is worth tightening.
It also teaches as it goes. Every finding says why it matters and links to the explainer, because the fastest way to learn how this stuff works is to see it on your own setup.
Everything it’s running.
The full feature list, laid out the way it should be. Watch the last column.
| Name | What it does | Network |
|---|---|---|
| See (5) | ||
| Live limits | Session, weekly and model limits for each tool, with time to reset and pace. | vendor API only |
| Menu bar number | The limit closest to the wall, as plain text next to the icon. | 0 B |
| Spend | Today, yesterday and 30 days, priced from your local logs by model. | 0 B |
| Inventory | MCP servers, agents, skills and guardrails across Claude Code, Claude Desktop, VS Code, Cursor, Windsurf, Gemini CLI and Codex. Names and counts only, never a key. | 0 B |
| Subscriptions | A ledger of what you pay, when it renews and a reminder before it does. | 0 B |
| Understand (6) | ||
| Forecast | When a limit runs out at your recent pace, and only when that comes before the reset. | 0 B |
| Your week | A 7 × 24 heatmap of when a limit really gets used. | 0 B |
| Clients & work areas | Spend by folder, rolled up to the client it was for, with budgets and CSV export. | 0 B |
| Session drill-down | The sessions behind any bar: times, totals and models. Never a prompt or a title. | 0 B |
| Plan vs pay-as-you-go | Whether each subscription beats what the same work would cost at API rates. | 0 B |
| Usage coaching | Top-heavy model mix, sessions left open for weeks, spend nobody sorted. | 0 B |
| Act (6) | ||
| Setup audit | A scored read of the whole setup, exportable as Markdown. | 0 B |
| One-click pin | Pins an unpinned MCP package to the version already in your local cache, after a preview. | 0 B |
| Budget guard | Alerts for a limit burning fast, a daily spend mark and a client going over budget. | 0 B |
| Weekly digest | One notification a week with what was spent and what renews next. | 0 B |
| Trust ratings | Matches your MCP servers against the MCP Trust Index. Off until you turn it on. | opt-in · 1 GET a day |
| Team mode | A headless agent and a self-hosted collector for a team view with policy checks. Reports carry counts and package names, never paths, prompts or clients. | your server only |
App privacy.
The label an app store would make us fill in, answered properly.
Data collected
None. There is no server to send it to.
No analytics, no crash reporting, no account, no update pings. The tracking module that came with the upstream project was removed, not switched off.
- Reads credentials your tools already stored, only to ask each vendor for your own limits. It never writes to the Keychain or Credential Manager.
- Reads metadata from local logs: models, token counts, folders, times. Never prompts, replies or conversation titles.
- Writes to another tool’s file in exactly one place: the pin you clicked Apply on.
The one exception
Trust ratings, if you switch them on.
Once a day the app downloads the public Trust Index list from staas.fund and matches it on your computer. The request has no parameters, so it says nothing about which servers you run. It is off by default.
- Everything else on the network is a tool’s own vendor API, the same calls the tool itself makes.
- The local API for your own scripts listens on this computer only, and the feeds with folder or client names stay off until you enable them.
- It is MIT licensed and built on two open-source projects you can read today, linked below.
Information.
- Made by
- Peter Saddington, StaaS Fund
- Version
- 0.1.0Private beta, unsigned builds
- Category
- Developer tools
- Compatibility
- macOS and Windows 10 / 11Built with Tauri: a Rust core and the system webview
- Size
- 26 MB installedAbout a 10 MB download on macOS
- Languages
- English, 中文, Русский
- Works with
- Claude, Codex, GitHub Copilot, Cursor, OpenRouter, ElevenLabs, Grok, DeepSeek, Kimi, Qwen, Z.ai, Ollama and moreChecked end to end on macOS so far: Claude and Copilot. The rest came across from Windows and are being verified.
- Price
- FreeNo in-app purchases. No upsell. MIT license.
- Standing on
- Pane by Jazii, the Windows tray app this grew from, and OpenUsage by Robin Ebers, the macOS original.Both MIT. Their provider research made this possible.
- Learn more
- The Library · MCP Trust Index · The Open Classroom
Want it on your machine?
Signed installers are the last thing left before a public release. Until then I am sending builds by hand to people who will tell me what breaks.