Built for Windows 10 & 11

Your agent has been waiting 11 minutes. You had no idea.

Vigil is a Windows-native always-on-top overlay that shows every Claude Code, Codex, Grok, Antigravity and Kimi session in real time — and lets you approve, answer and interrupt them without leaving what you're doing.

Interactive demo

This is Vigil, running.

The panel below is a live recreation of the Vigil overlay. When the agent asks for permission, answer it yourself.

A blank Windows desktop with the real taskbar, used as the backdrop for the interactive Vigil demo.

api-refactorWORKINGCLAUDEEdit src/auth/middleware.ts3m

Features

The demo shows the loop. Below is the rest of the cockpit — usage on the notch, policies per agent, history, and more.

See everything

Every session, every state, at a glance.

Real-time state for each agent: working, waiting, done, error, disconnected
Last 20 messages of any conversation, without opening the terminal
Sub-agents, grouped under the session that spawned them
Token statistics and usage windows — know how much quota you have left

Act without switching windows

Answer the agent from where you already are.

Approve or deny a permission request in one click
Answer an agent's question by typing directly in the overlay
Interrupt a running agent when it's going the wrong way
Jump straight to the right terminal when you do want to dive in

Stay out of the way

An overlay you forget about until it matters.

Fades back when agents are working, steps forward when they need you
Light enough to leave running all day
Drag it anywhere, on any monitor
Starts with Windows, or doesn't — your call

Usage on the notch

See remaining quota and usage windows on the notch itself — and pick which agent’s window you care about. Not buried in a secondary panel.

Permission policy, your way

Set approve, deny, always-approve, or manual defaults per tool — and override them per agent. Full control when one session is noisier than the rest.

Permissionsapi-refactor

Edit / Writeglobal

Readglobal

Web fetch / searchglobal

Otherglobal

Conversation without the terminal

Open the last messages of any session in the overlay. Read context, then jump to the terminal only when you need to type code.

api-refactorWORKING

JWT migration · last 20 messages

CLAUDETERMINAL×

14:02:01

Refactor auth middleware to support JWT.

14:02:04assistant

I'll inspect the current middleware and plan the change.

14:02:05tool

Read src/auth/middleware.ts

14:02:12assistant

Found cookie-only sessions. Proposing a JWT path next to cookies.

14:02:13tool

Grep verifyJwt / session

14:03:40

Keep cookies during migration if needed.

14:03:44assistant

Understood — dual mode until you cut over.

14:03:50tool

Edit src/auth/middleware.ts

14:04:02assistant

Edit ready: authorization header + verifyJwt.

14:04:18

Which strategy should we default to?

14:04:20assistant

Options: Session cookies, JWT, or keep both.

14:05:01

JWT for new clients; keep both for now.

14:05:06assistant

Defaulting to JWT; cookies remain accepted.

14:05:08tool

Bash npm run typecheck

14:05:41assistant

Typecheck clean. Adding a short note in README next.

14:05:55tool

Edit README.md

14:06:10assistant

Documented the dual-mode auth header.

14:06:22

Ship it when tests are green.

14:06:30tool

Bash npm test -- auth

14:07:02assistant

3 tests passed. Ready to continue.

in: 563k·out: 135k·cache: 21.7m·~$20.01

NotifsPermissions

Sub-agents in the notch

Child sessions stay grouped under the agent that spawned them — visible in the compact bar, not lost in a separate list.

api-refactorWORKINGCLAUDEEdit src/auth/middleware.ts3m+2

Notch display modes

Priority agent with overflow, multi-agent side by side, or a detailed single-agent row. Match the bar to how you run your fleet.

api-refactorWORKINGCLAUDEEdit src/auth/middleware.ts3m+2

Names and tags

Rename sessions and attach your own tags so the right agent is findable when you have more than a handful running.

History and resume

Browse past sessions and resume one when you need it — Vigil lands you back in the right terminal context.

Interrupt a running agent

Claude Code

Stop a session that is going the wrong way from the overlay — without hunting for the right window first.

And more…

The rest of the cockpit, without a full card each:

Customize the display — show or hide fields, pick colours, swap mascots
Sound cues when an agent needs you
Per-tool enable/disable — only hook what you use

Usage window priming

Claude Code

When your usage window is closed, Vigil can start it early so your next reset comes sooner. Opt-in, off by default, and it costs a few tokens each time.

A typical day · 5h Claude windowsstarts on your promptstarts automatically

Without priming3 windowsGaps waste reset time · conserve tokens each block

9:15 → 14:15

14:45 → 19:45

20:30 → 1:30

8:00 PC on

9:15 first prompt

14:45

20:30

With priming4 windowsResets land earlier · freer spend from first prompt

8:00 → 13:00

13:00 → 18:00

18:00 → 23:00

23:00 → 4:00

8:00 PC onautomatically

9:15 first prompt

13:00 new windowautomatically

18:00 new windowautomatically

23:00 new windowautomatically

Priming uses a few tokens at boot so the first window opens at 8:00 — not at your first prompt. Opt-in, off by default.

Know what's happening, everywhere

Get pinged on your phone when an agent needs you — on your terms.

Telegram notifications when an agent finishes, gets stuck, or fails
Approve or deny from your phone
Choose what notifies you globally, and override rules per agent — finish only, every wait, failures, or silence one session entirely
Your own bot, your own token — no Vigil server in the middle

Vigilbot

Today

✓ Task finished

api-refactor · Claude Code
Auth middleware JWT migration is done. Typecheck clean, 3 tests passed.

14:07

write-tests is still working — generating coverage for the new JWT path.

14:08

Permission request

api-refactor · Claude Code
Edit src/auth/session.ts

14:09

Works with your agents

Vigil talks to each CLI through its own hook system. What it can do depends on what each tool exposes — here's exactly what you get.

What each supported CLI lets Vigil do
Capability	Claude Code	Codex CLI	Grok Build	Antigravity CLI	Kimi CLI
Monitoring	Supported	Supported	Supported	Supported	Supported
Approve / Deny	Supported	Supported	Limited, see note 1	Not available	Limited, see note 2
Answer questions	Supported	Not available	Limited, see note 3	Not available	Not available
Interrupt agent	Supported	Not available	Not available	Not available	Not available
Sub-agents	Supported	Supported	Supported	Not available	Supported
API errors	Supported	Not available	Supported	Not available	Not available
History	Supported	Supported	Supported	Supported	Supported
Statistics	Supported	Supported	Supported	Not available	Supported
Usage windows	Supported	Supported	Supported	Supported	Not available
Usage window priming	Supported, see note 4	Not available	Not available	Not available	Not available
Mobile notifications, see note 5	Supported	Supported	Supported	Supported	Supported

Grok — Approve/Deny: full Approve/Deny from Vigil in always-approve mode. In default mode the request is shown read-only — confirm in the terminal.
Kimi — Approve/Deny: full Approve/Deny from Vigil in yolo mode. In manual or auto mode the request is shown read-only — confirm in the terminal.
Grok — Answer questions: shown read-only; type your answer in the terminal.
Claude Code — Usage window priming: opt-in — when your usage window is closed, Vigil starts it early so your resets come sooner (uses a few tokens each time). Enable it in Settings.
Mobile notifications: these work for every tool Vigil monitors. Which events reach you follows the rows above — a tool that exposes no API-error hook can't tell you it failed — and approving from your phone follows the Approve / Deny row.
Why the gaps? Each CLI exposes a different set of hooks. Where a cell is empty, the tool simply doesn't give Vigil anything to work with — we'd rather show you than let you find out after you install.

More on the way. New tools ship as free updates.

Up and running in a minute

Install
Download the installer and run it. It installs for your user only — no admin rights required. Windows will warn you the first time: Vigil isn't EV-signed yet. Click "More info", then "Run anyway".
Review, then allow
Vigil shows you exactly what it will add to your tools' config files. You approve, or you don't. Nothing happens behind your back.
Start an agent
Launch Claude Code as usual. It shows up in the overlay instantly.

Changed your mind? Uninstalling from Settings restores every config file to exactly how it was.

There is no Vigil server

Your prompts, your code and your conversations never pass through us — because there's no "us" to pass through. Here's every connection Vigil makes, and why.

No backend, no account, no telemetry. Vigil has no servers. You never create an account. We collect nothing about you — there's nowhere for it to go.

Your conversations stay on disk. Vigil reads your agents' local transcripts to show them to you. It never uploads them, anywhere.

Agents talk to Vigil over a local Windows named pipe. That's the entire architecture.

Every outbound connection

Goes to	What for	When
`api.anthropic.com`	Reads your Claude usage limits	Only if usage tracking is on
`chatgpt.com`	Reads your Codex usage limits	Only if usage tracking is on
`cli-chat-proxy.grok.com`	Reads your Grok usage limits	Only if usage tracking is on
`api.telegram.org`	Sends notifications to your bot	Only if you set up mobile notifications
`api.github.com`	Checks for a new Vigil version	On startup

Every one of these goes either to a service you already use, with credentials already on your machine, or to Vigil's own housekeeping — checking for updates. None of them go to us.

Questions

Is Vigil free?

Yes. Download it and use it. Vigil is free — nothing to activate, and nothing that expires.

Do I need an account or a licence?

No. There is no account and no licence. Install Vigil and it runs.

Is the source code public?

Yes. Vigil is open source (MIT). The code is on GitHub.

Does Vigil work without an internet connection?

Yes. Vigil runs locally, so it keeps working offline. Nothing phones home to unlock it, so a bad Wi-Fi connection will never lock you out. It reaches the network only for the things you turn on — usage limits, your own Telegram bot — and to check for a new version.

Does Vigil send my code or prompts anywhere?

Your code and prompts, never. Vigil does talk to a few services — your agents' own providers, for usage limits, and your own Telegram bot if you set one up. Here's the full list.

What exactly does Vigil change in my config files?

It adds hook entries to the config files of the tools you enable — nothing else. Vigil shows you the exact change before it makes it, and uninstalling restores every file to how it was.

How do I reinstall, or move to a new machine?

Install it again. Nothing to activate and nothing to turn off. Your settings stay on that machine's disk; on a new machine you start fresh, or copy them yourself. Use Vigil on as many machines as you like.

Which Windows versions are supported?

Windows 10 (version 2004 and later) and Windows 11, 64-bit. Vigil is Windows-native — there is no macOS or Linux build. Mac agent islands don't run on Windows; this product is built for Windows developers.

Why doesn't feature X work with tool Y?

Because that CLI doesn't expose the hook Vigil would need. The compatibility table above shows exactly what each tool supports.

Windows says it doesn't recognise the installer. Is that normal?

Yes. Vigil isn't signed with an EV certificate yet, so Windows SmartScreen warns you the first time you run the installer. Click "More info", then "Run anyway". Everything else about the install is ordinary: it goes into your user profile, it needs no admin rights, and uninstalling removes it cleanly.

How do I get support?

Email stevenpf@gmail.com, or open an issue on GitHub. Bug reports get a reply.

Stop waiting on agents that are waiting on you.

Free. No account, no licence.

FreeDownload Vigil

Open source (MIT)

Windows 10 & 11 · No Vigil server

Your agent has been waiting 11 minutes. You had no idea.Your ClaudeCodexGrokAntigravityKimi agent has been waiting 11 minutes. You had no idea.