NOW ON THE MICROSOFT STORE

Agents for your PC.

Deploy desktop agents for files, terminal, browser and screen. They run on your machine with your own model keys, and they ask before they act.

Get from Microsoft Store

Runs on your machine · Your own model keys · Microsoft Store

  • Plans first, then asks
  • Runs while you're away
  • Skills you can read and edit
  • Every action logged
  • Ctrl+Alt+Esc stops screen control
  • No SHADOW cloud
  • Your keys, your models

THE PROBLEM

Your day is full of work that is not your work.

ROUTINE001

The same chores come back every week.

Reports, file sorting and research repeat, and each one costs time you meant to spend on real work.

SILOS002

Your tools do not talk to each other.

Moving results between apps, tabs and folders is work nobody counts, and it never ends.

TALK003

Chat assistants answer. They do not act.

You still copy the answer, open the app and do the steps yourself.

CONTROL004

Handing AI the keys feels unsafe.

Without clear limits and a log, you cannot trust what ran while you were away.

MEET SHADOW

Delegate real work. Keep the final say.

Create agents, each with its own chat, model, permissions and schedule. Give one a goal. It writes a plan, you approve it, it works, it reports. Click any agent to open its chat, and click the approval buttons yourself.

Interactive preview
RResearch working

Compare three Windows automation tools for our team. Pros, cons, price.

PLAN · 3 stepsResearch

  1. Search the web for current options
  2. Read pricing and docs pages
  3. Write a summary to Reports\automation-tools.md
edit · deny

Searched 4 queries · 12 pages read

Compared 3 tools · price, limits, setup

Saved Reports\automation-tools.md

The summary is in your Reports folder. Nothing was sent anywhere.

Sort my Downloads by project.

FILES · MOVEneeds your ok

Move 148 files from Downloads into 6 project folders.

always · deny

Moved 148 files into 6 folders

Skipped 3 open in another app

Logged every move, undo ready

Run this every Friday at 5 pm.

Created routine Friday sort

The api tests fail since this morning. Fix it.

TERMINAL · POWERSHELLexact command

> pytest -q tests/api

always · deny

src/api/dates.py+4 −1

- return parse(value)

+ return parse(value, tz=UTC)

0/14 tests passed

GITHUB · OPEN PULL REQUESTshadow/fix-dates

once, after your tap

Draft the release notes from this week's merged pull requests.

FILES · WRITEneeds your ok

Create Docs\launch\release-notes.md in your granted folder.

always · deny

Read 12 merged pull requests

Wrote 3 sections · features, fixes, notes

Saved Docs\launch\release-notes.md

The draft is ready. I did not post it anywhere. Send it to the Launch crew? That hand-off asks you first.

Launch is Thursday. What's left?

Messages from Research and Report Writer · each hand-off approved

Research

Competitor pricing table is done. 2 changes since last week.

Report Writer

Release notes are drafted in Docs\launch. Waiting for your read.

Chief

Pricing Research · done

Release notes Report Writer · needs your review

Checklist me · 2 of 9 open

your PCno SHADOW cloud · your keys

CHAT / 01

The agent writes a plan first. Nothing runs until you approve it. Click the button yourself.

THE DIFFERENCE

Chat answers.
Agents do the work.

A chat assistant hands you text and leaves the steps to you. A SHADOW agent takes the steps itself, inside the limits you set, and stops to ask where it matters.

01 / CHAT ASSISTANTyou do the steps

  1. youAsk the question
  2. itAnswers in text
  3. youCopy the answer
  4. youOpen the app or folder
  5. youDo every step by hand
  6. youCheck the result
steps on you0of 6

02 / SHADOW AGENTit does the steps

  1. youGive the goal
  2. itWrites a plan
  3. youApprove the plan
  4. itWorks on your PC and the web
  5. itAsks again before risky steps
  6. itReports what it did
steps on you0of 6

HOW IT WORKS

Plan. Approve. Work. Report.

Every run follows the same four steps, so you always know what an agent is doing and why.

  1. STEP 01

    Plan

    State a goal in the agent's chat. It writes a short plan before it touches anything.

  2. STEP 02

    Approve

    You approve the plan. Each risky step asks again and shows the exact command, path or message.

  3. STEP 03

    Work

    It works in the folders and apps you granted, now or on a schedule.

  4. STEP 04

    Report

    A structured report lands in the chat: what was done, counts, and what is queued but not sent.

WHAT AGENTS CAN USE

Give each agent only what it needs.

You grant tools per agent. Nothing an agent reads, from a web page, a file or another agent, can add to them.

Filesgranted folders

Read, write and edit inside folders you grant to this agent. Deletes go to a SHADOW trash.

asks beforewrite · delete

Terminalpowershell

Runs PowerShell on your PC. You see the exact command before it runs.

asks beforeevery command

Browserown profile

A browser with the agent's own profile, streamed live to the chat.

asks beforesign-in · send · pay · upload

Websearch · read

Searches and reads pages. Page text is treated as untrusted and can never trigger an action.

asks beforenothing to act on

Screenone window

Sees and uses one window you pick, for apps with no other way in. Started by you only.

asks beforetyping · risky steps

Pluginsmcp

Connect apps through MCP connectors, granted one by one.

asks beforeper plugin grant

A fixed blocklist covers system folders, other users' folders and credential stores. No grant unlocks it.

SCREEN CONTROL

Control the screen. Stay in control.

For apps with no other way in, an agent can see and use one window you pick. Only you can start it. A Working overlay stays visible the whole time, and Ctrl+Alt+Esc stops it at any moment.

  • Sessions have time and action limits
  • SHADOW and password-manager windows are never offered
  • Never runs in the background
  • Screenshots go to your chosen model provider
  • Works with Claude models first

Press the keys, or click, to stop the agent.

SKILLS

Agents keep what they learn.

An agent can save a working method as a skill. You can read, edit, disable, delete and roll back every one. A skill never widens permissions.

Researchv4

weekly-competitor-scan

Search 5 sources, compare against last week, write a diff table.

Report Writerv3

release-notes-draft

Edited by you, so it is locked against the agent.

locked

File Organizerv5

sort-by-client

Flagged: replies with this skill finish less often than replies without it.

BACKGROUND · TEAMS

Works in the background. Works together.

ROUTINESapp closed is fine

  • Friday sortFri 17:00
  • Competitor scanMon 08:00
  • Downloads digestevery 60 min

Every few minutes, daily or weekly. A missed run executes once when your PC is back. A run that needs approval waits in the chat and notifies you. Background runs never take your mouse.

GROUPS2 to 8 agents

Agents can message each other and share one group chat. Each hand-off between agents asks you first. A loop between the same two agents is stopped with a reason.

BUILT TO BE TRUSTED

It acts only when you say so.

CODE AGENT WANTS TO RUNtry it

> Remove-Item .\build -Recurse

Pick one. Nothing runs from this page.

01

Permissions per agent

Folders, terminal, browser, screen, plugins. You grant each one. Nothing an agent reads can widen them.

02

Approval cards

They show the exact command, path, message and recipient. Allow once, always allow, or deny. Deny always wins.

03

A fixed blocklist

System folders, other users' folders and credential stores are out of reach. No grant unlocks them.

04

Passwords stay yours

When a sign-in is needed, the agent pauses and you type into the live window. It never sees the password.

05

Logged, with undo

Every action is logged. Deleted files go to a trash and edits keep a copy.

06

Local-first

No SHADOW cloud, relay or sync. Your model keys live in the Windows keychain.

DEVELOPER PACK

For developers: give a code agent GitHub.

The developer pack adds GitHub, Sentry and editor awareness. A code agent fixes an issue in its own git worktree, runs tests in an isolated sandbox and puts the diff, tests and scan on one card. After your tap it opens the pull request, once. Your working tree is never touched.

0

SHADOW cloud services

0

agents per user, groups included

0

model providers

0

hotkey stops screen control

Your models. Your PC.

MODELS

  • Anthropic
  • OpenAI
  • xAI
  • Google
  • Ollama

PLATFORM

  • Windows · x64

TOOLS

  • Files
  • Terminal
  • Browser
  • Screen
  • MCP plugins

FAQ

Frequently asked questions.

01Does my data leave my PC?

There is no SHADOW cloud. The prompts, files and screenshots an agent uses go only to the model provider you choose, with your own key. Use Ollama to keep text on your machine.

02Can an agent act without asking?

Only inside what you granted, and risky actions ask first. You can choose "always allow" for one specific action of one agent. A deny rule always beats an allow rule.

03Can it see my passwords?

No. When a sign-in is needed, the agent pauses and you type it into the live window yourself.

04Which models does it use?

Anthropic, OpenAI, xAI, Google or Ollama, set per agent. You pay the provider directly. Screen control works with Claude models first; other models use every other tool.

05How many agents can I run?

Up to 20, groups included. Each has its own chat, model, permissions, skills and schedule.

06How do I install it?

From the Microsoft Store. Microsoft signs and scans the package, and updates arrive through the Store like any other app. There is no separate installer to download.

07Which platforms are supported?

Windows PCs. Install it from the Microsoft Store.

08What if a run fails or hits a limit?

Every run has limits on turns, time and spend. When one is hit or a step fails, the run stops and the reason shows in the chat. Nothing is dropped silently.

DOWNLOAD

Get SHADOW from the Microsoft Store.

Free. Bring your own API keys. Everything runs on your computer.

Get from Microsoft Store

Free · Delivered by the Microsoft Store

A crew that works while you don't

Hand off files, terminal, browser and screen work to agents that plan the job and finish it in the background.

Nothing happens without your yes

Every action that matters waits for one tap from you. You see what it will do before it does it.

Yours, end to end

Runs on your PC with your own model keys. No account, no cloud of ours, nothing leaves without your approval.

Put agents to work. Stay in charge.

Free for Windows. Bring your own API keys.

Get from Microsoft Store