The AI worker's own computer

The idea

Crewshal can give an AI worker its own computer: a private Linux machine that keeps its tools and its work like a laptop. When an admin issues it from the worker's page, it stays until someone revokes it or the worker leaves the organization. A second person approves what is installed on it and which hosts it may reach: an owner, or an admin who did not issue it. The worker's own browser can open only those hosts, their subdomains and apps that run on the computer itself. The feature is built and switched off in production.

The problem
A worker without its own machine cannot keep sign-ins, tools or state.
Why it matters
The worker keeps its environment and accounts, and a revoke fences it at once.
Where it stands
Built, off in production as of Shown on a real guest computer on a local backend in owned runs on November 14 and 15, 2026; production keeps it switched off.

Crewshal team · · 5 min read

Recorded run · Owned QA, a real guest on a local backend · November 14–15, 2026

What happens on the worker's own computer?

A real guest computer is issued, approved, set up, checkpointed, restored and revoked. A second computer's browser reaches only approved sites. In production the feature is switched off.

The first computer, November 14, 2026

  1. Approval

    Request from Admin

    Waiting

  2. Onboarding job

    1. Provision
    2. Allowed hosts
    3. Install tools
    4. Health check
    5. Checkpoint

Checkpoint

none yet
Network boundary

The worker's computer

Requested

Linux engineer, version 1

  • gh
  • acli
  • node
  • codex
  • playwright
  • claude-code
  • agent-browser

Network proxy9 hosts requested

  • github.com
  • nodejs.org
  • registry.npmjs.org
  • 6 more

Step 1 of 12Lifecycle

AdminAn admin issues the worker a computer

An admin asks for a computer for one AI worker. The request names a built-in install manifest of pinned tools. The manifest names the 9 hosts that the computer may reach, such as github.com and nodejs.org. This computer is kept until someone revokes it. Nothing is set up yet.

InputThe request

Install manifest
Linux engineer, version 1: 7 tools with pinned versions, plus system packages
Allowed hosts
9, such as github.com, nodejs.org, registry.npmjs.org
Kept
Until revoked

CheckRecorded gates

  • An admin created it through the APIPassed

Where it ran. Two runs on November 14, 2026 and November 15, 2026 on a Mac (arm64, Docker 29.8.2): a local backend in product mode issued a real guest computer, a Docker container from the platform's guest image, through the product path. The checks called the platform and the worker's tools directly; no AI model ran. Production keeps worker computers off, and a production check on November 15, 2026 confirmed it.

The recorded files behind this replay
  • Real guest on a local backend: issue to revoke, 19 gates (November 14) sha256 c315d72bac0e
  • Real guest with a browser: allowed site, screenshot, refused sites, 19 gates (November 15) sha256 6b986f7eb85f
  • Release record: worker computers released switched off in production sha256 59a57e98f9be
  • Release note: the worker browser, its screen and site approvals sha256 ff550d5aef05
  • Production check: worker computers are off, 9 checks (November 15) sha256 3588fd4a8a00
  • Production capture: the worker Setup tab with no Computer group sha256 f31f44d9097b
  • Local screen check: the admin sees the worker screenshots (November 15) sha256 3fc4b65b0ed0
  • Local screen check capture: the admin Screen view sha256 1f1c5523b8cd

Why it matters

An AI worker that builds and checks real software needs a real machine: its tools, a browser, and a place to keep its work between tasks. A throwaway sandbox loses that state at the end of each task, and a machine without limits can reach anything on the internet.

The worker's own computer keeps its state inside fixed limits. The platform installs only the tools in its approved manifest. The computer reaches only the hosts on its approved list. A second person approves both, and one revoke takes the computer away. Inside those limits, the worker runs its own commands.

How it works

  1. An admin issues it. The request names one worker and a built-in install manifest of pinned tools. The manifest names the hosts that the computer may reach, and the admin can add sites.
  2. A second person approves it. The request goes to the owners. An owner, or an admin who did not issue the computer, can approve it. The admin who issued it cannot. One approval covers the computer's manifest and its hosts.
  3. One job sets it up. A background job provisions the guest, applies the allowed hosts, installs the tools and runs a health check (the doctor). It then takes a checkpoint and tests a restore. After a crash, the job goes on from its saved record. It does not start a step again while that step still runs in the guest.
  4. Only approved hosts. A network proxy lets the computer reach its listed hosts and their subdomains, and refuses the rest. A new host needs an admin's request and the approval of a second person, an owner or another admin. The proxy is changed and read back before the tools work again.
  5. Its own browser. The worker opens, reads, clicks and screenshots pages through governed tools. What a page says counts as outside content. A later step that runs a command, cannot be undone or acts outside the computer waits for a person when it repeats a command or a long passage from the page, or a link from the page to a host that the organization does not already work with, or when the page holds text that looks like an instruction. The platform checks each screenshot by its SHA-256 and keeps it for the organization's admins, up to 200 per computer for 30 days. The worker can also save a copy on its own computer.
  6. A revoke takes it away. The platform retires the guest and deletes its checkpoints, and the worker's next tool call is refused. A revoke does not remove access that a vendor, such as GitHub, gave the worker. An admin removes that access at the vendor.

What the run showed

The replay above draws on three recorded runs and one image from a separate local screen check. The first two runs were on November 14 and 15, 2026, on a Mac: a local backend in product mode issued a real guest computer, a Docker container from the platform's guest image, through the product path. The checks called the platform and the worker's tools directly; no AI model ran.

  • On November 14, an admin issued a computer with 9 allowed hosts, and the owner approved the high-risk request. One job provisioned the guest, applied the hosts, installed the tools (1 min 31 s) and ran the health check: 23 of 23 checks. All of onboarding took 2 min 25 s, on its first attempt.
  • The job took a checkpoint, wrote a marker file into the computer and restored the checkpoint. The marker was gone, the 7 tool versions were the same, and the health check passed 23 of 23 again. The check then looked in the guest itself and found no marker.
  • Hosts outside the list were refused. A second start did nothing, and a job replayed as if after a crash did no new work in the guest. After a revoke, the guest and its checkpoints were gone, and nothing of the run was left. All 19 gates passed.
  • On November 15, a second computer, with a browser, opened nodejs.org, read the page, scrolled and took a screenshot of 1279 by 656 pixels, which the platform served to admins only. Opening example.org was refused before it reached the computer, and a link to discord.gg was refused by the proxy. After a revoke, the next call was refused. All 19 gates passed.
  • The same day, a check on production with an operator test account found worker computers switched off: no computers, and a sign-in, a site request and the screenshot list refused. All 9 checks passed.

Limits

  • Off in production. Customers cannot use worker computers yet; production keeps the feature switched off.
  • A local guest. These runs used a Docker container on a Mac. The host for customers' computers is not decided yet.
  • No AI model in these runs. The checks called the tools directly. An owned lab run on November 16 used such a computer with a real model; see Correct once.
  • Sign-in is not shown. Signing the worker's own tools in with a device code was tested in a separate run that stopped at the code prompts.

Frequently asked questions

Can the worker reach any website?

No. The computer reaches only the hosts on its approved list and their subdomains. A new host needs an admin's request and the approval of a second person, an owner or another admin, and the platform changes the proxy before the tools work again.

What happens when a computer is revoked?

The platform retires the guest and deletes its checkpoints. The worker's next tool call is refused, because the worker has no computer. A revoke does not remove access that a vendor, such as GitHub, gave the worker; an admin removes that access at the vendor.

Who sees what the worker's browser shows?

The worker reads the pages that it opens. The platform keeps the worker's screenshots for the organization's admins, up to 200 per computer for 30 days, and checks each one by its SHA-256. The worker can also save a copy of a screenshot on its computer, for example to attach to a pull request.

Can I use this today?

Not yet. The feature is built and switched off in production, and a production check on November 15, 2026 confirmed that it is off.

  • One approval by a second person for the computer and its sites

    One approval by a second person (an owner, or an admin who did not issue the computer) binds the computer's install manifest and allowed hosts; onboarding runs as a resumable job with a doctor, snapshot and restore, and each added site needs another such approval.

  • Worker Console tab

    Each AI worker's Console shows what it does now, what needs a person, a closed list of health checks with read times, and its history; a check failing 5 minutes opens an alert.

  • Standalone autonomous AI employee

    A separate offering: one AI employee with its own private persistent computer and enrolled accounts that onboards itself, does a real engineering task alone and learns from owner corrections.

  • Independent verifier

    The platform checks an AI worker's pull request in its own container, out of the worker's reach, and approves it only when a new test fails without the fix.

    Recorded walkthrough

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.