Cross-platform build & QA
Give an agent clean Windows, macOS and Linux environments to install, test and validate the same release without contaminating a developer laptop.
Disposable labs · repeatable evidence
Secure computer-use for people and AI
Let AI work on computers. Keep control.
Give people and AI agents governed access to Windows, macOS and Linux—inside disposable labs or approved real devices. Watch every action, approve high-risk steps, take over instantly and keep the evidence in your environment.
Deny by defaultHuman takeoverCustomer-hosted
$ task
Improve this workstation
screen stream · commands
files · before/after report
Improve this workstation
Agent captured baseline
Plan ready for approval
Waiting for owner
$ evidencePlan ready for approval
Waiting for owner
screen stream · commands
files · before/after report
Windows workstationLIVE · VIEW ONLY
242processes
15startup items
1.7 GBmemory free
Approval required
Disable 4 startup items?
Disable 4 startup items?
Agent: Codex · lease 27 minHuman can take over
ONE GOVERNED LAYER ACROSS
⊞ Windows
⌘ macOS
⌁ Linux
›_ SSH & terminals
◎ Browsers & apps
The control plane for computer-use
From request to result, with a human in charge.
AnvilWork binds the task, identity, machine, tools and evidence into one governed work session. AnvilVDI is the cross-platform desktop layer underneath—not the whole product.
01 / SCOPE
Choose the environment
A clean lab, a team VM, an approved workstation or a private server. Nothing is discovered or granted implicitly.
02 / LEASE
Assign a person or agent
Name the operator, allowed tools and duration. Access expires automatically and one controller holds the lease.
03 / WORK
Watch and intervene
See the live desktop beside the terminal. Review actions, approve risky steps, pause or take over at any moment.
04 / PROVE
Keep the evidence
Export commands, screen evidence, files, approvals and a before/after report tied to the exact machine and task.
One product, real work
A safe place to build, test, operate and help.
Use the same governance model whether an engineer is testing an app, an agent is investigating a machine, or an operator is changing infrastructure.
Operations & migration
Combine desktop, terminal, SSH and files in one scoped session. Rehearse in a sandbox, review the plan, then apply only the approved change.
Test → approve → promote
Investigation & assistance
Audit a real computer, prepare a change plan, ask for consent and benchmark the result. The owner stays visible and can revoke control instantly.
Attended or time-limited accessSee who is doing what
Live work—not a mystery process.
A shared operational view keeps humans and multiple agents from stepping on each other. Every session has a named owner, target, tool and lease.
01
Live Wall
See active human and agent sessions across desktops without turning unrelated browser tabs into fake machines.
02
One controller, many observers
Prevent collisions with an exclusive control lease while reviewers watch or take over intentionally.
03
Desktop + terminal together
Follow the agent's reasoning and tool calls beside the live target instead of reconstructing work from screenshots.
04
Task-linked reports
Attach findings, plans, approvals and before/after measurements to the machine and task—not to a fragile chat transcript.
Windows performance reviewBASELINE CAPTURED
242running processes
15startup items
6 GBphysical memory
1,145 msshell launch baseline
Evidence identified a mechanical HDD and memory pressure as the primary constraints. No cleanup claim is made until the approved post-change benchmark is captured.
A real workflow, not a canned chatbot
Audit. Plan. Approve. Apply. Measure.
The agent does not jump from “slow PC” to deleting software. It collects a baseline, explains likely causes, classifies risk, waits for approval and measures the same schema afterward.
- Read-only assessment first
- Item-by-item approval for state changes
- Destructive operations excluded unless explicitly authorized
- Exportable Markdown and structured evidence
Security by architecture
Agents cannot reach what you did not grant.
AnvilWork is designed so production systems, new resources and real devices stay outside an agent's reach until an owner creates an explicit, expiring capability.
Deny by defaultNo ambient machine discovery
Named identitiesPerson, agent and lane recorded
Scoped capabilitiesResource, tools and time bounded
Immediate revocationThe next action is stopped
CUSTOMER BOUNDARY
control plane · relay · evidence · policy
control plane · relay · evidence · policy
Bring your own people. Bring your own agents.
Give them a safe place to work.
Build and test across operating systems. Operate approved infrastructure. Help on real devices. Keep control and evidence in one place.
Private beta · customer-hosted option · no spam