Human approval on sensitive actions
Sending a proposal or a contract requires your decision. Agents can ask for approval; they can never give it. And you choose how much autonomy each employee has.
AI employees that work in plain sight. You decide what matters.
Ask Sales, Legal, Accounting, Analysis, Operations, HR and your Assistant for work in plain language. Watch them split it up in a 3D office, approve what is sensitive and get a report of what was done and what it cost.
Already have an account? Sign in
Start free, no card needed. Pro is 29 USD a month and Max 99 USD; pay yearly and get two months free.
What ships today
- 7base employees, each with a role, tools and memory
- +6more professions you can hire from a template
- 4AI providers with your key: OpenAI, Claude, Kimi or DeepSeek, plus any OpenAI-compatible one
- 0keys the model sees: they live in an encrypted vault
The problem
Most AI agents work behind closed doors.
You hand over a task and get a result back. What happened in between stays in a black box.
What did they do?
A message history does not tell you who touched what, or what was sent outside.
How much did they spend?
The cost shows up later, on an invoice, when you can no longer decide anything.
Did they get it right?
If you cannot see the process or who they consulted, reviewing the result is guesswork.
AI Workforce OS starts from the opposite premise: everything an AI employee does is visible, has a spending cap and stops to wait for you when it matters.
How it works
From one sentence to a report, without letting go.
You write what you need. The Assistant builds the plan, the team works in parallel, you approve what is sensitive and you get the result.
-
You ask
In plain language, as you would ask a colleague. Before it runs you see a cost estimate.
-
It gets planned
The Assistant splits the work into tasks with dependencies and hands them out.
-
They work in parallel
Whatever does not depend on something else runs at the same time, and they consult each other.
-
You approve
Before anything goes out, the employee stops and asks for your go-ahead.
-
You get the report
What was done, who took part and what it cost, in one place.
The real app
Screenshots of the app, not mockups.
What you see on screen comes from real system events. These screenshots are from simulation mode.
Control
You set the limits. The backend enforces them.
These capabilities exist in the app's code today. Anything that depends on real integrations is listed under “Current status”.
Cost estimates and caps
You see an estimated range before anything runs. There are caps per request, per employee and per organization; when a request or employee cap is hit, the work pauses and you decide whether it continues.
Kill-switch and read-only mode
Freeze all activity or put the office in read-only mode. You can also pause a single employee or switch off a specific tool.
Audit trail
Every relevant action goes into a hash-chained log the app can only append to, never edit or delete. It can be exported and verified.
Credentials kept away from the model
Keys live in an encrypted vault (AES-256-GCM). The AI model never sees them: the backend runs the tools.
Projects with parallel work
A project is split into tasks with dependencies; whatever does not depend on something else runs at the same time.
See the exampleTemplates and business packs
Five workflow templates and four packs (agency, commerce, professional services and general) that set rules and autonomy.
Built in Spanish first
Interface and employees in Spanish (and English), with a configurable regional tone. Ships sample tax templates for Mexico and Spain, labeled as examples.
Parallel work
One project, several employees at once.
Example: the quarterly financial close, as in the Projects screenshot. The Accountant does not finish one thing before starting the other.
The team
The employees of your line of work. More when you need them.
When you sign up you pick what you do and start with the employees of that job plus an assistant. Each one has a role, tools, memory and a history of what it did, and you can hire other professions when you need them.
Sofía Lara
Assistant
Plans, coordinates the team and consolidates the final report.
Valeria Ríos
Sales
Analyzes the client and prepares proposals.
Elena Castro
Legal
Reviews contracts and flags risks.
Tomás Vidal
Accounting
Calculates margins and builds financial statements.
Nadia Ortega
Analysis
Assesses profitability and draws findings from data.
Iván Duarte
Operations
Checks capacity, deliveries and staffing needs.
Marcos Peña
HR
Prepares hiring processes and role profiles.
Hire from a template
Professions described as configuration, each with its risk level and approvals:
- Data analyst
- Instructional design
- Finance and treasury (analysis only)
- Internal audit
- Project manager
- Software engineering
Employees prepare and organize. They do not give legal, tax or medical advice: those decisions belong to a human professional.
Security
The model proposes. The backend decides.
Design principles you can check in the technical docs. They are not certifications.
- The backend decides, the model proposes
- The AI engine returns results and tool requests. The backend decides whether they run, need approval or are rejected.
- External content is data, not instructions
- Emails and documents an employee reads are handled as delimited data, so they cannot give orders to the system.
- Isolation per organization
- Data carries its organization identifier and the database enforces row-level security, using an application role that cannot bypass it.
- Least privilege
- An employee only uses the connections and capabilities you authorize, and every permission is recorded.
Current status
What works today, plainly.
We update it when something changes. Last reviewed in October 2026.
Works today
3D office, dashboard, projects, approvals and costs
End to end in simulation mode: scripted, realistic responses, no AI key and no cost.
Real AI with your own key
OpenAI, Claude, Kimi, DeepSeek or a compatible endpoint. It works, but we have tested it less than the simulation.
Accounts and separate offices
Email sign-up, a 24-hour trial account and invitations with roles (owner, admin, member, viewer).
Work that survives a restart
Requests in progress and pending approvals resume on startup, with a single server instance.
Built, not yet proven in real use
Import data from any system
A generic API source the agents can query, data inboxes (a URL with a token that any software can send JSON, CSV or files to) and import from a link (CSV, JSON or Excel, once or on a schedule). No integrations tied to one product in this version.
OIDC single sign-on, verified domains and SCIM
Implemented in the code; not yet validated with customers.
Plans and card payments
Free, Pro (29 USD a month) and Max (99 USD a month), paid through ONVO Pay. Tested in test mode; no real charges yet.
Planned
WhatsApp and other messaging channels
Designed on paper. They do not exist today.
SAML and multi-instance recovery
Pending. SSO is OIDC only today and recovery assumes a single server.
Getting started
Your first task, in five minutes.
Create your account and within five minutes you will have asked for something, watched the team work and approved your first action.
Ask
Type what you need in the office bar, the way you would ask a team lead. There are shortcuts to try.
Watch
Your AI employees split the work and you see it happen live: who does what, what it costs and how far along it is.
Approve
Anything sensitive stops and waits for your decision. You approve or reject from the tray; nothing important leaves without you.
Receive
You get a report, and every finished task earns coins to dress your team and decorate the office.
FAQ
Frequently asked questions
Does this use real AI?
That is up to you. Without a key it runs in simulation mode, with scripted results and no cost. With your own OpenAI, Claude, Kimi or DeepSeek key it uses a real model; that path works, but we have tested it less than the simulation.
Can it send emails or messages for me?
Not in this version. The agents prepare drafts and documents that you review and send. Connectors for specific products (Gmail, Google Calendar and Drive, Slack, LinkedIn, GitHub) exist in the code, but they are switched off in v1 and have not been tested with real accounts; they are turned on on request. WhatsApp is not built.
What happens to my data and my keys?
Keys go into an encrypted vault and the model does not see them. With real AI, task text is sent to the model provider you configure; check its terms before using sensitive data.
Can an agent approve something on its own?
No. An agent can ask for approval, but the decision is always a person's. You adjust what requires approval and how much autonomy each employee has.
Does it replace my team?
That is not the idea. It is a support team that prepares, organizes and consults; important decisions stay yours. It does not replace a lawyer, an accountant or any other professional.
What is “Try it without signing up”?
A throwaway trial account, with its own office, that stops working after 24 hours. It asks for no email: the username and password are shown once.
How much does it cost?
There is a Free plan to try it for real, no card needed: 1 USD of AI a month, 1 active project and 2 people. Pro is 29 USD a month (12 USD of AI, 5 projects, 10 people, 10 GB) and Max 99 USD a month (45 USD of AI, 20 projects, 50 people, 50 GB). Pay yearly and get two months free. The AI never spends past your plan's cap.
Is it suitable for a large company?
Not yet. OIDC single sign-on, SCIM and restart recovery exist, but we hold no certifications, there is no SAML and it has not been validated with large companies. Today it is aimed at small teams and Spanish-speaking professionals.
Ask for your first task today.
Create your account with your email and your company name: you land in your own office, separate from everyone else's. You start on the Free plan, no card needed. This page collects no data: signing up happens in the app.