It asks before. It says what it did. It leaves a receipt.
The sixteen layers: Key, Governor, Canary Camp, Vault, Restore, Panel, Warden, Lenses, Judge, Self Repair, Loop, Library Shield, Airlock, Readback, Black Box, Backstage.
Starpod runs your mission's home: hosting, email, progress, and agents. No promise of perfection: every error is written down, then becomes a rule that blocks it the next time.
Pick your mission. Pick how technical you want to be. Your pod adjusts.
Counted from our incident register, never typed into this page.
Each block below is in the control panel of your pod today.
Your agent's requests to spend money, publish a page, send to your contacts or delete your files wait for your answer.
When your agent proposes an automation, you read it, then confirm or decline it.
See what your pod remembers about you. Correct it, or make it forget.
See what your agent refused, and the rule behind it, in plain words.
Once your two passkeys are in place, you can pause it. While paused, your agent sends nothing, publishes nothing and answers no one on your behalf.
Once an independent audit of your pod has run, you see its grade and its date. A separate record lists what your agent did.
Before some approvals, you say in your own words what you are about to allow. Each check is kept.
Download your data, or ask to close your account.
A pod is a home for a mission. A mission is a pod template tuned to a trade: the same five systems, the same guards, set up for your reality. Pick yours, or bring your own.
Renovation, carpentry, trades. Your work is impeccable; your pod takes care of showing it: your site, your photos, your follow-ups.
Notary, broker, accountant, coach, consultant, trainer: your trade runs on words, and your pod drafts them. First replies, follow-ups, missing documents. The counsel stays yours.
Isolated pods, guards that step in before the action, receipts after. Built by people who run their own agents on it.
For the front desk of a dental office: the paperwork, the follow-ups, the online presence. The office side, never the care side. And never a patient testimonial: your professional order forbids it, and so do we.
Your trade is not here? A mission is something we build. Tell us yours at contact@starpod.ai.
An AI agent that acted without asking. That said “Done ✅” when nothing was done. That apologized, and did it again.
The worst part was not what broke: it was working with something you could not check.
Nobody can; whoever claims it is not telling you the truth. We build a direction, with mechanisms you can check at every turn.
It asks before the doubtful action, instead of apologizing after.
It says what it did.
It produces a receipt.
Every error that happens is written down and becomes a rule that blocks it the next time.
Receipt after receipt, the system becomes more accurate and more precise. Not an absolute: a measurable trajectory.
A pod is a home for a mission: hosting, email that lands, progress you can see, and agents that prepare your communications.
You deal with one manager, an AI. Not eight windows, not a menu of settings to learn. Zero technical decisions up front; you can adjust everything later. The first agent you meet introduces itself and says it is an AI before anything else.
This is how we run Starpod itself, every day. Your pod arrives with its guardrails already wired. Its thresholds are named up front: publishing, spending, sending. And nothing costly, public or irreversible leaves without a human hand authorizing it: yours.
Your pod keeps its word because the protections are built into the whole system, around the AI and not only inside it. There are sixteen today. Seven show on a panel, four work out of sight, and five more are the newest, each with its own receipt. For the first eleven, an illustration shows how it works. They belong to SIA (System Integrity Alignment), the framework of the Laeka Integrity Lab, where each one is defined in the open.
No spending beyond the budget set in advance, and no action that destroys without a backup copy, goes ahead until a person touches a physical key. The system prepares everything, but it cannot give itself that permission.
Before your pod acts, the action is put in a category. The category decides: the action goes on its own, or it stops and waits for a yes. What your pod will ask you is further down this page.
Every hour, the system checks each of its usual routes: it opens its public sites to see that they answer under their own name, and it sends test messages between its internal parts to see that they arrive. A broken route shows within the hour. One test is built to fail: if it ever passes, the checking is broken, not the system.
Passwords and access keys stay sealed in a vault. It opens for one precise action, and none of them is ever written out in plain text. Not every key is sealed this way. Two kinds stay on your pod, declared and written where you can read them: the keys of the nightly backup, so the copies run while everyone sleeps, and your pod's own keys, the ones that sign your sessions and link your pod to us.
A backup copy is made every night. Every week, a pod is rebuilt from a copy, to check that the copies really work on the day they are needed.
Your panel: one page, made for your phone. There you say yes to what your pod wants to do, you see its receipts and your invoice, and it emails you when a decision is waiting for you.
What you entrust to your pod serves your work and nothing else. Your pod is kept apart from other clients' pods and talks only to the services it has been allowed. Before one of your pages goes public, it is checked: if it shows someone's personal information, or a password, it waits for your yes. Your sensitive numbers (social insurance, credit card) never reach the AI provider as they are: they leave hidden and are put back in place on the way back.
You should not have to read forty pages to be protected: the protection is built into the machine.
When a turn changes the client's workspace, independent reviewers each read the deliverable from their own angle before it goes out. A defective deliverable is sent back for rework, up to twice; if it still fails, it goes out marked and a person is told. A reviewer whose answer cannot be read counts as a send-back, never as absent.
Once the answer has gone, a second opinion goes back over the exchange and writes down when it disagrees. Today it only watches: on purpose, it has no power over the answer. The Lenses read the answer; the Judge goes back over the exchange afterwards.
Today, two specific failures fix themselves. For the others, the system knows how to say it is broken, but not yet how to repair itself.
When an error happens, it is written down with what caused it, then turned into a guard that stops the same error the next time. More than two hundred errors written up so far. You only see it in what no longer happens.
Packages reach the pod only through its own mirror: the public registries are closed from inside it. npm installs a real package and answers not found for an invented name. Python's pip is not on the pod, so this page claims npm only.
Private data, untrusted content and a way out must not meet in one turn: after an untrusted read, the way out is cut and any send waits for a person. A test on a trial pod on 2026-11-03 showed the way out cut and a message held for a yes.
A person who clicks without reading is a rubber stamp. When this is switched on, the person approving a sensitive authorization says in their own words what it allows, and an evaluator judges the meaning. A wrong reading is refused and nothing is written; a right one goes on to the keys.
An audit from outside the pod. With no model, it reads the journal of the pod's actions and catches any logged action executed without its decision. Only this passive audit is proven today.
The other layers watch what the agent does; this one inspects what it is mounted with: its agent files, hooks, MCP servers, permissions and secrets. A scan runs every day from the pod's host; each pass plants three inert witnesses and fails if it misses one. A finding names the file and the rule, never the value of a secret.
The first public form of the specification lives at the Laeka Integrity Lab.
Inside your pod, Star works with your private files. Sometimes it pulls a page or a file from the internet, and that content can hide instructions written by someone else. The third ingredient is a way out: a message, a published page, a paid image.
The moment Star pulls something in, it remembers it for the whole conversation. Each time it works in that conversation, the pod's way to the internet is cut. A message it prepared does not leave without your yes, given after the outside content came in. A scheduled task that has read outside content sends nothing at all. If the cut cannot be confirmed, Star stops. It watches what Star pulls in itself; text you paste in counts as yours.
A pod does not ask permission for everything; that would bury you. It stops at three kinds of thresholds: what costs money, what becomes public, what cannot be undone. Below them, it runs. At them, it waits for a hand.
Illustration only: these are not your thresholds. Yours are set with you. And the hand that authorizes is yours.
An action that costs real money beyond a capped budget set in advance, or that destroys without a net, waits in a queue until a human touches the key. The AI prepares everything: the script, the backup taken before, the scope, the receipt to come. What it cannot do is give itself the permission. That boundary rests on a signature only the physical key can produce, not on a policy.
We run Starpod under our own keys today, and it is proven.
Every action that matters produces a receipt: what ran, when, what it touched, how it ended. A claim without a receipt does not count. The rule holds against us too: nothing goes out to a human before a reviewer passes it, and its refusals are final, including against the founder.
Our own board keeps the failed receipts visible next to the successful ones. Hiding them would be the first lie.
Two real receipts from our own board, in plain words.
Days without a major incident.
The number comes from the incident register of our own operation: this page counts the days since the date we publish. Read the full register, with the date of the last incident. It is not typed into this page. “Major” has a written definition:
The register keeps the minor incidents too, each with the rule it became. A counter can reset one day; counters do. The mechanism that turns every incident into a guard does not reset.
You never open an account with an AI provider, never paste a key, never read a token bill.
For agentic developers, small AI labs and automation consultants. The same pod and a high token ceiling. No human work time is included: you do the building.
We talk by email: contact@starpod.ai
Talk to us firstSeveral pods for one team. The same guards and the same receipts on every pod.
We talk by email: contact@starpod.ai
Talk to us firstMissions built for your organization. The scoping comes first, the written perimeter next, the work after. Nothing starts on a vague promise.
We talk by email: contact@starpod.ai
Talk to us firstThe fleet and the organization always start with a conversation, both ways, on purpose: we do not take a client we cannot serve. Builder too, until its access through the API is ready.
Prices in US dollars, plus applicable taxes.
Why $99 when site builders start at $25: an isolated pod, a guarantee, and your yes before every action that matters.
One email to start. You pick your plan and pay on the next page, at Stripe; no card number is asked on this site.
Privacy: this form sends your email and nothing else. It is never sold and never used for advertising; Stripe and Resend receive it, to take your payment and to send our emails. Questions about your data: privacy@starpod.ai. Privacy policy.