A pod is a bounded container with a pinned agent core, deterministic guards wired ahead of every tool call, a physical key for the actions that cost money or destroy, and receipts as first-class objects. We run our own company on it.
In your hands, obviously: the tech on screen. And if your lab grows, In many hands is right there.
It is the harness we work in every day, not a feature list written for a launch: the panel, shown tab by tab on laeka.ai, is ours, the receipts on the board are ours, and the reviewer whose refusals are final refuses us too.
Your agent proposes a deploy. The card sits in the queue with its script, its backup taken before, its scope. It waits for the touch; it cannot touch the key itself.
The model emits a destructive command. The guard refuses before execution, exit 2, the refusal is journaled, the agent gets a clear ⛔ instead of a polite apology after the damage.
You rotate a provider key. The vault opens for that one action; nothing gets written in clear.
A pod is rebuilt from last night's copy. The restore result lands on the panel, dated.
You read the journal: agents in a pod cannot whisper, their coordination is readable by design.
Example receipts, so you see the shape. The first one mirrors a real bench we run on our own guards; on a real pod, every receipt is a real event, and EXIT=FAIL stays on the board.
Sixteen named layers today, honest statuses: seven show on a panel, four work out of sight, and five more are the newest, each with its own receipt. Key (physical key beyond a capped budget, and on destruction), Governor (class decides what passes alone), Canary Camp (hourly self-mark with a probe that must fail), Vault (sealed secrets, except two declared kinds of keys that stay on the pod), Restore (nightly copy, weekly proven rebuild), Panel (ours today), Warden (isolated pods, public pages checked, sensitive numbers masked before the AI provider), Lenses and Judge (observation, journal only), Self Repair (thin, and said so), Loop (every error becomes a guard), Library Shield (packages only through the pod's own mirror, npm only), Airlock (after an untrusted read, the way out is cut and any send waits for a person), Readback (when switched on, the approver says in their own words what an authorization allows), Black Box (passive audit of the journal from outside the pod; thin, and said so), Backstage (daily scan of agent files, hooks, MCP servers, permissions and secrets). First public form of the spec at the Laeka Integrity Lab.
Your pod: $99 a month. Flat, tokens included, with a real person for technical support.
Builder, The fleet and The organization start with a conversation.
Changed your mind? The pod is refundable during its first 30 days, prorated. Cancel online, no call, no reason needed. Everything exports.
Prices in US dollars, plus applicable taxes.
We send you a sign-in link; then you pick your plan and pay at Stripe. Your pod arrives with its guardrails already in place. Its thresholds are named up front: publishing, spending, sending. And nothing costly, public or irreversible leaves without a human hand authorizing it: yours.
Privacy: your email is never sold and never used for advertising. The details are in our Privacy policy. Questions about your data: privacy@starpod.ai.