Skip to content
Sekos
Permission controls for AI agents

We keep your agents in the lanes you set, with proof they stayed there.

Your team sets the lanes. For enrolled capabilities at enforced resource boundaries, Sekos checks each requested action against those permissions and retains decisions and observed outcomes for inspection. Evidence covers those enforced capabilities; alternate routes and unknown outcomes need separate checks.

For engineering and security teams giving agents access to business systems.

Refund demo · 4 min 11 sec · Fictional customer dataTry it yourself ↗
Owner-defined permissions before the actionAllowed, refused & unknown stay distinctInspectable evidence after the action
01 / See the boundary

Useful work.
Visible limits.

The same agent can make a valid request and a request outside its permission. Inspect the result of each—not just its explanation.

The request

“Refund this order $25.”
Permission
Approved refund within the item value.
Confirmation
Separate from preparation.
Recorded case
N01 · Authorized refund
ALLOWED & CONFIRMED

The refund happened once.

Preparation caused no effect. After confirmation, the retained record reports one refund and one token consumed.

Effect: one refundRetained internal case

Static retelling of the September 29 Northstar package. Fictional customers; scripted model replies and simulated transport in the case pack. These are reported internal outcomes, not a fresh run or customer result.

Controls have a scope. Every alternate route and credential holder needs its own check. Reported internal examples. Live demo behavior must be checked within its own scope.
02 / How it works

Permission comes
from your team.

The model may propose. It cannot authorize itself. Your rules determine what may happen at the protected resource.

  1. 01

    Set the permission.

    Name the actions, resources and limits the agent may use. Its explanation cannot grant it more access.

    Governed operations describe the available actions and their rules.

  2. 02

    Check the exact action.

    Check the identity, target, arguments and current conditions before the protected consequence begins.

    The Resource Deputy guards that resource; Schemen Gate verifies the bound authority.

  3. 03

    Inspect what happened.

    Keep the decision and observed result. If the outcome cannot be confirmed, retain that uncertainty.

    Substrate supports typed computation and evidence workflows. Gate can be used independently.

03 / Choose your next step

Inspect the technology.
Test your workflow.

Understanding the boundary comes first. Choose the path that fits your current decision.

Technology path

Explore the controls.

Inspect the authority and enforcement primitives, then discuss the boundary your own resource needs.

  • Schemen Gate: public implementation
  • Resource Deputy: resource-side enforcement
  • Integration scoped to your deployment

Try the demos

The live refund sandbox lets you inspect its current decisions and fictional database effects. Opening its page does not grant access or execute a refund.

An integration discussion is not a ready-made production deployment or a universal security guarantee.

04 / Evidence & limits

Know what the
record establishes.

These examples report an internal synthetic case pack. They do not establish production customer outcomes, universal safety or satisfaction of the user's intent.

September 29 Northstar package · fictional data, scripted model replies and simulated transport
CaseReported decisionObserved effect in the case pack
N01Allowed, then confirmedOne $25 refund; one token consumed
N07Above-item-value refund refusedNo refund; no token consumed
N23Outcome unavailableUnconfirmed; reconcile before another mutation
Not proved by the Gate alone. Alternate routes, unmediated credentials, host isolation and whether an authorized action meets the user's intent require separate checks.

The internal source package is retained privately.

A separate AI-assisted review of public Schemen Gate commit f35ef8ad8242 reported 375 targeted tests passed and 7 skipped. Those are the review's reported results, not customer acceptance or security certification. Read the review · Inspect the reviewed commit.

New · Live demo

Write it in plain English.
Query it like a database.

OntQL is the query language of the Substrate, the memory engine under Sekos. In the demo you write an ontology in short English sentences. Definitions compile into types, and every fact is checked against them before anything is stored. Undefined words and wrong types are refused, with the line and column. Each answer names the sentence it came from. When two facts conflict, or a fact was never recorded, the answer is UNKNOWN, not a guess.

Your own private Substrate, kept in memory for 20 idle minutes. No language model reads your sentences, and no sign-in is needed. The server runs 25 checks of the tour before it starts.

05 / Research & Substrate

Go deeper when
the question calls for it.

Explore the research behind the tools, including the experiments that failed and the limits of the results.

Substrate is proprietary. Public descriptions and reviewed interfaces remain separate from implementation and protected representations.

What should your agent
be allowed to do?

Start with one useful action and one boundary it must respect.

Discuss your workflow
UX NOTES · 0
Click anything on the page: the element and its internals are captured and copied to your clipboard immediately. Add a note if you want, or keep clicking. Esc pauses picking. Nothing leaves this browser until you paste it.

Content version hash

SHA-256 of this text, with whitespace normalized. This identifies the recorded text; it does not verify the claim or an action.