Skip to content
For enterprise

Let agents act. Keep the controls.

OrchStack is the operating system for the AI-native company. At its core is OrchKernel: one gate between every agent and everything it touches, on your infrastructure.

The questions

What security asks before agents act

Once an agent can send email, issue refunds or change records, two questions matter. Who is it acting for, and who said it could?

01

Who holds the keys?

Each agent holds its own API keys to the CRM, billing and email. Every new agent adds another set to track.
02

Who approved it?

Risky actions happen first and get noticed later.
03

What does it know?

Nothing is shared. Each agent starts from zero.
04

Can you prove it?

There is no trustworthy record of what an agent did and why.
Governance

Every action passes one gate

Agents propose. Rules decide. The kernel applies. The gate runs the same checks, in the same order, for every agent.

01 Kill switches
Has someone switched this off?
02 Arguments
Are the inputs valid for this tool?
03 Delegation
Does the person it acts for allow this?
04 Budgets
Is there budget left for the agent, team and task?
05 Rules
Does a company rule allow, hold or deny it?
06 Risk against trust
Is the risk within this agent's trust tier?
AllowedNeeds approvalDenied, with the reason

Example: a $240 refund waits for an admin, who sees the exact payload. Once approved, it runs exactly once.

01

Delegation only narrows

An agent never has more than the person it acts for. Revoke it and everything stops.
02

Trust is earned

Trust is earned: sandboxed, standard, trusted. Riskier actions need more trust. Ready-made agents start at standard; modules your company uploads start sandboxed.
03

Budgets and kill switches

Budgets per agent, team and task. A kill switch is the first check at the gate, so the next action stops there.
Audit

Proof an auditor can check

Know what happened, show it, and change the rules the way you change code.

A tamper-evident log

People, agents and systems write to one event log. Each event is chained to the one before, and ok audit verify names the first break.

Replay any run

Any agent run can be replayed from the log, so a question about what happened has a checkable answer.

Policy as code

Export rules, budgets and kill switches to Git. Check them in CI and apply changes through review.

Policy dry run

Ask the kernel to explain what a rule change would allow, hold or deny before you apply it.

Self-hosting

Runs where your data lives

You host the kernel. You pick the models. You decide which hosts it may call.

Self-hosted

One binary and one database file, with Docker and Kubernetes assets.

Your model providers

Each model is cleared for a data class. Restricted data reaches only the models cleared for it.

Sealed secrets

Vendor credentials are sealed at rest. Agents never hold them; the kernel runs each approved action with its own credentials.

Checked outbound traffic

Every outbound request is checked against your list of allowed hosts.

Agents you already use

Claude, Cursor or your own code get their own identity and act only through the gate.

Multi-tenant hosting

A kernel per tenant, under a commercial license from OrchStack.

Enterprise questions

Put one real workflow through the gate

Design partners run OrchKernel on their own systems, with a direct line to the engineers building it.

Runs on your own servers. The repository opens soon.