Resources
Where the limits are
Each limit is in code or in the operating system. None depends on a model following a rule in a prompt.
On the machine
Two processes run on each machine. They have different jobs and different permissions.
- The agent does not run commands
The agent talks to the server, reports facts and receives orders. It never runs a command. It does not hold the signing keys, so a compromised agent cannot make the executor run an unsigned recipe.
- The executor cannot reach the network
A separate root process runs the recipe. The unit sets
PrivateNetwork=yesandRestrictAddressFamilies=AF_UNIX. It verifies the signature with keys built into the binary.- A closed catalog of recipes
A recipe is a signed document. Each step is a fixed list of arguments with typed parameters. No step is a shell string. A model can select a recipe and fill its parameters. It cannot write a new step.
- Live facts before the run
The executor checks the precondition of the recipe again on the machine, with live facts, before it runs.
Decisions in code
- Five gates
Each proposal goes through five gates, all in code. Each answer is allow, ask or deny, and each answer is written on the timeline of the incident.
- The Guardian only raises the risk
The Guardian reads the rendered steps. Its rules and its model can raise the risk of a recipe. They cannot lower it.
- A classifier that never sees the diagnosis
Gate 5 is a safety classifier. It reads the recipe and the machine, never the diagnosis, so the text of a diagnosis cannot argue with it.
- The mode can only go down
Audit reads and reports. Shadow shows what Kosken would do. Live runs approved recipes. A limit at workspace level lowers a machine, and nothing raises it.
- No tool to close an incident
The diagnosis can end with
conclude, which takes no action. A model has no tool to close an incident. An incident is resolved only after the check of the fix passes.
People and approvals
- A person decides
A risk above low, a recipe with no trust on that machine, or a gate that cannot decide sends the request to a person.
- High risk needs two people
A high-risk request needs a second person with write access, who confirms their identity again before they approve. When the workspace has one writer, that person confirms their identity again.
- Sign in
Email and password, passkeys, Google, or single sign-on.
Workspaces, keys and models
- Tenant isolation
Row-level security separates workspaces in the database. Each workspace has its own keys, derived with HKDF and sealed.
- Your model key
You bring your own key for OpenAI, Anthropic or DeepSeek. Kosken does not issue model keys.
- Fixed endpoints, redacted context
The endpoints for model calls are fixed in code. Context is redacted before it leaves.
An audit log that shows tampering
Each audit entry includes the hash of the entry before it. A signed anchor is written each day. If someone edits an old entry, the chain no longer verifies.
This detects tampering. It does not prevent a person with full database access from deleting the whole log and its anchors. Anchors stored outside the database reduce that risk.
More of Kosken
- Incident responseFrom a failed signal to a root cause report and a checked fix.
- MonitoringMonitors for services, ports, URLs, containers, disks and logs.
- Recipes and approvalsSigned fixes, five gates in code, and a person who decides.
- Tasks and maintenanceScheduled checks, planned changes, and the jobs of each machine.
- Fleet and modesMachines, workloads, and the audit, shadow and live modes.
- Ask KoskenAsk about your fleet in plain words, with your own models.