Start with an outcome
Ask for a bounded result. Kerna can retrieve relevant local context and discover configured MCP tools.
A visible control loop
The agent can plan.
You keep the authority.
Kerna sits between an AI model and MCP tools. It owns orchestration, policy checks, budgets, approval decisions, and the evidence of what occurred.
What happens in one task
Every important step has a place to stop.
Choose a tool
The model proposes a call. The proposal does not become an action merely because a model requested it.
Check the boundary
Permissions fail closed. Kerna evaluates the allowed tool, routine scope, approval policy, and hard budget before invocation.
Isolate the connector
MCP plugins run as child processes. A hung plugin is bounded and reaped instead of quietly taking over the task.
Ask when it matters
Approval-required work waits in a local queue. Approve once or deny; a background routine will not silently bypass a required choice.
Keep the receipt
Trace the task afterward: policy result, approval decision, budget event, tool lifecycle, and redacted structured records.
THE APPROVAL PATH
A proposed write
“Create a meeting note” is a request, not a command.
A reviewed policy can allow safe reads to proceed while making a write wait. The person sees the proposed effect, decides once, then inspects the finished task receipt.
Try it locally✓
Decision sequence
Task event trail
- Tool requested
- policy checked
- Approval
- asked
- Your choice
- recorded
- Tool run
- bounded
✓ a denial never starts the action
What Kerna does not do
It is not a new agent framework or a cloud account.
Kerna deliberately keeps domain logic in MCP plugins. That means calendar, notes, search, and other capabilities can evolve independently while the trust boundary stays consistent.
Your workspace
Task history, policies, and receipts live in the workspace you initialize.
Tool capability
External plugins provide domain work through a constrained process boundary.
Your model route
Choose a provider or use local/demo modes; policy remains outside the model.
Your evidence
Inspect the decision trail instead of relying on an opaque “done.”