Four checks around every action.
The model can propose an action, but routing, execution and delivery still have to pass these controls.
Access control
Check whether the contract allows the capability and whether data may leave the device.
Confirmation
Pause when the side effect, risk or contract policy requires a human decision.
Cost control
Reserve and inspect tokens, external calls and currency usage before every call.
Audit gate
Record every routing, call, decision and delivery in an append-only event ledger.
Narrow the set before the model plans.
Instead of stuffing the entire capability registry into context, LibreTask retrieves, ranks and restricts candidates first.
Enrich
Searchable profiles and examples.
Retrieve
Vector and lexical matching.
Rank
Relevance, history, cost and latency.
Restrict
Hard policy and data filters.
Plan
LLM works within the shortlist.
Confirm
Automatic or human-gated action.
The interface can show why a capability passed, required confirmation or was blocked by a hard constraint.
No card, no match. No boundary, no trust.
Capabilities declare their contract, cost, health and risk before the execution engine can call them.
search.web · v1.2.0
MCP is the transport, not the whole trust model.
MCP standardizes discovery and calls. LibreTask adds the identity, timeout, cost, error and human approval semantics required for production execution.
| Concern | LibreTask layer | Result |
|---|---|---|
| Identity | Invocation context | Scope is injected and logged for the call. |
| Time and retry | CapabilityCard policy | Each capability defines its own timeout and retry budget. |
| Failure meaning | Normalized errors | Failures can be evaluated instead of being opaque text. |
| Data exposure | Boundary policy | External destinations are checked before execution. |
Task memory stays on the device. Website operations target European hosting, and exports remain portable.