locale

Agent policy layer for safe tool use across MCP, A2A, and orchestrated agents.

Locale gives hosts, IDEs, and orchestrators a narrow policy boundary between agent intent and tool execution. It issues capabilities, verifies them offline, preflights low-risk actions locally, and evaluates higher-risk actions without turning into another orchestrator.

[](#trust-and-verification) [](./docs/chronology/2026-03-v0.3/manifest.json) [](./docs/verification-workflow.md) [](./LICENSE)

Locale is the project brand. The current package and runtime artifact is `capability-policy-authority`.

Agents can call tools, delegate work, and cross system boundaries faster than policy and verification layers have caught up. Most teams do not need another orchestration system. They need a policy layer that fits into existing MCP hosts, A2A flows, and orchestrators without taking over execution.

Locale is built for that gap:

• issue short-lived capability tokens to agents and orchestrators
• verify capability and decision artifacts offline
• preflight covered low-risk actions locally
• evaluate uncovered or approval-gated actions remotely
• attach decision artifacts to logs, traces, receipts, and audit workflows

An agent policy layer sits between an agent deciding to act and the host actually executing the tool call. Locale keeps that layer narrow on purpose.

• it is not an orchestrator
• it does not execute downstream tools
• it does not broker secrets
• it does not handle payments

That narrow boundary is what makes it natural to integrate into existing systems instead of forcing a platform rewrite.

See [